Understanding the embodied teacher : nonverbal cues for sociable robot learning

Abstract

As robots enter the social environments of our workplaces and homes, it will be important for them to be able to learn from natural human teaching behavior. My research seeks to identify simple, non-verbal cues that human teachers naturally provide that are useful for directing the attention of robot learners. I conducted two novel studies that examined the use of embodied cues in human task learning and teaching behavior. These studies motivated the creation of a novel data-gathering system for capturing teaching and learning interactions at very high spatial and temporal resolutions. Through the studies, I observed a number of salient attention-direction cues, the most promising of which were visual perspective, action timing, and spatial scaffolding. In particular, this thesis argues that spatial scaffolding, in which teachers use their bodies to spatially structure the learning environment to direct the attention of the learner, is a highly valuable cue for robotic learning systems. I constructed a number of learning algorithms to evaluate the utility of the identified cues. I situated these learning algorithms within a large architecture for robot cognition, augmented with novel mechanisms for social attention and visual perspective taking. Finally, I evaluated the performance of these learning algorithms in comparison to human learning data, providing quantitative evidence for the utility of the identified cues. As a secondary contribution, this evaluation process supported the construction of a number of demonstrations of the humanoid robot Leonardo learning in novel ways from natural human teaching behavior.