GigaBrain-WBC-0.5

Behavior world model for robust humanoid whole-body tracking with environment interaction. A single policy covers motion tracking, fall recovery, and terrain interaction.

• Paper: GigaBrain-WBC-0.5: A Behavior World Model for Robust Humanoid Whole-Body Tracking with Environment Interaction
• Project page: https://shepherd1226.github.io/gigabrain-wbc-0.5/
• Code and documentation: https://github.com/open-gigaai/giga-brain-wbc
• Training and benchmark data: https://huggingface.co/datasets/open-gigaai/GigaBrain-WBC-0.5-TerrainData

Contents

checkpoints/unitree_g1_model/
  checkpoint.pt          training checkpoint, for fine-tuning
  model_config.yaml      observation configuration of this checkpoint
  exported/
    policy_encoder.onnx
    policy_decoder.onnx
checkpoints/maker_l01_model/
  amp_terrain.onnx       AMP locomotion policy for the base walking states
  policy_encoder.onnx    GIGA-WBC reference encoder
  policy_decoder.onnx    GIGA-WBC action decoder

Usage

The evaluation runner and the deployment controller in the code repository load the ONNX encoder/decoder pair directly. See its documentation for the full evaluation, simulation, and real-robot workflows.

Licence

Released under Apache-2.0, the same licence as the code repository.

Citation

@article{cheng2026gigabrainwbc,
  title   = {GigaBrain-WBC-0.5: A Behavior World Model for Robust
             Humanoid Whole-Body Tracking with Environment Interaction},
  author  = {Cheng, Ziyang and Tang, Tianshu and Lan, Jinxin and Chen, Xinze and
             Gong, Yuhan and Liu, Zhichao and Wu, Changzhong and Mao, Yahao and
             Deng, Zongyan and Ma, Mingxuan and Xi, Huasen and Liu, Yilong and
             Wu, Yutong and Wang, Xiaofeng and Zhang, Borui and Yu, Bingyao and
             Wang, Yang and Ye, Yun and Huang, Guan and Jin, Xiaojie and Zhu,
             Zheng and Lu, Jiwen},
  journal = {arXiv preprint arXiv:2608.18234},
  year    = {2026}
}