Learning from demonstration is an approach in which an agent acquires behaviour by observing examples performed by a teacher. It is used to bootstrap policies in robotics and control without hand-specified reward functions.
Semantic Classification
Content
- Learning from demonstration, also termed programming by demonstration, lets a system infer a task from recorded expert trajectories rather than explicit instructions. The teacher may be a human operator, a scripted controller or a higher-quality policy.
- Approaches range from directly imitating the observed actions to inferring the underlying reward and then optimising it. The method is common in robotics where defining a reward by hand is difficult but demonstrations are easy to provide.