GEAR-SONIC

NVIDIA's whole-body controller for the Unitree G1, trained to track any human motion

License

Apache 2.0

Split licensing. Code: Apache-2.0. All released model weights: NVIDIA Open Model License, which allows commercial use and derivative models but requires attribution and compliance with NVIDIA's Trustworthy AI terms. BONES-SEED, the main training dataset: BONES-SEED licence, granted only to non-profit academic users and to companies under $1M annual revenue; others need a commercial licence from Bones Studio. Checked 2026-10-16.

Overview

GEAR-SONIC is a whole-body controller for humanoid robots from NVIDIA's GEAR lab, released with its training code, C++ deployment stack and checkpoints in the GR00T-WholeBodyControl repository. Rather than training one policy per skill, it trains a single policy to track human motion, using motion capture as the supervision, and scales that up: the paper, published in Science Robotics in 2026, grows the network from 1.2M to 42M parameters on more than 100M frames from 700 hours of motion capture. The result walks, runs, kneels, crawls and gets up on a Unitree G1. Every input, whether a kinematic planner steered by keyboard or gamepad, a VR headset or a vision-language-action model, is turned into the same motion tokens, so one controller serves all of them. Training runs in Isaac Lab, simulation checks run in MuJoCo, and deployment runs on the G1's onboard Jetson Orin with TensorRT. Code is Apache-2.0; the weights are under the NVIDIA Open Model License.

SONIC treats humanoid control as motion tracking. During training, the policy is shown a short window of a reference motion, a human clip retargeted to the robot, and is rewarded for following it in simulation. Because the targets come from motion capture rather than hand-written rewards, adding a skill means adding data. Several encoders map different kinds of reference, robot joint trajectories, SMPL human poses or VR teleoperation keypoints, into a shared 64-dimensional token, and one decoder turns that token into joint targets for the G1's 29 joints at 50 Hz.

The repository covers the whole path. The gear_sonic package trains with PPO in Isaac Lab, either from scratch on the BONES-SEED motion dataset (142,000 motions, about 288 hours, already retargeted to the G1) or by fine-tuning the released checkpoint. The authors recommend 64 or more GPUs for training from scratch; a single 8-GPU node works but is much slower. The gear_sonic_deploy package is a C++ stack that runs the exported ONNX models through TensorRT on a desktop GPU or on the robot's Jetson AGX Orin or Thor, with a MuJoCo sim-to-sim mode to try a policy before it touches hardware. A guide covers adding another robot, using the Unitree H2 as its worked example.

On top of the controller sit the ways to drive it. A kinematic planner generates walking, running, crouching and crawling in several styles from keyboard or gamepad commands. A PICO VR headset, or NVIDIA's IsaacTeleop, streams the operator's whole body for teleoperation and data collection. A documented workflow collects teleoperated demonstrations, fine-tunes GR00T N1.7 on them and runs the result with SONIC underneath, so the language model decides what to do and SONIC handles the balance.

Three G1 checkpoints are released on Hugging Face: the original controller, a low-latency teleoperation model that looks 80 ms ahead instead of 200 ms, and SONIC v1.1, trained with wrist-pose augmentation for steadier teleoperation. The same repository also hosts NVIDIA's earlier Decoupled WBC, used under GR00T N1.5 and N1.6, and a preview of MotionBricks, a real-time motion generator.

Why it matters

Most humanoid controllers are small policies trained for a handful of behaviours, each with its own reward terms. SONIC is the clearest open evidence so far that the scaling recipe behind language and vision models also works for low-level humanoid control: more motion data, a larger network and more compute gave steadily better tracking and better handling of motions the policy had never seen. Because it is released end to end, with training code, a real-robot deployment stack and teleoperation tooling, a lab with a Unitree G1 can use it as the motor layer under its own planner or VLA instead of building locomotion from scratch.

Problem it solves

Getting a humanoid to move naturally across many behaviours normally means designing a separate controller or reward for each skill, and those controllers do not share an interface, so a planner, a teleoperator and a learned policy each need their own. SONIC replaces them with one motion-tracking policy trained on large-scale motion capture, and gives every kind of command one token interface to that policy.

Key takeaways

• One policy tracks human motion for the Unitree G1's 29 joints at 50 Hz, covering walking, running, kneeling, crawling and getting up
• The paper scales the network from 1.2M to 42M parameters on 700 hours of motion capture and finds tracking keeps improving with data and compute
• A kinematic planner, a VR headset and a VLA model all drive it through the same 64-dimensional motion tokens
• Released end to end: Isaac Lab training, a C++ TensorRT deployment stack for the G1's Jetson, MuJoCo sim-to-sim checks and VR teleoperation
• Three checkpoints on Hugging Face: the original controller, a low-latency teleoperation model and SONIC v1.1
• Code is Apache-2.0 and the weights are under the NVIDIA Open Model License, which allows commercial use; the BONES-SEED training data has its own, narrower licence

Who it is for

• Labs and companies with a Unitree G1 that want a general whole-body controller rather than per-skill policies
• Researchers training humanoid controllers from motion capture who want a scaled baseline to start from
• Teams collecting humanoid teleoperation data or running a VLA that needs a low-level motor layer
• Researchers porting a motion-tracking controller to another humanoid

Limitations

• Released checkpoints are for the Unitree G1 only; other robots need retargeted motion data and a full training run, documented with the H2 as the example
• Training from scratch is a cluster job: the authors recommend 64 or more GPUs, and the Isaac Lab training environment needs a CUDA 12 GPU (L40 recommended)
• Deployment needs TensorRT at a specific version per platform and, on the G1's onboard Orin, a reflash to JetPack 6
• The BONES-SEED training data is licensed only to academic users and companies with less than $1M in annual revenue; anyone else needs a commercial licence to retrain on it
• Real-robot use carries the usual humanoid risks: the documentation asks for a clear safety zone and an operator ready to stop the robot