The question of ‘what to imitate’: inferring goals and intentions from demonstrations

Abstract

Getting a robot to imitate is a difficult task. Currently, researchers are tackling difficult issues such as the 'correspondence problem' (Nehaniv and Dautenhahn, 2002) of how to map different aspects of a demonstration to one's own imitative response, how to decide whom, when and what to imitate, and how to evaluate success (Dautenhahn and Nehaniv, 2002a, b). From our perspective, the most interesting of these issues is the question of what to imitate, that is, which aspects of a demonstration robots should copy. This is because in many situations, instead of copying others' actions and their results on the environment exactly, it is often more useful to extract and adopt the goal of the demonstrator's actions, and only copy the actions and results of the demonstrator if they are intended and relevant or essential to the task.