Summary

GR00T Whole-Body Control (WBC) provides a unified platform for developing and deploying advanced humanoid robot controllers, targeting engineers and researchers. It offers decoupled WBC models and the GEAR-SONIC generalist controller, enabling natural, generalized whole-body movements and efficient teleoperation for humanoid robots.

How It Works

The core approach centers on the SONIC foundation model, trained on extensive human motion data using motion tracking as a scalable task. This enables a single policy to generate diverse, natural whole-body movements and generalize beyond observed data. It integrates a decoupled WBC (RL for lower body, IK for upper body) and the latest GEAR-SONIC iteration for advanced control. A kinematic planner facilitates real-time locomotion, and VR teleoperation allows direct human-to-robot motion transfer.

Quick Start & Requirements

• Primary install: Clone the repository using Git LFS (git clone https://github.com/NVlabs/GR00T-WholeBodyControl.git), then navigate into the directory and run git lfs pull.
• Prerequisites: Git LFS is required for cloning. Specific Python versions, GPU, or CUDA requirements are not detailed in the setup section but are implied for advanced robotics ML tasks.
• Documentation: Full documentation is available at https://nvlabs.github.io/GR00T-WholeBodyControl/.

Highlighted Details

• GEAR-SONIC: The latest generalist humanoid whole-body controller series, featuring pretrained policy checkpoints and a C++ inference stack.
• SONIC Foundation Model: Achieves natural, whole-body movement and generalization through large-scale human motion data training.
• VR Teleoperation: Real-time whole-body teleoperation via PICO VR headset for intuitive data collection and interactive control, supporting various complex motions.
• BONES-SEED Dataset: Open-sourced large-scale human motion dataset (142K+ motions, ~288 hours) with Unitree G1 MuJoCo-compatible trajectories.
• C++ Inference Stack: Updated for motor error monitoring, temperature reporting, and streamed token input via ZMQ protocol v4.

Maintenance & Community

For questions and feedback, contact the GEAR WBC team at gear-wbc@nvidia.com. Recent updates in March 2026 indicate active development. No specific community channels (e.g., Discord, Slack) or explicit sponsorship details are provided.

Licensing & Compatibility

This project employs dual licensing:

• Source Code: Licensed under Apache License 2.0.
• Model Weights: Licensed under the NVIDIA Open Model License. The NVIDIA Open Model License permits commercial use with attribution and requires adherence to NVIDIA's Trustworthy AI terms.

Limitations & Caveats

Key components such as training scripts, data collection workflows, and preprocessed datasets are not yet open-sourced. The project is under active development, as indicated by recent updates and outstanding TODO items, suggesting certain functionalities may still be evolving or incomplete.