What it is
Pretrains robot policies from heterogeneous videos by learning 3D-grounded latent actions, then retargets the pretrained model to robot-specific controls. Instead of relying on scarce action-labeled demos or forcing incompatible control spaces together, it learns 3D-grounded latent actions from video transitions and trains a VLA to predict them, then adapts that pretrained policy to each robot's native actions.
Key results
- 98.8% success on LIBERO
- 84.7% on LIBERO-Plus without benchmark-specific adaptation
- 88.6% average success on RoboTwin 2.0 bimanual benchmark
Shows latent-action supervision can scale robot policy pretraining beyond native action-labeled data across embodiments.