SONIC

NVIDIA's general-purpose humanoid whole-body control foundation model, built by scaling up motion tracking training.

SONIC was released by NVIDIA's GEAR lab (Zhengyi Luo, Linxi Fan, Yuke Zhu, and others) in November 2025, later published in Science Robotics (2026). The premise is that humanoid controllers can also get better simply by scaling up, and motion tracking — having a robot reproduce a piece of human motion in real time — is well suited to that, since motion-capture data already comes with dense supervision and needs no hand-designed reward. The team scaled the network from 1.2 million to 42 million parameters, trained on about 700 hours of motion-capture data (over 100 million frames) using roughly 21,000 GPU-hours, and deployed it on the Unitree G1. Through a unified token interface, the same policy can take motion commands from a joystick plus a real-time motion planner, VR teleoperation, video imitation, or text- and music-driven motion generation, and it can also sit behind a VLA such as GR00T N1.5 to perform whole-body mobile manipulation. The code is open-sourced in the GR00T-WholeBodyControl repository.

ExampleAn operator wearing VR gear tracks only their head and hands; SONIC uses a motion planner to fill in the lower body, controlling the G1 to complete a task like pushing a lawnmower while walking.

Also called

GEAR-SONIC, SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control

Related

Motion Tracking · Learning-Based Whole-Body Control · Unitree G1 · NVIDIA Isaac GR00T N1 · NVIDIA Generalist Embodied Agent Research Lab · GR00T-WholeBodyControl

Sources

SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control (arXiv 2511.07820)
GEAR-SONIC project page

As of

2026-09