Learning a Universal Humanoid Motion Representation for Versatile Physics-Based Control

A universal motion representation that encompasses a comprehensive range of motor skills for physics-based humanoid control, learned by distilling from a large-scale motion imitator and using a learnable prior.