furonghuang-lab/mu0-policy

μ₀ as Policy — RoboCasa Checkpoints

Downstream control-policy checkpoints trained on top of the frozen μ₀ trace world model using RoboCasa demonstrations. The action expert learns flow-matching action chunks conditioned on μ₀'s predicted 3D-trace features.

• 🌐 Project page: https://mu0-wm.github.io
• 📄 Paper: arXiv:2606.13769
• 💻 Code: https://github.com/Yoonkyo/mu0

Checkpoints

mu0_policy.tar contains two policies:

Checkpoint	Tasks	Steps
atomic8_60k_s42	8-task atomic subset	60k
atomic65_all_150k_s42	all 65 atomic tasks	150k

Download

wget https://huggingface.co/furonghuang-lab/mu0-policy/resolve/main/mu0_policy.tar
tar -xf mu0_policy.tar -C outputs/robocasa/

Resulting tree:

outputs/robocasa/mu0_policy/
├── atomic8_60k_s42/final/checkpoint.pt
└── atomic65_all_150k_s42/final/checkpoint.pt

Evaluating these also requires the μ₀ world-model release (furonghuang-lab/mu0) for the trace checkpoint and normalization stats. Full training and evaluation instructions are in docs/release/TRAINING_POLICY.md.

Citation

@article{lee2026mu0,
  title={$\mu_0$: A Scalable 3D Interaction-Trace World Model},
  author={Lee, Seungjae and Jung, Yoonkyo and Lee, Jusuk and Shin, Jonghun and
          Shahidzadeh, Amir Hossein and Lee, Yao-Chih and Kim, H. Jin and
          Huang, Jia-Bin and Huang, Furong},
  journal={arXiv preprint arXiv:2606.13769},
  year={2026}
}