GR00T-WholeBodyControl
Mirror of NVIDIA's GR00T Whole-Body Control codebase: training, evaluation and deployment code for humanoid whole-body controllers.
What it is for
The repo copies the GR00T-WholeBodyControl code repo. It holds the Decoupled WBC controller used in GR00T N1.5 and N1.6 (RL for the lower body, IK for the upper body), the GEAR-SONIC training stack and C++ inference stack, and a preview of MotionBricks, a real-time latent generative model for motion control. It also holds VR teleoperation, data collection and kinematic planner tooling for the Unitree G1 robot. The released GEAR-SONIC checkpoints live in NVIDIA's own model repo.
Provenance
- Upstream source: NVlabs/GR00T-WholeBodyControl, with documentation.
- Checkpoints: nvidia/GEAR-SONIC. Whitepaper: GEAR-SONIC, cited as arXiv 2511.07820. MotionBricks: project page.
Use
hf download chibifire/GR00T-WholeBodyControl --local-dir GR00T-WholeBodyControl
cd GR00T-WholeBodyControl
python check_environment.py
Each use case (training, MuJoCo simulation, VR teleoperation, data collection, robot deployment) has its own environment and install script. See the upstream README and the documentation for full setup and usage.
Licence
Apache 2.0, the licence of the GR00T-WholeBodyControl code repo. Terms: LICENSE. The weights also carry the NVIDIA Open Model License. The LICENSE file holds both texts.
- Downloads last month
- 9