EnRI III Encontro de Robótica Inteligente 14 a 20 de julho de 2006 Campo Grande, MS Anais do XXVI Congresso da SBC Robot Manipulator Control with Neuro-Fuzzy Friction Compensation Sebastião C. P. Gomes, Daniel S. Guimarães Jr., Cláudio M. Diniz, Vitor I. Gervini Fundação Universidade Federal do Rio Grande FURG-DMT Núcleo de Matemática Aplicada e Controle, Av. Itália km 8, 96201-900, Rio Grande, RS dmtscpg@furg.br, danieljr@gmail.com, cmdiniz@furg.br, gervini@ita.br Abstract. The main objective of this paper is to propose a new friction compensation mechanism applied to robotic actuators, and to confirm it through experimental results. Friction is a phenomenon that changes with time and with actuator’s operational conditions. To deal with these parameters variations, it is proposed a neuro-fuzzy algorithm for friction identification and compensation. A Neural Network (NN) was trained off line. The NN output (compensation friction torque) is multiplied by a gain, obtained with a Fuzzy inference algorithm, to deal with friction parameters variations and to adjust the compensation torque. Experimental results showed good performance, indicating that the actuator becomes approximately linear. Resumo. O principal objetivo do presente artigo é propor um novo mecanismo de compensação de atrito aplicado a atuadores robóticos e ainda, testa-lo a partir de resultados experimentais. O atrito é um fenômeno que varia com o tempo e com as condições operacionais do atuador. Para lidar com tais variações, está sendo proposto um algoritmo neuro-fuzzy para a identificação e a compensação do atrito. Uma rede neural artificial (RNA) foi treinada off line. A saída da RNA (torque de compensação do atrito) é multiplicada por um ganho, obtido a partir de um algoritmo fuzzy, a fim de lidar variações nos parâmetros do atrito e ajustar o torque de compensação. Resultados experimentais mostraram um bom desempenho, indicando que o atudor tornou-se aproximadamente línea. 1. Introdução At present, there are many applications of neural networks (NN) in the science domain (Jung and Hsia (1998), Kaynak, O. and Ertugru, M. (1997)). This subject has been object of great attention of the scientific community. In Miller et al. (1995), for instance, there is an important description of the history of the so-called neural networks. This paper investigates the identification of the friction torque of a geared motor drive joint robotic actuator using neural networks. The main motivation is the difficulty in obtaining a very realistic drive joint dynamic model mainly due to the internal non- linear friction characteristics of the actuators (Armstrong (1988)). In spite of NN application in robotic been relatively old (approximately fifteen years ago), NN applications to drive joint non-linear friction estimation are more recent. Dapper et al. 193