Third Origin's synced release. 40 takes of shoe-factory work, 36.2 hours of head-mounted video, each take narrated action by action and rated by an expert. We used it as the ego pre-training stage for an SO-ARM101 policy twice, once as released and once with 143 minutes removed, then ran both on six bench tasks, 127 rollouts each. The edited version won.
A same-size random cut scored 23.6%, so the gain is from which minutes were removed, not from removing minutes. Fold and lid, the two tasks with the most matching footage, gained most.
Edited minus full, with its 95% range. The pooled lift clears zero. Single tasks at 127 rollouts mostly do not, and the chart says so.
Measured from the release. 40 synced ego and exo takes, 23,650 narrated actions with a timestamp, an expert rating on speed, precision and fluency for every take. 15 takes are rated expert, 22 proficient, 3 beginner. Scenarios are shoe assembly 21, sole attachment 9, shoe packaging 8, stitching 1, shoe finishing 1. The seller already grades its own workers, which is rare, and it makes the corpus a clean test of whether our score agrees with a human expert.
Measured, from the annotations. The narrator marks moments where the wearer is not doing the task, and in one take writes that the camera was removed at 23 minutes while the file ran on for 57 more. Those minutes plus the moments spent waiting, fetching, being taught or talking, plus the three takes rated beginner, are the cut. The three takes with a minor annotated mistake are kept; one smear of glue in 107 minutes is not a reason to lose the take.
| Why it was removed | segments | minutes |
|---|---|---|
| camera off, still recording | 3 | 57.7 |
| waiting, fetching, being taught, talking | 49 | 14.8 |
| whole take, annotator rates the worker a beginner | 3 | 70.3 |
| Removed | 142.8 | |
| Kept | 2032 |
Head camera, twelve seconds each, at the moment the annotation names. Six removed, two kept for contrast.