Machine Learning

Motus2: A Self-Evolving General World Model for Dexterous Manipulation

Hongzhe Bi, Zihao Zhou, Yihang Tang, Jingrui Pang, Shuhe Huang, Haitian Liu, Runqing Wang, Shuai Huang, Yichen Wang, Yiming Cheng, Ruowen Zhao, Zhenghua Li, Hengkai Tan, Xiaolong Liu, Jinhui Wan, Jiabao Liu, Min Zhao, Fan Bao, Jun Zhu

Featured September 7, 2026

Original Paper

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

A robot brain called Motus2 learns to do tricky hand tasks by imagining what might happen if it tries different actions, deciding which outcome is best, and then using that feedback to get smarter, just like practicing in its head.

In depth

Motus2 introduces a self-evolving general world model for dexterous manipulation by unifying policy, simulator, and evaluator interfaces within a single shared-parameter model. This architecture enables a closed decision-and-learning loop where the model proposes actions, predicts their visual consequences, evaluates outcomes, and uses this feedback for policy improvement, moving beyond static imitation learning. The approach also leverages hierarchical egocentric data scaling and tactile feedback for robust physical grounding.

Key Takeaways

1
A shared-parameter General World Model unifies policy, simulator, and evaluator interfaces, enabling a closed decision-and-learning loop for dexterous manipulation.
2
The model employs value-guided self-evolution through Model-Based Reinforcement Learning (MBRL) and Best-of-N planning, using predicted outcomes and their values to refine action policies.
3
A hierarchical data-scaling curriculum progresses from large-scale monocular to stereo egocentric human data, followed by robot-domain adaptation and integration of tactile feedback for contact-sensitive control.

Conceptual Flow

HIGH LEVEL

Methodology: Self-Evolving General World Model

The robot brain learns to act, imagine, and judge all at once, using its own predictions to get smarter over time.

Robot Sees

Robot Hears Goal

Think & Plan

Try Action

Imagine Future

Judge Outcome

Results: Improved Dexterous Manipulation

By learning from lots of human videos and practicing in its head, the robot can now do complex tasks much better than before.

Old Robot Brain

Limited Practice

Learn & Improve

New Robot Brain

Better Hand Skills

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.