Developing affect-aware robot tutors

Abstract

In recent years there has been a renewed enthusiasm for the power of computer systems and digital technology to reinvent education. One-on-one tutoring is a highly effective method for increasing student learning, but the supply of students vastly outpaces the number of available teachers. Computational tutoring systems, such as educational software or interactive robots, could help bridge this gap. One problem faced by all tutors, human or computer, is assessing a student's knowledge: how do you determine what another person knows or doesn't know? Previous algorithmic solutions to this problem include the popular Bayesian Knowledge Tracing algorithm and other inferential methods. However, these methods do not draw on the affective signals that good human teachers use to assess knowledge, such as indications of discomfort, engagement, or frustration. This thesis aims to make understanding affect a central component of a knowledge assessment system, validated on a dataset collected from interactions between children and a robot learning companion. In this thesis I show that (1) children emote more when engaging in an educational task with an embodied social robot, compared to a tablet and (2) these emotional signals improve the quality of knowledge inference made by the system. Together this work establishes both human-centered and algorithmic motivations for further development of robotic systems that tightly integrate affect understanding and complex models of inference with interactive, educational robots.