A. L. Thomaz, C. Breazeal.

Abstract

Socially Guided Machine Learning explores the ways in which machine learning can be designed to more fully take advantage of natural human interaction and tutelage. In this article we present a framework for studying the role real-time human interaction plays in training robots to perform new tasks. We have results from an initial user study using our experimental platform, Sophie’s World, to understand how people administer reward and punishment to teach a simulated robot a new task through Reinforcement Learning (RL). Based on this study, we identified three modifications to a standard RL algorithm to make it more amenable to learning from real-time human interaction: an embellished communication channel with both guidance and feedback, transparency behaviors, and responsiveness to errors. We are evaluating these modifications in a follow-up study.