Donggeun Lim

1. Introduction Recent advances in physics-based character control have shown remarkable progress in generating realistic human motions using simulated agents and reinforcement learning [YMG*23, BWL*23, HGW*23]. Hierarchical control strategies [MHG*18, MTA*20, KFPM21] offer a scalable approach to obtain and utilize a general motion prior, instead of manually selecting reference data tailored for each scenario. A hierarchical controller is trained in two phases: imitation learning and task learning.