Learning versus analytical approach to contact estimation in assembly tasks with robots

Abstract

In order to automatically execute assembly tasks with robots following a fine-motion plan, it is usually necessary to estimate the current contact situation to select the corresponding robot command. This contact estimation is generally based on configuration and force sensory data. Two methods for this purpose are presented here: a) an analytical method which explicitly takes into account all the uncertainty sources that may affect the task, and which computes the sets of configuration and forces compatible with each possible contact situation; and b) an inductive learning approach based on a backpropagation neural net that uses simulated contact-situation examples for the training phase. The methods are illustrated by the simple assembly task of positioning a block into a corner considering three degrees of freedom. The advantages and disadvantages of each approach are also discussed. Key Words - Robotic Assembly, Fine-motion Planning, Contact Estimation, Learning, Neural Nets. 1. I...