What it is
Large real-robot mobile manipulation dataset with teleoperated HSR trajectories split into independently trainable primitive actions and linked task metadata. 1.18M successful primitive-action episodes and 180.9M video frames from teleoperated Toyota HSRs, structured so each primitive is trainable on its own while metadata preserve its short-horizon task context.
Key results
- 1,184,259 successful primitive-action episodes
- 180,905,084 frames across 5,025 recorded hours
- Collected from 44 robots at 5 physical sites
Useful for training imitation policies for mobile manipulation from primitive actions up to short-horizon tasks.