Combining analysis, imitation, and experience-based learning to acquire a concept of reachability in robot mobile manipulation

Abstract

Analytic modeling, imitation, and experience-based learning are three approaches that enable robots to acquire models of their morphology and skills. In this paper, we combine these three approaches to efficiently gather training data to learn a model of reachability for a typical mobile manipulation task: approaching a worksurface in order to grasp an object. The core of the approach is experience-based learning. For more effective exploration, we use capability maps as analytic models of the robot's dexterity to constrain the area in which the robot gathers training data. Furthermore, we acquire a human model of reachability from human motion data and use it to bias exploration. The acquired training data is used to learn Action-Related Places. In an empirical evaluation we demonstrate that combining the three approaches enables the robot to acquire accurate models with far less data than with our previous exploration strategy.