- Clamp or bolt both SO-101 bases to a rigid work surface.
- Place the hashtag board flat and prevent it from sliding.
- Place one pickup-zone frame on each side used during data collection, keeping both completely inside the corresponding arm's safe reach.
- Put six red X tokens in the X zone and six light-grey O tokens in the O zone.
- Fasten the camera tower to the table, route USB cables down its rear channel and keep them outside both arm sweeps.
- Mount the top camera as
observation.images.top; mount the wrist camera on the arm used by the recorded follower asobservation.images.wrist. - With torque disabled, compare both views to several episodes in the public LeRobot dataset viewer.
The dataset records the actual trained bench but this release does not yet include a dimensioned base-to-board fixture drawing. Do not infer millimetre coordinates from this document. The pretrained behavior is sensitive to arm base position, board/pickup-zone position, top-camera pose, wrist-camera pose, lighting, token colour and follower calibration.
Until a dimensioned fixture is released, treat visual matching as a diagnostic, not a calibration method. Perform a torque-off reach check, then a bounded single-move rollout with the physical E-STOP in hand. If geometry differs, record/fine-tune for that bench rather than increasing controller limits.
The top camera is rotated 180 degrees relative to the model's task coordinates:
| Top-camera cell | Model/task cell |
|---|---|
| 1 | 9 |
| 2 | 8 |
| 3 | 7 |
| 4 | 6 |
| 5 | 5 |
| 6 | 4 |
| 7 | 3 |
| 8 | 2 |
| 9 | 1 |
This mapping is enforced by the deterministic controller and must not be compensated by swapping camera names or mirroring images.