GR00T Whole-Body Control is a unified platform for developing and deploying advanced humanoid robot controllers.
The platform addresses the challenge of creating sophisticated control systems for humanoid robots by providing decoupled whole-body control models that separate lower-body and upper-body control strategies. The decoupled approach uses reinforcement learning for lower-body locomotion while applying inverse kinematics for upper-body manipulation, allowing each subsystem to be optimized independently. The codebase includes model checkpoints, training scripts, and evaluation tools to support both development and real-world deployment of these controllers.
The tool is suited for robotics researchers and engineers working with humanoid platforms who need production-ready control architectures. It provides implementations of multiple controller variants including the decoupled WBC used in NVIDIA's GR00T models, the GEAR-SONIC series for generalist whole-body control, and MotionBricks for real-time latent generative motion control. Teams building on humanoid robot platforms will find value in the pre-trained checkpoints and modular design that allows selective use of lower-body and upper-body components.
The project maintains active development with regular updates to support new humanoid robot versions and controller architectures. The codebase is written in Python and backed by comprehensive documentation including whitepapers and project pages for each controller variant. The platform's focus on providing both research implementations and deployment-ready code suggests ongoing refinement of the control algorithms and their practical application to real robotic systems.
