X-WBC: A Cross-Embodiment Foundation Model for Humanoid Whole-Body Control

CoRL 2026

Project Page · Paper · Video · Repository

Release status: This repository currently hosts the official project page and release announcement for X-WBC. The implementation, checkpoints, and training assets are still being prepared for public release. Star or watch the repository to follow future updates.

Overview

X-WBC studies how humanoid whole-body control experience can be shared across robots with different morphologies, joint layouts, dynamics, and action spaces. It combines a shared temporal motion backbone with lightweight embodiment-specific modules, allowing multiple humanoids to contribute to one joint training process while retaining robot-specific execution.

The framework supports three human-centered command routes—full human motion, retargeted robot motion, and sparse five-point VR observations—and is evaluated across nine simulated humanoid embodiments. We additionally demonstrate the same sparse-VR interaction format on four physical humanoid platforms.

Highlights

• One shared motion Transformer trained with mixed multi-robot rollouts.
• Robot-specific state encoders and action decoders for heterogeneous bodies.
• Unified command tokens for human motion, robot references, and sparse VR.
• Evaluation on nine simulated embodiments and deployment on four real robots.
• Separate training-distribution studies and frozen external-style evaluation.

Planned Release

• Project page and demo video.
• Official paper link.
• Training and evaluation code.
• Checkpoints and deployment examples.

Code and model weights are being prepared for release.

Citation

If you find X-WBC useful, please cite:

@inproceedings{zhang2026xwbc,
  title     = {X-WBC: A Cross-Embodiment Foundation Model
               for Humanoid Whole-Body Control},
  author    = {Zhang, Juntong and Gu, Chun and Zhang, Li},
  booktitle = {Conference on Robot Learning},
  year      = {2026}
}

Institutions

Tongji University · Fudan University · Shanghai Innovation Institute

License

The license for the future code and model release has not yet been finalized. Unless stated otherwise, no license is granted for unreleased implementation or model artifacts.

Acknowledgements

We thank the authors and maintainers of the humanoid-learning community whose open research has made reproducible whole-body control possible. Detailed third-party acknowledgements and inherited licenses will accompany the code release.