KRVLA real-robot dataset
859 episodes, 185269 frames at 25 FPS. RGB video dimensions: 384x384. Human leader/follower dynamic pick-and-place.
Existing episode data, videos and quality files are retained byte-for-byte when new episodes are appended. Metadata and aggregate statistics describe the cumulative dataset. Previous revisions remain in Git history.
DOM-compatible LeRobot v2.1 export. Object pose/velocity are visible-surface estimates; orientation is assumed. Operator success is not independent placement verification. Inspect quality sidecars and export_report.json. Use explicit episode IDs in meta/splits.json. Raw RGB-D archives remain local.
KRVLA training
Use the processed training branch with revision="training". This main branch preserves raw exports. The collection workflow refreshes the processed branch with depth-corrected labels, quality handling, explicit splits and train-only statistics after each upload.
- Downloads last month
- 885