Yuzhu Sun

Human demonstrations, even suboptimal ones, dramatically boost robot adaptation in dynamic environments when integrated via a real-time digital twin and a dual-actor RL framework, outperforming direct imitation loss methods.