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

SONIC scales humanoid motion tracking by increasing network capacity from 1.2M to 42M parameters and training on 700 hours of motion capture data. The model uses dense supervision from 100M+ frames to learn human motion priors, eliminating the need for manual reward engineering in whole-body control.