A Correspondence Metric for Imitation

Abstract-- Learning by imitation and learning from demonstration have received considerable attention in robotics. However, very little research has been in the direction of providing a quantitative and objective measure of the quality of imitation. In this paper we present initial work in this direction. We make use of a graph theoretic encoding for the configuration of a robot/agent, represented as a kinematic chain, termed a pose. We provide an algorithmic measure of distance between any two poses, provided their underlying graphs are homeomorphic. This distance measure imposes a pseudometric on the class of poses of homeomorphic graphs. We demonstrate this metric on various configurations of the Sony AIBO dog, a human, and a synthetic dolphin-like skeleton. In order to illustrate the metric, we embed these onto a two dimensional Euclidean space by employing Multi-dimensional scaling (MDS).

• Type: Conference
• Year: 2004
• Where: AAAI
• Authors: R. Amit, Maja J. Mataric
• Keywords: AAAI 2004, Considerable Attention, Homeomorphic Graphs, Intelligent Agents, Objective Measure