Driving · 2026-10-28

ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning

Chaehyun Kim, Chi-Pin Huang, Jisoo Kim, Min-Hung Chen, Yu-Chiang Frank Wang, Seungryong Kim, Fu-En Yang

arXiv:2610.37666PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

ThinkWorld is a VLA framework that grounds action reasoning in world dynamics using Chain-of-Dynamics (CoD): a sequence of discrete motion tokens. It initializes CoD with action-labeled demonstrations, then predicts future representations in a frozen DINO feature space and uses consequence-grounded reinforcement learning (rewarding imagined vs observed futures) to learn from videos without motion annotations, enabling human-to-robot transfer and generalization.

Why it matters

The authors address a trade-off in human-video learning: videos are scalable but lack explicit action labels, while robot action-labeled data provides grounding but is limited. By reinforcing intermediate action thoughts based on agreement between imagined and observed future consequences, ThinkWorld turns unlabeled state transitions into supervision and aims to improve generalization across tasks, environments, and embodiments.

Method

  • Represent each action thought as a Chain-of-Dynamics (CoD): a sequence of discrete motion tokens shared across human and robot, learned autoregressively from instruction and visual context.
  • Train CoD-conditioned future imagination by predicting future visual features in a frozen DINO feature space using learnable future queries, instead of reconstructing RGB pixels.
  • Use consequence-grounded RL (world-imagination reward) to reinforce sampled CoDs whose imagined futures better match observed future outcomes; all other components remain frozen during RL and CoD generation/future prediction is not required at deployment.
Abstract (from arXiv)

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 from them faces a fundamental trade-off between scalable data and explicit action grounding. We introduce ThinkWorld, a VLA learning framework that grounds action reasoning in world dynamics and strengthens it through reinforcement learning with world imagination. ThinkWorld represents each action thought as a Chain-of-Dynamics (CoD), a sequence of discrete motion tokens that captures the intermediate dynamics connecting the current world to a future state and serves as a common action representation across embodiments. We first initialize CoD with action-labeled demonstrations, then predict the future induced by each CoD and reinforce those whose predictions better match the observed outcome. This enables videos without action annotations to provide direct supervision from observed state transitions, allowing human and robot experience to refine the same reasoning space. Experiments in simulation and on real robots demonstrate effective human-to-robot transfer and strong generalization to unseen tasks and environments.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-27 · Indexed by Pub-AI: 2026-10-28 UTC

preprint

ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning

Chaehyun Kim, Chi-Pin Huang, Jisoo Kim, Min-Hung Chen, Yu-Chiang Frank Wang, Seungryong Kim, Fu-En Yang

Preprint posted: 2026-10-27 · Original record ↗

arxiv
2610.37666
Explicit source versions (1)
  • v1 · 2026-10-27
  1. Preprint available

    Source date: 2026-10-27 · First detected: 2026-10-28 UTC · Historical baseline

    Source evidence (2)
Abstract · arxiv

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 from them faces a fundamental trade-off between scalable data and explicit action grounding. We introduce ThinkWorld, a VLA learning framework that grounds action reasoning in world dynamics and strengthens it through reinforcement learning with world imagination. ThinkWorld represents each action thought as a Chain-of-Dynamics (CoD), a sequence of discrete motion tokens that captures the intermediate dynamics connecting the current world to a future state and serves as a common action representation across embodiments. We first initialize CoD with action-labeled demonstrations, then predict the future induced by each CoD and reinforce those whose predictions better match the observed outcome. This enables videos without action annotations to provide direct supervision from observed state transitions, allowing human and robot experience to refine the same reasoning space. Experiments in simulation and on real robots demonstrate effective human-to-robot transfer and strong generalization to unseen tasks and environments.