What it is
2,000 hours of high-fidelity robot-free bimanual manipulation data with synchronized 6-view video, trajectories, gripper states, and language. 2,000 hours of robot-free bimanual demos with 6 synchronized camera views, calibrated 20D state/action trajectories, gripper states, task text, and QC metadata across 480+ scenes.
Key results
- 2,000 released hours from a 20,000+ hour corpus
- 6 synchronized camera views per episode
- CC BY 4.0; Parquet tables + MP4 videos
Useful for pretraining and post-training robot manipulation policies from human demonstrations without robot teleoperation.