μ₀: A Scalable 3D Interaction-Trace World Model
μ₀ is a trace world model that operates in 3D space rather than pixel or action space. Given an image, a language instruction, and a short history of keypoints, μ₀ predicts the future 3D traces of semantic interaction points — objects, tools, hands, and contact regions — i.e. what must move to accomplish a task. Because these traces are embodiment-agnostic, μ₀ can be pretrained from video alone and transferred across robot embodiments.
- 🌐 Project page: https://mu0-wm.github.io
- 📄 Paper: arXiv:2606.13769
- 💻 Code: https://github.com/Yoonkyo/mu0
This repository hosts the release artifacts for μ₀ — the trained world-model
checkpoint, normalization stats, and the evaluation test set. The downstream
policy checkpoints live at
furonghuang-lab/mu0-policy.
Contents
| Path | Description |
|---|---|
final_ckpt/ | Released world-model checkpoint (config.json, meta.json, model.safetensors) |
normalizer_stats.json | Delta/depth normalization stats used in training and evaluation |
test_set.tar | Evaluation episodes (extracts to test_set/) |
Download
pip install -U "huggingface_hub[hf_transfer]"
hf download furonghuang-lab/mu0 --local-dir mu0_release
cd mu0_release && tar -xf test_set.tar # only needed for evaluation
You can also fetch individual files via their resolve URLs, e.g.
https://huggingface.co/furonghuang-lab/mu0/resolve/main/final_ckpt/model.safetensors.
See docs/release/EVALUATION.md
in the code repo for usage.
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}
}
Acknowledgements
μ₀'s training, dataset, and policy infrastructure are built on LeRobot (Apache-2.0) by Hugging Face. If you use these artifacts, please also cite LeRobot.