Physical AI · Egocentric Manufacturing
Factory capture, labeled for training
Dataset teaser, hand-pose samples, and subtask-level scene segmentation from live manufacturing floors.
Teaser
Mosaic preview of live factory headset captures from the egocentric manufacturing corpus.
Hand detection
Two recovered hand-pose overlays on factory capture — the motion signal used for robot retargeting.
Hand pose
Scene segmentation
Every clip is cut into ordered subtasks with sub-second boundaries, skill labels, per-step outcome, and process narration. Play a clip or jump to any step.
Review becomes a query
Because every step carries a skill, an outcome, and exact boundaries, finding failures across the corpus is a filter — not a scrub through video.
where skill = "inspect"
and outcome = "fail"
return clip_id, step, t0, t1
and outcome = "fail"
return clip_id, step, t0, t1
Loading matches…