Learning to predict resistive forces during robotic excavation

Abstract

Few robot tasks require as forceful an interaction with the world as excavation. In order to effectively plan its actions, a robot excavator requires a method to predict the resistive forces experienced as it scoops soil from the terrain. This paper presents methods for a robot to predict resistive forces and to improve its predictions based on experience, using "learning" methods. A simple analytical model of a flat blade translating through soil is extended to account to for phenomena specific to motions of an excavator. In addition, the paper examines how representation and methodology affect prediction performance.