Robotics
Researchers Scale Humanoid Robots to 42M Parameters, Achieving Natural Whole-Body Movement
A new foundation model trained on vast motion-capture datasets demonstrates that larger neural networks and more diverse training data unlock more natural and robust control for humanoid robots.
Humanoid robot control has lagged behind the scaling successes seen in large language models. While foundation models in other domains have grown to billions of parameters on thousands of GPUs, neural controllers for humanoid systems have remained modest, handling only narrow task sets and trained on minimal compute. A new research paper positions motion tracking as a path to close that gap.
Scaling for Better Control#
Researchers built SONIC, a generalist humanoid controller that improves performance by growing along three dimensions: network size (from 1.2 million to 42 million parameters), training data volume (over 100 million frames drawn from 700 hours of motion-capture recordings), and computational resources (21,000 GPU hours). The key insight is that motion tracking—using dense supervision from real human motion data—naturally provides training signal without requiring manual reward engineering.
The approach treats existing motion-capture libraries as a scalable source of learning material, allowing the model to acquire human movement patterns directly from observation.
Practical Applications Beyond Motion Capture#
The researchers demonstrated downstream utility by integrating the motion-tracking foundation with a real-time kinematic planner, which bridges tracking to practical tasks like autonomous navigation. A unified token space within the policy also enables virtual reality teleoperation and compatibility with vision-language-action (VLA) models—all operating under a single learned controller.
In an autonomous demonstration, the VLA-driven system coordinated full-body loco-manipulation, executing tasks that required simultaneous planning of hand and foot placement.
Validation and Generalization#
Performance gains held steady as compute and data diversity increased, suggesting the approach avoids performance plateaus. The trained policies also generalized to motion sequences not seen during training, indicating that the scaling benefits establish motion tracking as a practical foundation for future humanoid systems.
