Models of Intention for Human-Robot Interaction

Abstract

As demand for robots grows in non-industrial settings, there is a corresponding need to develop systems that engage with humans on a social level. A key component of this social interaction is the process of inferring humans' intentions from their observed behavior. In this dissertation, we define the "intent recognition problem" in robotics, describing what it is and why it matters. We then describe a series of systems that we have designed and deployed on several physical robot platforms. This includes a system based on hidden Markov models, a system that incorporates contextual information from such sources as a parse of the simplified English Wikipedia, and systems based on Hewitt's actor model. For each system, we describe its design and evaluate its performance in simulation or on one of several physical platforms, including wheeled mobile robots and humanoids. As a result of our evaluations, we describe several features required for the successful operation of an intent recognition system. In particular, we demonstrate through multiple systems the importance of modeling social contextual information in order to interpret and predict human actions in unstructured environments. We also offer guidance on important challenges that are main to be solved as roboticists attempt to build more socially capable systems.