Remove Track 1 ground-truth annotations
#3
by dzou-nv - opened
Summary
- remove the 30 supplied Track 1 sequences (episodes 37-66)
- remove all object meshes and ground-truth human/object annotations
- retain 37 RGB-video episodes in the existing LeRobot v2.1 layout
- preserve frame/episode/task indexes and original action descriptions
- reduce per-episode metadata to camera, sequence, object, prompt, and video key
Validation
- 37 episodes and videos
- 25,856 contiguous frames
- 33 unchanged task indices and action descriptions
- retained Parquets contain only timestamp, frame_index, episode_index, index,
task_index, and next.done - no GT feature schemas, GT stats, mesh references, or removed sequence IDs remain
The retained video blobs are unchanged and were not downloaded or re-encoded.
zhuxh-nvidia changed pull request status to merged