SO-101 Sword on Stand Clean32
LeRobot v2.1 dataset for the task:
Pick up the sword and place it on the sword stand.
Contents
- Robot: LeRobot SO-101 follower
- Episodes: 32
- Frames: 8,636 (about 4.80 minutes at 30 Hz)
- Cameras:
observation.images.fixed,observation.images.wrist - Capture: RGB, 640×480, 30 FPS, AV1 MP4
- State/action: calibrated 6D joint positions
- Joint order:
shoulder_pan,shoulder_lift,elbow_flex,wrist_flex,wrist_roll,gripper - Stored action representation: absolute calibrated joint-position target
The source recordings were not modified. This clean derivative excludes three operator-reported recovery trajectories and one accidental 1.2-second static recording. Videos were retained without re-encoding.
Audited split
The physical files contain all 32 episodes in one LeRobot dataset. Use the
explicit split below rather than the broad 0:32 entry in meta/info.json:
- Train: derived episodes
0..29(30 episodes) - Validation: derived episodes
30,31(2 episodes)
Only episodes 0..29 should be used to compute normalization statistics.
See meta/pi05_split_manifest.json and meta/source_episode_map.json.
Important limitations
This is a narrow, nearly single-setting dataset. Object placement, camera pose, camera focus, illumination, and robot start pose should be treated as in-domain constraints. It is not evidence of broad spatial generalization. The wrist camera must remain sharply focused at deployment time.
For OpenPI π0.5 notes, read PI05_BC_HANDOFF_CN.md before writing a training
configuration.
- Downloads last month
- 215