Reproduce “ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chaehyun Kim, Chi-Pin Huang, Jisoo Kim, Min-Hung Chen, Yu-Chiang Frank Wang, Seungryong Kim, Fu-En Yang
ThinkWorld addresses robot-learning by grounding robot actions and annotated human hand motion in a shared discrete Chain-of-Dynamics (CoD) vocabulary. It trains on action-unlabeled videos by rewarding tokens whose predicted futures match observed futures. At deployment, a separate expert predicts continuous actions. In real-robot trials, it achieved 94.4% macro success across six task-condition combinations and succeeded in 52/60 OOD trials.
Key insights
- Observed future frames can provide a training signal for action reasoning even when the video has no action labels and candidate actions are never executed.
- In ablations on LIBERO-PRO, adding future prediction, CoD, and reinforcement learning successively raises performance from 40.1% to 55.6%, suggesting each component contributes.
- Human demonstrations can communicate the order of familiar robot skills, enabling the tested robot to perform some composite sequences without a robot demonstration of the full sequence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.