pnp/egg: setup.json

pnp/s2rg_egg_0514/setup.json

@@ -0,0 +1,41 @@
+{
+  "task_family": "pick-and-place (pnp)",
+  "video_gen_model": "Wan2.1-Fun-V1.1-1.3B (Mixed-DROID-AGIBOT-3View-DualControl-PrevFrame, droidnew_continue_depthpatch_lora step-8859)",
+  "lora_repo": "zychf/s2r",
+  "inference": {
+    "clip_length": 81,
+    "overlap": 1,
+    "num_inference_steps": 50,
+    "cfg_scale": 5.0,
+    "height": 240,
+    "width": 416,
+    "views": [
+      "ext1",
+      "wrist"
+    ]
+  },
+  "depth_patch": {
+    "ext1": "robot-motion-bbox extended downward = NO-TOUCH zone; sample 1-4 rects outside, each >=10% area, total <50%, fixed across all frames in trajectory",
+    "wrist_chunk1": "frames 0-80: 1-2 rects in depth-empty regions (<=40th percentile of frame-0 wrist depth, >=80% overlap with bg mask)",
+    "wrist_rest": "frames 81+: 1-2 rects in NO-TOUCH side strips (gripper x-range x full height, with small horizontal buffer)"
+  },
+  "prompt_format": "'This video shows in the background, <scene>. In the foreground, on <surface>, there are <task objects>. There are also <N distractors> on the same foreground table. The robot is instructed to <task>.'",
+  "source": "wensi-ai/s2rg_<obj>_0514 (lerobot) under /vision/u/wsai/data/lerobot/",
+  "schema_per_episode": {
+    "ext1_depth_patched.mp4": "patched depth input to model (ext1)",
+    "ext1_mask.mp4": "robot RGB mask input (ext1)",
+    "ext1_generated.mp4": "model output, ext1 view",
+    "wrist_depth_patched.mp4": "patched depth input (wrist)",
+    "wrist_mask.mp4": "robot RGB mask input (wrist)",
+    "wrist_generated.mp4": "model output, wrist view",
+    "patch_info.json": "chosen rect coords + NO-TOUCH zones",
+    "prompt.txt": "exact prompt fed to the model"
+  },
+  "created_utc": "2026-05-23T02:31:28Z",
+  "dataset_short": "egg",
+  "task_string": "Pick up the egg and place it inside the bowl.",
+  "target_episodes": [
+    0,
+    500
+  ]
+}