NOJIMA21/so101-egg-cracking

SO-101 egg-cracking teleoperation: task success, data-capture failure

One public LeRobot v3-format teleoperation episode recorded with a Hiwonder SO-ARM101 follower arm.

Result

The egg-cracking manipulation itself succeeded under human teleoperation. The episode is published as an honest engineering record: the physical task succeeded, while the camera-data pipeline did not remain reliable for the entire recording.

• 1 episode / 4,301 frames / approximately 143 seconds at 30 FPS
• Robot action and state vectors for six joints, including the gripper
• H.264 video streams for fixed overhead and wrist views
• No audio

What went wrong

The wrist-camera stream is valid at the beginning of the episode but later contains an unintended overhead-camera view after a Windows camera-routing issue. The dedicated overhead_camera stream is recorded throughout the episode.

This makes the episode unsuitable as-is for training that assumes a stable wrist viewpoint. It remains useful for inspecting the successful task, the action/state trajectory, and the failure mode of a practical USB-camera data pipeline.

Corrective actions for subsequent recordings

1. Resolve cameras at launch using immutable USB VID/PID values instead of remembered OpenCV indices.
2. Exclude the notebook's built-in webcam from robot-camera selection.
3. Force the OpenCV DirectShow backend on Windows instead of automatic backend selection.
4. Before long recording, run a two-camera preview and verify both views remain distinct for several minutes.
5. Reject or label any episode whose camera stream changes viewpoint mid-recording; do not mix it with clean imitation-learning data.

Task

Manipulate and crack an egg under human teleoperation.

Privacy

Published with the recorder's authorization. It contains workspace imagery; please do not use it for identity, surveillance, or unrelated personal-data analysis.