open-gigaai/GigaBrain-WBC-0.5-TerrainData

GigaBrain-WBC-0.5 Terrain Data

Training and benchmark data for GigaBrain-WBC-0.5, a behavior world model for robust humanoid whole-body tracking with environment 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
• Released policy: https://huggingface.co/open-gigaai/GigaBrain-WBC-0.5

Contents

The data ships as two uncompressed tar archives rather than as a loose file tree, because the tree holds ~29k files that average only a few kilobytes each. Both archives unpack to repository-relative paths:

bundles/data.tar             459 MB
  data/data_terrain/
    dataset/        7,332 terrain-interaction motion clips (PKL)
    terrain/        21,996 per-clip terrain annotations and primitives (JSON)
    terrain_types/  8 terrain categories used for sampling
  data/data_exclude/
    exclude_flat.json      12,565 clips excluded from the flat set
    exclude_terrain.json   150 clips excluded from the terrain set
bundles/evaluation_data.tar   29 MB
  evaluation/data/terrain/
    robot/          150 benchmark reference motions (NPZ)
    terrain/        150 matching terrain primitive files (JSON)
bundles/SHA256SUMS                      checksums for both archives

Extract both at the root of the code repository:

tar xf bundles/data.tar
tar xf bundles/evaluation_data.tar

python giga_brain_wbc/scripts/download_release.py --training-data --benchmark-data fetches, verifies and extracts them for you, so a manual download is only needed if you want the archives for something else. The archives are plain tar, not gzip: the payload is PKL and NPZ, which compresses by at most 30%.

dataset/ merges clips from Bones-SEED and from MotionMillion / MotionDecode, retargeted to the Unitree G1 (29 DOF) and re-timed to 30 FPS. Clips are stored as root translation, root rotation, per-joint axis-angle poses, joint positions, and frame rate.

Scope

This release covers terrain interaction only. Flat-ground training motions and the AMASS benchmark motions are not included, because their upstream licences do not allow redistribution. The code repository documents how to prepare both locally.

Licence and attribution

This dataset is a heavily re-processed derivative of the original motions: every clip was filtered, retargeted to the Unitree G1, re-timed, split, and annotated with terrain primitives and per-frame terrain types. Our processing and annotations are released under CC-BY-4.0, while the upstream licences continue to apply to the underlying motions. The source motions come from Bones-SEED, MotionMillion, and MotionDecode, with ChingMu credited as the motion source for the latter. If you use this data, please cite GigaBrain-WBC-0.5 together with the original Bones-SEED, MotionMillion, and MotionDecode papers. We thank the Bones-SEED, MotionMillion, and ChingMu teams for releasing their motion capture data, which made this dataset possible.

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}
}