Learning from Demonstration

Learning from Demonstration studies how an agent can acquire skills, goals, or task rules from examples rather than relying only on hand-coded instructions or rewards. Common approaches include behavior cloning of demonstrated actions, inverse reinforcement learning, reinforcement learning guided by demonstrations, and extracting reusable skills or symbolic plans from video and interaction traces. Work spans robot manipulation and navigation as well as GUI workflows, with evaluation in simulation, on real hardware, and on held-out tasks or objects. A central challenge is turning limited, imperfect demonstrations into safe, closed-loop behavior that generalizes across scenes and embodiments; data quality and offline prediction scores do not always reveal whether a policy will succeed at deployment.