Quality and review
All 5,308 training frames are checked against source-frame mappings. Full video decode and standalone reader checks are recorded in validation.json. Face detection ran on every displayed training frame and returned zero detections for these excerpts. Audio and source location/container metadata are omitted from public videos.
The assistant reviewed RGB storyboards at approximately two-second intervals and clip endpoints, and estimated-overlay frames near the start, middle and end. No visible faces or identifying private text were found in inspected frames. This is a sampled visual review, not an exhaustive manual privacy audit or an independent annotation-accuracy evaluation.
Estimated annotation coverage
Hand keypoints were requested at 10 Hz and object detection at 5 Hz. Counts below describe timestamps with emitted estimates; they do not measure recall against visible objects/hands. Unknown percentages use temporal hand instances (one track at one timestamp), not keypoint rows. A track is not an independent person or physical entity.
| Excerpt | Raw hand timestamps | Unknown handedness | Max temporal hands at one timestamp | Object timestamps |
|---|---|---|---|---|
| remove-tape | 290 | 24.2% | 2 | 145 |
| brush-front-grille | 300 | 50.0% | 2 | 150 |
| brush-housing | 300 | 100.0% | 2 | 137 |
| brush-panel-channels | 300 | 99.3% | 3 | 144 |
| wipe-panel | 300 | 13.7% | 3 | 150 |
| wipe-panel-corners | 280 | 24.0% | 3 | 135 |
The raw hand candidate selection caps detections at two per timestamp. Temporal interpolation can briefly retain extra tracks. The public tables preserve raw and temporal estimates separately, with null z/world coordinates and explicit provenance. Some visible hands are missed; finger placement, small-tool boxes and handedness are imperfect. An object track may change ID during one manipulation. Review overlays intentionally expose these errors. Overlays use a constant-frame-rate display clock and are visual aids, not the alignment reference.
Language and grouping
The assistant prepared the instructions and 13 action intervals from visual review. Boundaries are approximate to two seconds and have not been independently checked by a human annotator. The intervals describe observed actions, not task success or robot commands. One conservative session group covers all six excerpts; do not split excerpts from this session between training and evaluation and then claim independent generalization.
Capture limitations and next collection
This release covers one camera wearer, one appliance/work area and one apparent session. Viewpoint is chest-mounted according to the owner. The public 1920 x 1080 images are tone-mapped from iPhone HLG HDR; full original MOV files remain with the owner. Exposure changes, motion blur, self-occlusion and repetitive activity are visible. Metric geometry, forces, tactile measurements and robot control/state are unavailable.
For a paid capture pilot, agree on target tasks, full task sequences, multiple workers/sites, tool visibility and acceptance criteria. A labeled evaluation subset and held-out sessions would support measured annotation quality and downstream utility. This sample establishes delivery structure and inspectable footage; it does not report training gains.