Inferring user intent for learning by observation

Abstract

Despite the numerous advances in human-robot interaction, most development systems still require that users have substantial knowledge of procedural-programming techniques as well as the specific robot system at hand. For the vast majority of the population, this effectively precludes the use of robots in most cases. If robots are to make headway into everyday situations, then users must be able to program robots in a more natural and intuitive manner. This dissertation explores a method of programming robots to automate motor tasks by inferring the intent of users based on demonstrations of a task. In order to understand such a system, we decompose it into simpler components: modeling user subgoal selection and the response of users to different conditions. We have developed a learning algorithm that constructs a statistical model of user subgoal selection based on previous observations. After deriving the algorithm, we provide theoretical guarantees about the model. To validate the theoretical underpinnings of the algorithm, we isolate the performance of modeling user subgoal selection by removing extraneous factors such as sensor noise and environment considerations. To this end, we present experimental results in predicting the waypoints of manipulator-robot programs. We show that the algorithm produces submillimeter prediction errors on real-world data. We hypothesize about the response of users to different conditions with a model of sequenced linear dynamical systems. We first develop the concept that a single dynamical system can represent a simple trajectory using a closed-form least-squares procedure. Since our approach is based on the least-squares principle, it is simple to combine multiple demonstrations, giving the system a better generalization of “what the user would have done” in novel conditions. To represent more complicated trajectories, we segment it and represent each segment by a single dynamical system. These algorithms form the core of a mobile-robot system that learns motor skills by observing users demonstrating a task. From these observations, the system extracts task subgoals and automatically associates them with objects in the environment, so that as the objects move, the subgoals are updated accordingly. This system can learn from multiple demonstrations, as well as demonstrations performed in different environment configurations. In laboratory experiments, we show that the system accurately infers user intent.