ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning

Researchers introduced ThinkWorld, a method that reinforces chain-of-dynamics reasoning in Vision-Language-Action (VLA) models by using world imagination to address the limited diversity and coverage of action-labeled robot demonstrations. The approach draws on human videos for broader behavioral and environmental experience, according to the paper's description of the work. The research targets generalization in robotic manipulation, a persistent constraint for VLA models trained on robot demonstrations.

Vision-Language-Action VLA models have made substantial progress in robotic manipulation, but their generalization remains constrained by the diversity and coverage of action-labeled robot demonstrations. Human videos provide much broader behavioral and environmental experience, while learning fro