What it is
A large VLA model trained with 100k+ hours of real-world trajectories and auto-labeled state-transition language, then aligned to robot embodiments and instructions. Instead of relying on scarce teleoperated robot data with manual task labels, it pre-trains on 100k+ hours of UMI trajectories auto-labeled with state-transition language, then post-trains to map those skills to robot embodiments and imperative instructions.
Key results
- RoboCasa365: 57.6% success vs prior SOTA 46.6%
- RoboDojo: 20.07 avg score vs prior SOTA 13.07
- Real-robot out-of-box success: 75% with full pretrain vs 26% no action pretrain
Shows that scaling real-world action pre-training with automatic language labels transfers to stronger zero-shot robot control and data-efficient adaptation.