A study of back-error propagation networks for a trainable tactile pattern classifier

Abstract

Current robotic taction applications have largely been limited to presence/absence detection via “simple touch” or the use of force-torque sensors. Although artificial tactile sensors have been developed, advanced industrial applications have not matured due to time and resource constraints in manufacturing environments to develop and maintain application specific pattern recognition algorithms. This difficulty can be resolved if the sensor modules permit users to simply “train” the classification of object impressions specific to the tasks of individual work-cells. A description is given of artificial neural networks (ANNs) consisting of coupled back-error propagation (BEP) networks that perform feature extraction, clustering, and categorization of tactile surface impressions. The network characteristics, including a study using noisy input patterns, are presented. Simulation results indicate that regarding geometry-, size- and activation-constrained grey-scale patterns, a BEP classifier is sensitive to both additive low-amplitude spike noise and additive white Gaussian noise. Most of the misclassifications occur among patterns that differ only by small variations in force gradients. The network's performance gradually improves when noisy patterns are included in the training set, but large training sets will be required to achieve robust performance in the tactile domain.