A Vision-Based Approach to Imitation Using Heterogeneous Demonstrators

Contents

Imitation learning is a powerful mechanism used by humans and other creatures. In imitation learning, the actions of others form the basis for desirable behaviour, and an imitation learner must be able to recognize the outcomes of the actions of others, understand how these relate to its own abilities, and ultimately duplicate the final outcome of a series of actions. We are interested in supporting this type of learning in general populations of robots, where a two important complications arise. First, physical variation between demonstrator and learner may require the learner to carry out different action(s) from the demonstrator to achieve the same results. Second, since demonstrators’ skills may differ as much as their physiology, agents must be able to compare the demonstrations of a number of different individuals, in order to give greater weight to better demonstrators. Being able to integrate multiple demonstrations from different demonstrators allows a learner to deal with these problems as well as encouraging the creation of more general behaviours, rather than simply mimicking the actions of a single agent with no ability to generalize. In this paper we describe an approach to imitation learning based on global vision, which deals with these problems.