# Motus2: a self-evolving world model for dexterous manipulation

> Motus2 is a research project that couples three heads of one shared-weight model into a closed loop, a policy that proposes actions, a simulator that predicts their visual consequences and an evaluator that scores them, and it is being released as code and checkpoints progressively through September 2026. The repository today holds documentation, a project page, an arXiv paper and a citation block, so the interesting reading is the method, not the software.

**shengshu-ai/Motus2** — Motus2: A Self-Evolving General World Model for Dexterous Manipulation

- Repository: https://github.com/shengshu-ai/Motus2
- Stars: 66 · Forks: 1
- Language: Unknown
- License: not declared
- Published: 2026-10-21 · Updated: 2026-10-21 · Language: en
- Canonical page: https://hysenlabs.com/projects/shengshu-ai-motus2

## The complaint is about how world models are wired up

The project states its motivation before it states its method, and the motivation is a design critique rather than a performance claim. General embodied agents should perceive, predict, act, evaluate and improve within a unified system. World models have shown promise in building such agents, yet, the argument runs, existing models typically append an action output head to a world simulator without coupling them into a closed decision and learning loop for policy improvement. The distinction is precise and it matters. In the common arrangement the world model is a forward model of what happens next and the policy is a separate module that consumes it, so the policy improvement loop has to be bolted on afterwards, with its own data, its own objective and its own failure modes. Motus2's claim is that the coupling should be structural rather than an arrangement of separate parts.

## One model, three control interfaces

The mechanism is model scaling, and it is stated with unusual clarity in the overview. A single model with shared weights exposes three control interfaces. The first is the policy, described as a world-action model, which proposes candidate action chunks. The second is the simulator, an action-conditioned world model, which predicts the visual consequences of those proposed chunks. The third is the evaluator, a value model, which assesses the predicted outcomes. The three are then coupled: the policy proposes, the simulator predicts what would be seen, the evaluator scores that prediction, and the result feeds back into the policy. That is the closed decision-and-learning loop the introduction argues existing work lacks. The interesting consequence is that the same weights have to be useful for three different jobs, and the data each job needs is not the same. The formulation handles that asymmetry explicitly, since it uses curated expert demonstrations for action learning, while failed and suboptimal interactions provide the evidence for dynamics modelling and value learning. Failures, in other words, are not waste, they are training data for two of the three heads.

## Data scaling in three stages

The second axis of the work is data scaling, and the stages are laid out in order. The project moves from large-scale monocular egocentric data to synchronized stereo egocentric data, and then to robot-domain adaptation using robot trajectories and supplementary human-robot alignment data. The logic is a narrowing of the domain. Human first-person video is abundant and can be collected without a robot, but it is monocular and it is not robot data, so it teaches the model about objects and manipulation without telling it about its own embodiment. Stereo data adds depth. The robot-domain adaptation stage then reconciles that human data with the actual robot, and the alignment data is there to bridge the remaining gap between a human hand and a dexterous hand. The scaling in both axes, more model and more data, is what the project presents as a path toward general dexterous manipulation rather than a collection of separate tricks.

## Context extensions, tactile feedback and a biomimetic platform

Three extensions are studied on top of the core formulation. The first concerns context handling, with the project studying global-autoregressive and hybrid-memory extensions of its sliding-window context, which is the usual way to trade memory for fidelity in a long video prediction. The second is tactile feedback for contact-aware control, which is where dexterous manipulation differs most sharply from pick and place, since the interesting failures happen at the moment of contact and not in the approach. The third is the physical instantiation: a fully biomimetic platform with stereo vision, dual arms, dual dexterous hands and tactile sensing. Putting a world model and a robot in a loop on hardware with that much sensing is a demanding test of the method, and it is also what makes the evaluation meaningful, because the platform can produce the kind of contact-rich failures that the training data is supposed to cover.

