A comparison of distance measures for learning nonparametric motor skill libraries

Abstract

Autonomous robots that interact with the environment to learn new motor skills need to continuously memorize and compare fresh knowledge with past experience. Traditional approaches assume that experts label skills. In this paper, we introduce a new framework for autonomously learning a nonparametric skill library. Crucial design choices are the space in which the motor skills are represented, a distance measure to evaluate the similarity between known skills and new observations and a clustering policy. These aspects determine the structure and amount of skills learned by the library. In this paper, we use a probabilistic skill representation to support a large variety of measures as well as useful library features. Clustering is done via a decision tree. We evaluated the learning process with respect to 19 distance measures used to separate four manipulation skills on a KUKA LWR arm. The results suggest comparisons of trajectories in the end effector space, and that the use of measures for distributions depends on the amount of available data.