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

Technical Analysis “SONIC” shows that scaling up model capacity, data, and compute yields a generalist humanoid controller capable of creating natural and robust whole-body movements. SONIC proposes a foundation model for motion tracking by scaling along three axes: network size (from 1.2M to 42M parameters), dataset volume (over 100M frames, 700 hours of high-quality motion data), and compute (9k GPU hours). High-level Analogy: Imagine you have a brilliant human dancer who has watched millions of hours of all types of human movements – from walking and running to complex martial arts and intricate dances. This dancer is so skilled that you can simply show them a video, tell them what to do, play some music, or even move your own body, and they can immediately understand the essence of the motion and flawlessly replicate it, adapting on the fly. They don't need new lessons for every single step or style; they have an innate 'universal dance language' that lets them interpret any instruction and translate it into smooth, natural movement. SONIC is like training this master dancer for a robot, allowing it to interpret diverse human movement commands and perform them with incredible fluidity and adaptability. Motivation of the Work Current humanoid robots often struggle to perform natural and varied movements. Typically, their controllers are small, specialized programs that are good at one specific task, like walking forward. To teach a robot a new skill, like dancing or…