| --- |
| license: other |
| license_name: attribution-noncommercial-sharealike-4.0-international |
| license_link: LICENSE |
| tags: |
| - WAM |
| - WM |
| - world model |
| - world action model |
| language: |
| - en |
| - zh |
| base_model: |
| - Wan-AI/Wan2.2-TI2V-5B |
| --- |
| |
| <div align="center"> |
|
|
| <img src="assets/logo1.png" width="180" alt="Thinker-WM Logo"> |
|
|
| <h1>Thinker-WM: A Unified World Model with Shared Multimodal Backbone for Embodied Intelligence</h1> |
|
|
| **Thinker Team, Ubtech** |
|
|
| <a href="https://thinkercosmos.ubtrobot.com/thinker-wm/"><img src="https://img.shields.io/badge/Project-Page-1E9BD7?style=for-the-badge" alt="Project Page"></a> |
| <a href="https://github.com/UBTECH-Robot/Thinker-WM"><img src="https://img.shields.io/badge/GitHub-Repository-black?logo=github&style=for-the-badge" alt="GitHub"></a> |
| <a href="https://arxiv.org/abs/2610.18447"><img src="https://img.shields.io/badge/arXiv-Paper-b31b1b?logo=arxiv&style=for-the-badge" alt="arXiv"></a> |
| <a href="https://github.com/UBTECH-Robot/Thinker-WM/blob/main/LICENSE"><img src="https://img.shields.io/badge/License-CC_BY--NC--SA_4.0-green?style=for-the-badge" alt="License"></a> |
|
|
| </div> |
|
|
| ## Model Description |
|
|
| Thinker-WM is a unified multimodal world model that jointly models and generates **RGB, depth, and action** through a shared intermediate Transformer backbone with modality-specific input/output experts. It is pretrained on **20K hours of egocentric data** and **10K hours of robot data** (≈4K hours with aligned depth supervision), and ranks **1st** on three public benchmarks: **EWMBench**, **World Olympiad**, and **World Arena**. |
|
|
| This repository contains the official **inference release** of the model weights. For full documentation, installation, and usage, see the [GitHub repository](https://github.com/UBTECH-Robot/Thinker-WM). |
|
|
| ## Supported Inference Modes |
|
|
| | mode | inputs | outputs | |
| |---|---|---| |
| | `wam` | prompt + first_frame + robot | video + actions (+depth) | |
| | `wm` | prompt + first_frame | video (+depth) | |
| | `vla` | prompt + first_frame + robot | actions | |
| | `a2v` | first_frame + actions + robot | video | |
| | `v2a` | video + robot | actions | |
| | `uncond` | first_frame + robot | video + actions | |
| |
| |
| |
| |
| ## Evaluation Results |
| |
| Thinker-WM is evaluated on embodied world-model benchmarks covering real-robot and simulated environments, together with controlled evaluations of action conditioning and geometric prediction. |
| |
| ### Main Benchmarks |
| |
| | Benchmark | Setting | Metric | Thinker-WM | Rank | |
| | ----------------- | --------------------------- | ---------: | ---------: | ------: | |
| | **EWMBench** | Embodied world model | Overall ↑ | **4.66** | **1st** | |
| | **WorldOlympiad** | Robotics sub-leaderboard | Overall ↑ | **0.739** | **1st** | |
| | **WorldArena** | Track-1, 12-metric protocol | EWMScore ↑ | **71.73** | **1st** | |
| |
| ### Detailed Results |
| |
| #### EWMBench |
| |
| | Model | SceneC ↑ | HSD ↑ | Dyn ↑ | nDTW ↑ | Diversity ↑ | BLEU ↑ | CLIP ↑ | Logics ↑ | Overall ↑ | |
| | --------------- | ---------: | ---------: | ---------: | ---------: | ----------: | ---------: | ---------: | ---------: | --------: | |
| | Veo3 | 0.8415 | 0.2130 | 0.1932 | 0.1613 | 0.0221 | 0.2139 | 0.8965 | 0.9474 | 3.49 | |
| | Wan2.6 | 0.6712 | 0.2034 | 0.0900 | 0.1715 | 0.0502 | 0.1616 | 0.8743 | 1.0000 | 3.22 | |
| | Kling26 | 0.8211 | 0.3272 | 0.1822 | 0.3423 | 0.0173 | 0.2591 | 0.9014 | 1.0000 | 3.85 | |
| | LTX-2 | 0.7850 | 0.2076 | 0.1283 | 0.2443 | 0.0120 | 0.1425 | 0.8869 | 0.5000 | 3.01 | |
