A Practical Comparison of Three Robot Learning from Demonstration Algorithm

Research on robot Learning from Demonstration has seen significant growth in recent years, but the field has had only limited evaluation of existing algorithms with respect to algorithm usability by naïve users. In this article we present findings from a user study in which we asked non-expert users to use and evaluate three different robot Learning from Demonstration algorithms. The three algorithms selected—Behavior Networks, Interactive Reinforcement Learning, and Confidence Based Autonomy—utilize distinctly different policy learning and demonstration approaches, enabling us to examine a broad spectrum of the field. Participants in the study were asked to teach a simple task to a small humanoid robot in a real world domain. they controlled the robot directly (teleoperation and guidance) instead of providing retroactive feedback for past actions (reward and correction). We present our quantitative findings about: (a) the correlation between the number of user-agent interactions and the performance of the agent and (b) the correlation between agent’s final performance and its perceived accuracy by the participant. Comparatively, the strongest correlation was found in CBA data. We also discuss the possible reasons of our qualitative results. Additionally, we identify common trends and misconceptions that arise when non-experts are asked to use these algorithms, with the aim of informing future Learning from Demonstration approaches. Our results show that users achieved better H.B. Suay ( ) · R. Toris · S. Chernova Worcester Polytechnic Institute, 100 Institute Dr., Worcester, MA 01609, USA e-mail: [email protected] R. Toris e-mail: [email protected] S. Chernova e-mail: [email protected] performance in teaching the task using the CBA algorithm, whereas the Interactive Reinforcement Learning algorithm modeled user behavior most accurately.