Learning by observation with mobile robots: a computational approach

Abstract

We present a computational approach to learning by observation (LBO) that allows users to program mobile robots by demonstrating a task. Unlike previous approaches, our system incorporates statistical-learning techniques and concepts from control theory to reduce the amount of domain knowledge needed to infer the intent of the user. To improve the generalization ability of the system, the user can demonstrate the task multiple times. We extract task subgoals from these demonstrations and automatically associate them with objects in the environment. As these objects move, the subgoals are updated accordingly. This gives our system the ability to learn from demonstrations performed in different environments. In this paper, we present the concepts used in our LBO system as well as experimental laboratory results in learning motor-skill tasks.