Imitation learning from multiple demonstrators using global vision

Imitation learning enables a learner to expand its own skill set with behaviours that it observes from others. Most imitation learning systems learn from a single class of demonstrators, and often only a single demonstrator. Such approaches are limited, however: in the real world, people have varying levels of skills and different approaches to solving problems, and learning from only one demonstrator would be a very limited perspective. In the context of robots, very different physiologies make learning from many types of demonstrators equally important. A wheeled robot may watch a humanoid perform a task, for example, and yet not be able to perfectly approximate its movements (e.g. stepping over small obstacles). This thesis describes an approach to learning a task by observing demonstrations performed by multiple heterogeneous robots using global (overhead) vision, incorporating demonstrators that are different in size, physiology (wheeled vs. legged), and skill level. The imitator evaluates demonstrators relative to each other, which gives it the ability to weigh its learning towards the more skilled demonstrators. I assume the imitator has no initial knowledge of the observable effects of its own actions, and begin by training a set of Hidden Markov Models (HMMs) to map observations to