SONIC: Supersizing motion tracking for natural humanoid whole-body control

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Foundation models have scaled significantly in size and capability, but this has not translated to humanoid control.
• Current humanoid controllers are limited in size, behavior scope, and computational resources.

Purpose of the Study:

• To demonstrate that scaling model capacity, data, and compute can yield a generalist humanoid controller.
• To establish motion tracking as a scalable task for acquiring human motion priors without manual reward engineering.

Main Methods:

• Trained a foundation model for motion tracking by scaling network size (1.2-42M parameters), dataset volume (>100M frames), and compute (21,000 GPU hours).
• Leveraged dense supervision from diverse motion-capture data.
• Developed a real-time kinematic planner and a unified token space for downstream applications.

Main Results:

• Achieved a generalist humanoid controller capable of natural, robust whole-body movements.
• Demonstrated downstream utility via a kinematic planner for navigation and a unified policy for VR teleoperation and vision-language-action (VLA) models.
• Showcased autonomous VLA-driven whole-body locomanipulation with coordinated hand and foot placement.

Conclusions:

• Scaling motion tracking improves performance steadily with compute and data diversity.
• Learned policies generalize to unseen motions, establishing large-scale motion tracking as a foundation for humanoid control.
• This approach enables natural, interactive, and versatile humanoid control systems.