LinhanWang/ABC-130k-lerobot

ABC-130k (LeRobot v3)

LeRobot v3-format conversion of XDOF/ABC-130k, the bimanual YAM-station manipulation dataset, converted from the original MCAP logs.

Contents

• 343 self-contained LeRobot v3 datasets, one per task unit (large tasks are split into __partNN chunks — a multi-dataset loader consumes the full list).
• 118,685 episodes / 199 tasks.
• 232,430,937 frames (~3,228 h at 20 Hz).
• 3 cameras: observation.images.top, observation.images.left_wrist, observation.images.right_wrist, each 240×320 (source frames center-cropped to 4:3).
• State 14-dim, action 14-dim, resampled to a 20 Hz grid.

Layout

abc_<task>_lerobot/            # or abc_<task>__partNN_lerobot/ for chunked tasks
  meta/   info.json, stats.json, modality.json, tasks.parquet, subtasks.jsonl, episodes/
  data/   chunk-000/...        # parquet, per-frame state/action
  videos/ observation.images.{top,left_wrist,right_wrist}/...   # mp4

Each subfolder is a standalone LeRobot v3 dataset and can be loaded on its own.

Conversion notes

Converted with the pipeline in examples/abc130k/ of the StarVLA repo. Every episode is verified so the video frame count matches the parquet row count (length == round((to_ts - from_ts) * fps)); 0 units failed verification and 0 frame-drops.

License & attribution

Released under Apache-2.0, matching the source dataset XDOF/ABC-130k. Please cite the original ABC-130k authors when using this data.