## What is actually released right now

This is where an engineer's expectations need adjusting. The repository contains a single top level entry, the README, and the open-source roadmap is a list of unchecked boxes. Stage 1 pretraining checkpoints, stage 2 pretraining checkpoints, the video pretraining code for stage 1, the video-action value training code covering stage 2 pretraining, mid-training and post-training, the MBRL code, the memory and variable-length training infrastructure, and the tactile code are all marked as not yet released. The updates section says the team plans to release the Motus2 code and checkpoints progressively throughout September 2026, and the arXiv entry is listed at version 2. The links section points to a project page, the paper and a Hugging Face organisation for the models. So the artefacts to look at today are the paper and the project page, with the model weights on Hugging Face as the first code-adjacent thing to check. The README does not describe an installation procedure, a configuration file, a dataset layout or a licence, and the repository records no licence at all.

## Citation expectations and how to follow the work

The project asks users to cite the work, and the README carries a BibTeX block ready to paste into a paper, with the title, a long author list, the year 2026, the eprint identifier 2608.30237, the arXiv archive prefix and the primary class cs.RO:

```tex
@misc{bi2026motus2,
  title         = {Motus2: A Self-Evolving General World Model for Dexterous Manipulation},
  author        = {Hongzhe Bi and Zihao Zhou and Yihang Tang and Jingrui Pang and
                   Shuhe Huang and Haitian Liu and Runqing Wang and Shuai Huang and
                   Yichen Wang and Yiming Cheng and Ruowen Zhao and Zhenghua Li and
                   Hengkai Tan and Xiaolong Liu and Jinhui Wan and Jiabao Liu and
                   Min Zhao and Fan Bao and Jun Zhu},
  year          = {2026},
  eprint        = {2608.30237},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2608.30237}
}
```

The affiliation implied by the repository owner and the model organisation on Hugging Face is a research group rather than a long-standing open source project, and the repository carries no contribution guide, no tests and no build files, which is consistent with a research code drop in progress. The practical takeaway for an engineer is that the evaluation claims in the overview are claims from a paper, not from something you can run, and the fastest way to check them is to read the arXiv paper and to watch the Hugging Face organisation for the pretraining checkpoints.

## Conclusion

Motus2 is worth following if you work on robot learning with dexterous hands, because the closed decision and learning loop is a different bet from the usual world model with an action head bolted on, and the biomimetic platform with stereo vision, two arms, two hands and tactile sensing gives the evaluation a concrete setting. It is not adoptable software today, since the repository holds only the README and every item on the open-source roadmap is still unchecked. Before planning around it, watch the Hugging Face models and the project page for the stage 1 and stage 2 pretraining checkpoints, read the arXiv paper rather than the summary, and check the licence terms, which the repository does not currently record.

## FAQ

### What is Motus2?

Motus2 is a self-evolving general world model for dexterous manipulation, presented through a paper, a project page and a Hugging Face organisation for models. It couples a policy, a simulator and an evaluator into a closed decision and learning loop for policy improvement.

### What are the three control interfaces in Motus2?

They are the policy, which is a world-action model proposing candidate action chunks, the simulator, an action-conditioned world model predicting their visual consequences, and the evaluator, a value model assessing the predicted outcomes. All three come from a single model with shared weights.

### Is the Motus2 code and checkpoints released?

Not yet in the repository. The open-source roadmap lists stage 1 and stage 2 pretraining checkpoints, the video pretraining and video-action value training code, the MBRL code, the memory infrastructure and the tactile code, all unchecked, and the updates section says the team plans to release code and checkpoints progressively through September 2026.

### How should Motus2 be cited?

The README provides a BibTeX entry with the title, the author list, the year 2026, the eprint 2608.30237, the arXiv prefix and the primary class cs.RO. The paper itself is listed on arXiv at version 2.

## Sources

- [Issues](https://github.com/shengshu-ai/Motus2/issues)
- [README](https://github.com/shengshu-ai/Motus2/blob/main/README.md)
- [shengshu-ai/Motus2 on GitHub](https://github.com/shengshu-ai/Motus2)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/shengshu-ai-motus2