Learning Social Interaction from the Wizard: A Proposal

Learning from demonstration (LfD) is a widely-used technique for programming robot behavior. Despite its flexibility and applicability to a wide range of tasks, LfD is most typically used to learn non-social behaviors. To our knowledge, no prior work has learned robot behavior from demonstration for interaction that is primarily social in nature. In this paper, we conjecture that LfD can be used to develop robot behavior capable of interacting socially with humans, in scenarios such as a sharing the use of a playful and educational app. These demonstrations, we argue, should be provided within a Wizard-of-Oz paradigm. Thus, we refer to LfD for social human-robot interaction as learning from the wizard (LfW). We describe the corresponding algorithmic and experimental framework and explore challenges in designing LfW systems. Finally, we describe two ongoing projects involving LfD for human-robot interaction, along with a robotic platform and cognitive architecture that are common between