ORIGINAL ARTICLE F. Nagata (*) · T. Mizobuchi Department of Mechanical Engineering, Faculty of Engineering, Tokyo University of Science, Yamaguchi, 1-1-1 Daigaku-dori, Sanyo-onoda 756-0884, Japan e-mail: nagata@ed.yama.tus.ac.jp T. Hase · Z. Haga R&D Center, Meiho Co., Nogata, Japan K. Watanabe Department of Intelligent Mechanical Systems, Graduate School of Natural Science and Technology, Okayama University, Okayama, Japan M.K. Habib Mechanical Engineering Department, School of Sciences and Engineering, American University in Cairo, Cairo, Egypt This work was presented in part at the 15th International Symposium on Artificial Life and Robotics, Oita, Japan, February 4–6, 2010 Artif Life Robotics (2010) 15:101–105 © ISAROB 2010 DOI 10.1007/s10015-010-0776-9 Fusaomi Nagata · Takanori Mizobuchi · Tetsuo Hase Zenku Haga · Keigo Watanabe · Maki K. Habib CAD/CAM-based force controller using a neural network-based effective stiffness estimator 1 Introduction Many kinds of robot systems for polishing and deburring have been developed in various manufacturing fields. In the manufacturing process of small metallic molds such as LED lens molds, a 3D CAD/CAM system and an NC machining center are generally and widely used, and these have drasti- cally rationalized the design and the manufacturing process. However, the final finishing process after the machining processing has hardly been automated at all because the target mold has several concave areas to be polished in almost all cases. That means that the target mold is not axis- symmetric, so that conventional effective polishing systems, which can only deal with axis-symmetric work pieces, cannot be applied. Accordingly, such axis-asymmetric lens molds are polished by skilled workers in related industrial fields. To automate the finishing process of metallic molds, we have already proposed a polishing robot for PET bottle blow molds, 1 and a desktop NC machine tool for LED lens molds. 2 The control strategy for these two types of system is almost the same. A ball-end abrasive tool attached to an arm tip is controlled by a CAD/CAM-based position/force controller. The force controller used is called the impedance model force controller, in which the desired damping is tuned to be in a critical damping condition. The desired damping is one of the impedance parameters, and has a large influence on force control stability. The critical damping condition is calculated using the effective stiffness, that includes the characteristics of the robot or the NC machine tool itself, a force sensor, an abrasive tool, a jig, a floor, etc. However, undesirable nonlinearities exist in the effective stiffness. In this article, an impedance model force control using neural networks is proposed to deal with the nonlinear effective stiffness. First, the nonlinearity is examined by simple static position and force measurements. The effective stiffness is easily obtained by these measurements. The desired damping in the impedance model force control law is statically calculated from the critical damping condition of the force control system. The neural networks learn the Abstract In industries manufacturing metallic molds, various NC machine tools are used. We have already pro- posed a desktop NC machine tool with compliance control capability to automatically cope with the finishing process of LED lens molds. The NC machine tool has the ability to control the polishing force acting between an abrasive tool and a work piece. The force control method is called imped- ance model force control. The most effective gain is the desired damping of the impedance model. Ideally, the desired damping is calculated from the critical damping condition after considering the effective stiffness in the force control system. However, there is a problem in that the effective stiffness of the NC machine tool has undesir- able nonlinearity. The nonlinearity has a bad influence on the force control stability. In this article, a fine tuning method of the desired damping is considered using neural networks. The neural networks acquire the nonlinearity of effective stiffness. The promise is evaluated through an experiment. Key words Neural network · Nonlinear effective stiffness · Force control · NC machine tool Received: April 1, 2010 / Accepted: April 8, 2010