Can World and Action Modeling Benefit Each Other under Unified Causal Training?
OneWM’s joint training improved both robot policy performance and action-conditioned visual prediction, the authors report.
The framework uses a shared video-action architecture, with actions serving as inputs for world prediction and targets for policy learning. Learning the visual consequences of specified actions improved policy performance and generalization across changing scenes. Action learning also improved multi-view world-prediction fidelity within the same architecture. The evaluation covered LIBERO, LIBERO-Plus and RoboCasa365, alongside real-world policy experiments.
HF Daily Papers' note
The framework uses a shared video-action architecture, with actions serving as inputs for world prediction and targets for policy learning. Learning the visual consequences of specified actions improved policy performance and generalization across changing scenes. Action learning also improved multi-view world-prediction fidelity within the same architecture. The evaluation covered LIBERO, LIBERO-Plus and RoboCasa365, alongside real-world policy experiments.
HF Daily Papers' note
score 5