ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning

TL;DR

ThinkWorld introduces a vision-language-action framework that uses world imagination and chain-of-dynamics reasoning to generalize robotic manipulation beyond action-labeled demonstrations. By representing actions as sequences of motion tokens that model intermediate dynamics, it leverages unannotated human videos for direct supervision via predicted state transitions, enabling robust human-to-robot transfer and generalization.

Why It Matters

This approach overcomes the scalability bottleneck of action-labeled robot data by utilizing rich, unstructured human videos as training signals, significantly improving generalization across tasks and environments without requiring explicit action annotations.

Key Findings

• Chain-of-Dynamics (CoD) provides a unified, embodiment-agnostic action representation grounded in world dynamics.
• World imagination via prediction of future states enables reinforcement learning from unannotated human videos.
• ThinkWorld achieves strong zero-shot generalization to unseen tasks and environments in both simulation and real-world robotics.