| --- |
| license: apache-2.0 |
| task_categories: |
| - robotics |
| tags: |
| - LeRobot |
| - so101 |
| - so_follower |
| configs: |
| - config_name: default |
| data_files: data/*/*.parquet |
| --- |
| |
| # pick-red-cube |
|
|
| A teleoperated SO-101 (`baby_gewu`) manipulation dataset: **"Grab the red cube |
| and put it in the box."** Recorded with a leader/follower SO-101 pair; the |
| follower's joint states and three rig cameras (top / wrist / side → |
| `camera1` / `camera2` / `camera3`) are captured per episode. Intended for |
| finetuning SmolVLA-class policies. |
|
|
| - **Robot:** `so_follower` (SO-101), 6-DoF (`shoulder_pan`, `shoulder_lift`, |
| `elbow_flex`, `wrist_flex`, `wrist_roll`, `gripper`) |
| - **Episodes:** 122 |
| - **Frames:** 50,355 |
| - **FPS:** 30 |
| - **Cameras:** 3 × 640×480 (top / wrist / side) |
| - **Format:** [LeRobot](https://github.com/huggingface/lerobot) v3.0 |
|
|
| ## Episode breakdown |
|
|
| All episodes share the same task string; the failure-recovery split below is |
| tracked by episode index, not encoded in the per-episode task field. Episodes |
| 92–121 were added to teach corrective behavior (the clean-only set never shows |
| the policy how to recover from a bad approach). |
|
|
| | Episodes | Kind | Count | |
| | --- | --- | --- | |
| | 0–91 | **Pristine** — clean approach, clear the cube, grasp center-of-faces | 92 | |
| | 92–101 | **Corner grasp** — grabbing a corner, cube slides, recover + regrasp | 10 | |
| | 102–111 | **Pushing into the table** — fingertip contact, back off, re-approach | 10 | |
| | 112–121 | **Air grasp** — closing beside/above the cube, re-center + regrasp | 10 | |
|
|
