Journal of Intelligent and Robotic Systems 20: 71–86, 1997. 71 c  1997 Kluwer Academic Publishers. Printed in the Netherlands. A Fuzzy Classifier for Tactile Sensing L. CAROTENUTO, ⋆ D. FAMULARO and P. MURACA Universit´ a della Calabria, Dip. di Elettronica, Informatica e Sistemistica (DEIS), 87036 Rende, Italy, e-mail: luciano.carotenuto@unical.it. G. RAICONI Universit´ a di Salerno, Dip. di Informatica e Applicazioni (DIA), 84081 Baronissi, Italy, e-mail: gianni@udsab.dia.unisa.it. (Received: 10 November 1996; in final form: 5 February 1997) Abstract. In this paper we propose a fuzzy rule-based algorithm for solving classification problems related to tactile sensing. A tactile sensing system is a robotic device which gives an image of contacting objects. While the fine form recognition problem has been widely discussed and several techniques have been proposed for its solution (Bayesian approach or Neural algorithms), less attention has been paid to the problem of deciding whether the object belongs to a particular class or set of objects that share a common feature, also known as tactile primitive. As input data we consider the sum of the normal stresses at the sensing sites. Three levels of classification, hierarchically connected, are analyzed and, for each level, different basis variables with their membership functions are proposed and calibrated using a training procedure. The output is an answer regarding the features of the object at each level and is related to the truth values of the fuzzy classes. The numerical experiences show that, at least for data affected by low noise level, the algorithm has a very high percentage of correct answers. Key words: fuzzy logic, tactile sensing, object classification 1. Introduction A tactile sensor [4–6, 8, 10] is a robotic device able to reproduce the tact charac- teristics. In its basic configuration it can be regarded as an elastic parallelepipedon (Figure 1) in which a regular two-dimensional array of sensing elements is embedded. It is modelled by a linear static input-output relation mapping the displacements (assumed only normal) of the active face, due to a contacting object, into the components of the stress tensor at the sensing sites. Each sensing element then provides signals carrying information about the local value of the stress: according to the specific technology, all the six components of the stress tensor, or only the normal stresses, or a linear combination of the normal stress- es can be recovered, thus forming a stress image of the contacting object. Such data have been used to solve the inverse problem of reconstructing an estimate of the actual displacements: although different approaches (Bayesian, Tikhonov regularization), and different solution algorithms have been proposed [2, 3, 7], the final result is a fine form reconstruction of the object. ⋆ Corresponding author. Tel: +39-984-49 4723; fax: +39-984-49 4713. VTEX(OV) PIPS No.: 134833 MATHKAP JINT1399.tex; 8/08/1997; 12:16; v.7; p.1