Shape from Motion Decomposition as a Learning Approach for Autonomous Agents

Abstract

This paper explores Shape from Motion Decomposition as a learning tool for autonomous agents. Shape from Motion is a process through which an agent learns the "shape" of some interaction with the world by imparting motion through some subspace of the world. The technique applies singular value decomposition to observations of the motion to extract the eigenvectors. We show how shape from motion applied to a fingertip force sensor "learns" a more precise calibration matrix with less effort than traditional least squares approaches. We also demonstrate primordial learning on a primitive "infant" mobile robot.