ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning

Overview & Motivation

Every video records the consequences of an action

Robot data is scarce. Human video is abundant but has no robot actions. ThinkWorld learns from both.

Latent action models

Any unlabeled video
May encode camera or appearance

Explicit motion supervision

Physically grounded
Needs pose estimation & retargeting

ThinkWorld

Motion tokens shared by humans & robots
Supervised by observed futures, no labels

Learning from human and robot experience

Eight datasets of egocentric human video and robot demonstrations.

Human egocentric videoRobot demonstrations

4,276 hof video

44%egocentric human video, no action labels

8datasets, one shared motion vocabulary

Method

Reason in dynamics, then check the reasoning against the world

**Two training stages.** (a) Learn to write a Chain-of-Dynamics and imagine its future. (b) Reinforce the chains whose imagined future matches the real one.

Reasoning at training time, fast at deployment

Chain-of-Dynamics and imagination are used only during training, so inference runs as fast as a standard VLA.

Real-World Experiments

Generalization on a bimanual robot

Three bimanual tasks. Each is trained on in-domain setups and tested on changes the policy never saw.

Color Matching

“Put the fruits in the matched color bowl.”

In-domainOut-of-distribution

Cup Hanging

“Put the cup on the hanger.”

In-domainOut-of-distribution

Cup Hanging, new environment — New environment

Sorting

“Sort the same objects into the baskets.”

In-domainOut-of-distribution

Sorting, new trajectory — New trajectory

Success rates

ThinkWorld succeeds in 60/60 in-domain and 52/60 out-of-distribution trials.

Side-by-side rollouts

Same scene, four policies. Baselines break down when the scene changes; ThinkWorld keeps succeeding.

Show teleoperated training demonstrations for this task

Skill composition from human video

Robot data covers only atomic skills; the full sequence appears only in human videos. ThinkWorld performs the whole sequence in one rollout.

Human demonstration · egocentric

ThinkWorld rollout · robot, single episode

Robot training data: atomic skills only

Analysis

What does Chain-of-Dynamics capture?

CoD drives the imagined future

Only the generated CoD moves the imagined future toward the real one. Shuffled, empty or missing CoDs stay close to the current frame.

CoD matters where the scene moves

Similarity to the true future on the 25% of patches that change the most.

Colors are PCA projections of DINOv3 features.

CoD attends to what moves

CoD tokens focus on hands, grippers and the objects they move.

Overlay

Simulation Benchmarks

Robust to task and environment shifts

Trained on LIBERO, tested zero-shot on LIBERO-PRO, which changes objects, positions, instructions, goals and environments.

Success rate (%). Bold: best, underline: second best. † reproduced by the authors.

97.3%LIBERO average, best overall

55.6%LIBERO-PRO average, best overall

+11.1pts on Task shifts vs. VLA-JEPA

+19.9pts on Environment shifts vs. VLA-JEPA

Additional Results

Imagination stays on track in closed loop

During closed-loop rollouts, the imagined future (+1 s) keeps tracking what actually happens. Scrub through an episode to compare.

Citation

BibTeX

@article{kim2026thinkworld,
  title   = {ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning},
  author  = {Kim, Chaehyun and Huang, Chi-Pin and Kim, Jisoo and Chen, Min-Hung and Wang, Yu-Chiang Frank and Kim, Seungryong and Yang, Fu-En},
  journal = {arXiv preprint arXiv:2610.37666},
  year    = {2026}
}