Agent Skill Learning

Agent skill learning studies how agents acquire, represent, retrieve, and combine reusable behaviors, from motor primitives in robotics to tool-use procedures in language-model agents. Common approaches include hierarchical reinforcement learning and latent skill discovery, imitation from demonstrations, and external skill libraries refined from successful and failed trajectories using verification or reinforcement learning. Work is tested in settings such as LIBERO and ALFWorld for embodied tasks, WebArena for web use, and benchmarks that measure continual learning and skill transfer. A central challenge is making skills reliable and useful on held-out tasks: more updates or broader libraries can improve coverage but also cause drift, costly retrieval, or unsafe behavior, so evaluation increasingly examines transfer, execution quality, and lifecycle safety.