| | Sora2 | 0.8526 | 0.2807 | 0.3494 | 0.2754 | 0.0314 | 0.2466 | 0.9100 | 0.9474 | 3.89 | |
| | Cosmos | 0.7963 | 0.2500 | 0.2052 | 0.2533 | 0.0803 | 0.1230 | 0.8458 | 0.7333 | 3.29 | |
| | GigaWorld | 0.8707 | 0.3050 | 0.0849 | 0.2783 | 0.0278 | 0.2048 | 0.8873 | 0.9000 | 3.56 | |
| | LVP | 0.8795 | 0.4248 | 0.0433 | 0.6226 | 0.0093 | 0.2179 | 0.8995 | 0.9524 | 4.05 | |
| | Vidar | 0.7341 | 0.1877 | 0.1520 | 0.1769 | 0.0653 | 0.1607 | 0.8821 | 0.9411 | 3.30 | |
| | WoW | 0.8866 | 0.2494 | 0.0529 | 0.2566 | 0.0266 | 0.1932 | 0.9001 | 0.9524 | 3.52 | |
| | Qwen-RobotWorld | 0.9142 | 0.5660 | 0.3429 | 0.6708 | 0.0114 | 0.2079 | 0.8834 | 1.0000 | 4.60 | |
| | **Thinker-WM** | **0.8877** | **0.5638** | **0.3630** | **0.7102** | **0.0117** | **0.2691** | **0.9020** | **0.9524** | **4.66** | |
| |
| #### [WorldOlympiad](https://alibaba-damo-academy.github.io/WorldOlympiad/#leaderboard) |
| |
| **Robotics leaderboard** |
| |
| | Model | Physical ↑ | 3D Cons. ↑ | CLIP ↑ | Interact. ↑ | All ↑ | |
| | ------------------ | ---------: | ---------: | --------: | ----------: | --------: | |
| | Uranus | 0.925 | 0.511 | 0.304 | 0.730 | 0.722 | |
| | Cosmos-Predict-2.5 | 0.937 | 0.479 | 0.321 | 0.721 | 0.712 | |
| | LingBot-World | 0.949 | 0.393 | 0.314 | 0.710 | 0.684 | |
| | Rolling Forcing | 0.870 | 0.389 | 0.329 | 0.566 | 0.608 | |
| | LongLive | 0.857 | 0.472 | 0.327 | 0.470 | 0.600 | |
| | Yume-1.5 | 0.851 | 0.288 | 0.312 | 0.624 | 0.588 | |
| | WoW | 0.787 | 0.272 | 0.288 | 0.447 | 0.502 | |
| | **Thinker-WM** | **0.923** | **0.531** | **0.305** | **0.764** | **0.739** | |
| |
| #### WorldArena |
| |
| **Track-1, 12-metric evaluation protocol** |
| |
| | Model | 12-Metric EWMScore ↑ | |
| | ------------------------- | -------------------: | |
| | DexWorldEngine | 71.01 | |
| | BWM-Fast | 70.97 | |
| | UNIS | 70.79 | |
| | MotuBrain | 70.18 | |
| | Wan2.6 | 67.43 | |
| | Cosmos-Predict2.5(Action) | 64.40 | |
| | **Thinker-WM** | **71.73** | |
| |
| For the WorldArena Track-1 evaluation, the reported EWMScore is the unweighted mean of 12 metrics covering visual quality, content consistency, physics adherence, 3D accuracy, and controllability. Dynamic Degree, Flow Score, and Motion Smoothness are excluded because of their sensitivity to frame rate and temporal sampling density. |
| |
| ### Additional Evaluations |
| |
| | Evaluation | Metric | Result | |
| | ----------------------------------------- | -------------- | -----------: | |
| | Action-conditioned generation on EWMBench | Overall ↑ | **5.03** | |
| | Action-conditioned generation on EWMBench | HSD ↑ | **0.7369** | |
| | Action-conditioned generation on EWMBench | nDTW ↑ | **0.8736** | |
| | Generated depth on AgiBot | AbsRel ↓ | **0.1360** | |
| | Generated depth on AgiBot | RMSE ↓ | **0.1790 m** | |
| | Closed-loop control on RoboTwin | Success Rate ↑ | **75.0%** | |
| | Motion Probe on held-out AgiBot windows | Action MSE ↓ | **0.116** | |
| |
| |
| |
| |
| ## Citation |
| |
| If you find Thinker-WM useful, please cite: |
| |
| ```bibtex |
| @article{thinkerwm2026, |
| title={Thinker-WM: A Unified World Model with Shared Multimodal Representations for Embodied Intelligence}, |
| author={Thinker Team, Ubtech}, |
| journal={arXiv preprint}, |
| year={2026}, |
| eprint={2610.18447}, |
| archivePrefix={arXiv}, |
| url={https://arxiv.org/abs/2610.18447}, |
| } |
| ``` |
| |
| ## Contributors |
| |
| **Core Authors**: Chen Liao, Jiyuan Wang, Xiongwei Zhao |
| |
| **Authors**: Peng Gao, Yiming Ma, Daqin Luo, Xiangming Xu |
| |
| **Project Lead**: Baiyu Pan |
| |
| **Advisors**: Jichao Jiao, Hailin Shi |
