NVIDIA and KAIST's ThinkWorld VLA Hits 94.4% Real-World Macro Success Learning From Human Video

NVIDIA and KAIST unveil ThinkWorld, a VLA framework that represents actions as discrete 'chain-of-dynamics' motion tokens shared across human and robot embodiments. Trained on 769K episodes and 390M frames, it reaches 94.4% real-world macro success - 80% out-of-distribution on color matching versus 26.7% for MolmoAct2 - and 55.6% on LIBERO-PRO, 11.1 points over VLA-JEPA. Consequence-grounded RL lets it learn compositional skills from action-unlabeled human video.