Presentation

· Contributors · Organizations · Search Program · My Schedule · Happening Now · Maps

SuperPADL: Scaling Language-directed Physics-based Control With Progressive Supervised Distillation

DescriptionWe present a framework for scaling physics-based text-to-motion models to datasets containing thousands of motions. Our approach being with using reinforcement learning to train a large number of expert tracking policies. We then progressively distill these experts into larger, more capable networks using supervised distillation objectives.

Authors

University of Toronto

Xue Bin Peng

NVIDIA

Simon Fraser University