The author presents the DuSkill framework, utilizing offline skill diffusion with a guided diffusion model to enhance policy learning for tasks in different domains.
Offline skill diffusion framework enhances policy learning for tasks in different domains by generating diverse skills.
DuSkill framework enhances policy learning by generating diverse skills for different domains.
EXTRACT enables robots to efficiently learn new tasks by unsupervisedly extracting a discrete set of semantically meaningful, parameterized skills from offline data, which can be effectively leveraged for downstream reinforcement learning.