NVIDIA and UT Austin Open-Source Humanoid Model
Researchers from UT Austin and NVIDIA have open-sourced SONIC, a behavior foundation model for humanoid robots. The model enables real-time, whole-body motion for complex loco-manipulation tasks. It supports both teleoperation and Vision-Language-Action (VLA) inference, aiming to accelerate the real-world deployment of generalist humanoid robots.
- The model, named SONIC, was developed by scaling up three key areas: network size (to 42 million parameters), dataset volume (over 700 hours of motion data), and compute (9,000 GPU hours). This large-scale training on human motion-capture data allows the model to learn "human motion priors" without the need for manual reward engineering for each specific skill. - The research is led by UT Austin Associate Professor Dr. Yuke Zhu, who also co-leads NVIDIA's Generalist Embodied Agent Research (GEAR) group, the team behind the project. His work focuses on the intersection of robotics, computer vision, and machine learning to create general-purpose autonomous robots. - SONIC is designed as a "System 1" controller, providing fast, reactive whole-body motor skills, analogous to human reflexes. It can be paired with a "System 2" model, like NVIDIA's GR00T N1.5 Vision-Language-Action (VLA) model, which handles slower, high-level reasoning and planning to decide what tasks to perform. - A core technical innovation is its "universal token space," which allows a single, unified policy to be controlled by diverse inputs without retraining. This enables real-time control from VR teleoperation rigs, human video feeds, text commands, and even music. - To bridge the gap between the learned model and interactive control, SONIC includes a real-time kinematic planner. This allows an operator to use a gamepad or keyboard to guide the robot's locomotion with various styles like running, sneaking, or crawling, with the planner regenerating motions in under 5 milliseconds. - The model's capabilities were demonstrated on the Unitree G1 humanoid robot, showcasing a direct path from simulation to real-world hardware deployment.
