ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning
ThinkWorld uses videos without action labels to train robot action reasoning from observed changes in the world.
The framework represents actions as sequences of discrete motion tokens connecting a current state to a future one. It first learns from action-labeled robot demonstrations, then rewards sequences whose predicted outcomes better match what a video shows. This gives human and robot experience a shared representation for refining action reasoning. The authors report effective human-to-robot transfer and strong generalization to unseen tasks and environments in simulation and real-robot experiments.
HF Daily Papers' note
The framework represents actions as sequences of discrete motion tokens connecting a current state to a future one. It first learns from action-labeled robot demonstrations, then rewards sequences whose predicted outcomes better match what a video shows. This gives human and robot experience a shared representation for refining action reasoning. The authors report effective human-to-robot transfer and strong generalization to unseen tasks and environments in simulation and real-robot experiments.
HF Daily Papers' note
score 5