Field notes
Notes from the refinery floor.
Engineering notes on capture, consent, QA gates, and export pipelines. Written for the people who will train on this data.
Why 10-hour takes beat 10,000 clips
A clip ends where the interesting part begins: the transition, the interruption, the recovery. What shift-length capture keeps, with segment statistics from the packaging line.
Read the note → ACTIONS · 2026-06-18Recovering TCP pose from a camera on a gripper
The X3 rides the gripper and becomes its own tracking system. Per-frame TCP pose and width, and the sync gate that keeps them honest.
Read the note → CONSENT · 2026-06-04Consent as a schema field, not a PDF
DPDP Sec-6 artifacts that join sellability at query time. Withdrawal reaches the catalog by itself, not through a legal inbox.
Read the note → QA · 2026-05-21Reason-coded rejects, or why we log every retake
A gate that discards in silence is a filter no one can audit. The four codes: sync, blur, occlusion, task validity. What the reject log tells a buyer.
Read the note → EXPORT · 2026-05-07One master, three formats: exporting to LeRobot v3, RLDS, and GR00T
Exports are views over the masters. The schema mappings between the three formats, and what re-export instead of re-collection saves in practice.
Read the note → HANDS · 2026-04-23MANO at shift scale: hand pose over 10 continuous hours
Hand pose is easy for forty seconds and hard for ten hours. Occlusion, drift, and the per-take thresholds we hold ourselves to.
Read the note →