BotFails, H.264 224x224 LeRobot v3 copy
Derived from kantine/BotFails at revision
3478e49d91e1737eb76dfee2d81bb22617039c13 (Apache-2.0). Transcode only; all credit to the original authors
(FIDeL, Rolland, Mayran de Chamisso, Mouret, ICRA 2026, arXiv 2604.13788). Please cite the original dataset and paper.
What changed: the 10 tasks x (expert + test) LeRobot v2.0 datasets (AV1 640x480, one file per episode) were merged into one LeRobot v3.0
dataset (all 178 expert episodes first, then the 145 test episodes), videos re-encoded to H.264 224x224 (area scaling, all-intra,
CRF 18, yuv420p), with frame counts and timestamps checked per file. Cameras observation.images.logitech_1 / logitech_2, 30 fps,
action and state 6-dim. Parquet values are unchanged.
Extras: botfails_map.json (global episode id -> split, task, original episode, frames, label file), labels/epNNNNNN.npy (per-frame
0/1 anomaly labels of the test episodes, copied from the original CSVs), split.json (the world-model split used in
road-to-robotics). Known issues of the source: domotic_setTheTable test episode 7
has no label file; industrial_screws_sorting test episode 7 has 1491 labels for 1492 frames.
The labels mark anomalies (benign and failures); the failure-vs-benign split is not public.
- Downloads last month
- -