Towards an Embodied, Physically-Consistent Generative World Model

WoW (World-Omniscient World Model) is a 14B-parameter generative world model trained on 2 million real-world robot interaction trajectories. It is designed for physically consistent imagination, reasoning, and action in robotics.

• We've updated the WoW-WAN2.1 Gradio ‘demo’ in the demo folder. A more user-friendly inference interface is now available. Just download the checkpoint and run the code to try it out!
• We release the DiT postraining checkpoints of WoW，includes DiT-2B based on Cosmos-Predict2, DiT-7B based on Cosmos-Predict1, and DiT-14B based on the Wan2.1

For Wan based models, follow the demo/README.md (recommand)

For Cosmos based DiT models:

pip install -r dit_models/wow-dit-2b/requires.txt

• Run the Wan demo:

python demo/wan_infer_demo.py 

• Example: Inference with 2B DiT model

python scripts/infer_wow_dit_2b.py --help

• For 7B model:

python scripts/infer_wow_dit_7b.py --help

• For custom input or parameters, please refer to comments in the corresponding demo scripts.

We have released the following models and datasets on Hugging Face:

WoW is being released in phases to promote transparency and collaboration. Below is the current open-source progress:

• Paper released on arXiv:2509.22642
• Project website launched: wow-world-model.github.io

• Model Weights (2B, 7B, 14B WoW-DiT)
• Inference Scripts & Colab Demo
• Baseline Inverse Dynamics Model
• Baseline Model Weights (SVD, CogVideoX, Cosmos1&2)

• 3D-Flow-Mask Inverse Dynamics Model
• Training Pipeline
• SOPHIA Framework Code
• WoWBench benchmark design & evaluation metrics

• Continuous release of real/simulated trajectory data
• Expansion to multimodal inputs (audio, tactile, etc.)
• Universal fine-tuning API for downstream tasks
• Community challenges and leaderboard

• Submit issues or feature requests
• Improve code or documentation
• Run experiments and submit results
• Contribute real-world robot data

• Email: [email protected]
• Project website: wow-world-model.github.io

If you use WoW in your research, please cite:

@article{chi2025wow,
  title={WoW: Towards a World omniscient World model Through Embodied Interaction},
  author={Chi, Xiaowei and Jia, Peidong and Fan, Chun-Kai and Ju, Xiaozhu and Mi, Weishi and Qin, Zhiyuan and Zhang, Kevin  and Tian, Wanxin and Ge, Kuangzhi and Li, Hao and others},
  journal={arXiv preprint arXiv:2509.22642},
  year={2025}
}