What it is
Large simulated humanoid manipulation dataset with synchronized actions, proprioception, language tasks, and multi-view egocentric video for object pickup. 352,696 simulated Unitree G1 pickup episodes with 52.0M frames, 902 language instructions, 10 fps actions/proprioception, and head+wrist egocentric videos in LeRobot v3 format.
Key results
- 352,696 episodes and 52,008,151 frames
- 902 language instructions for pickup tasks
- 2 cameras at 384×224 plus low-dim state; ~510 GB
Useful for training and evaluating language-conditioned humanoid pickup policies from video and proprioception.