Daily Paper Briefing

Abstract Overview

ThinkWorld is a vision-language-action framework that represents action reasoning as a Chain-of-Dynamics (CoD), a sequence of discrete motion tokens intended to capture how the world evolves from the current state toward a future one. The method first grounds this shared motion-token space using action-labeled robot demonstrations and annotated human motions, then predicts future visual features conditioned on CoD. It subsequently reinforces sampled CoDs by rewarding agreement between imagined future states and observed future outcomes, which allows temporally ordered videos without action labels to supervise action reasoning. Across simulation and real-robot experiments, the paper demonstrates human-to-robot transfer, robustness to distribution shifts, and compositional skill acquisition from human demonstrations.

Novelty

The paper’s main novelty is formulating intermediate action reasoning as world-grounded Chain-of-Dynamics tokens shared across human and robot embodiments, rather than embodiment-specific controls or unconstrained latent actions. It also introduces consequence-grounded reinforcement learning, where alignment between predicted and observed futures provides the reward signal, enabling learning from action-unlabeled videos while restricting future prediction solely to training time.

Results

ThinkWorld achieves the highest reported average success rates on LIBERO (97.3%) and LIBERO-PRO (55.6%), outperforming VLA-JEPA particularly under task shifts (31.9% vs. 20.8%) and environment shifts (60.9% vs. 41.0%). In real-robot evaluations across in-domain and out-of-distribution settings, it achieves 94.4% macro and 93.3% micro success rates, exceeding MolmoAct2 (62.2% macro, 60.0% micro). Empirical analyses confirm that CoD conditioning improves future prediction accuracy on dynamic regions and that incorporating human video data consistently enhances generalization.

Key Points

1. ThinkWorld introduces Chain-of-Dynamics, a discrete motion-token representation that serves as a shared action-reasoning interface across human and robot data.
2. The training objective reinforces CoDs using agreement between imagined and observed future world states, enabling action-unlabeled videos to provide supervision without explicit motion annotations.
3. The method demonstrates strong generalization under simulation and real-world distribution shifts, and enables the composition of full robot skills guided by human demonstration sequences.

References

• arXiv: https://arxiv.org/abs/2610.37666v1
• Fugu-MT: https://fugumt.com/fugumt/paper_check/2610.37666v1
• Hugging Face Papers: https://huggingface.co/papers/2610.37666
• Project: https://thinkworld-project.github.io/