ONLINE IDENTIFICATION OF A ROBOT USING BATCH ADAPTIVE CONTROL Björn Bukkems, 1 Dragan Kostić, 2 Bram de Jager, 3 and Maarten Steinbuch 4 Technische Universiteit Eindhoven, Department of Mechanical Engineering, Dynamics and Control Technology Group, P.O. Box 513, 5600 MB Eindhoven, The Netherlands 1 B.H.M.Bukkems@tue.nl, 2 D.Kostic@tue.nl, 3 A.G.de.Jager@wfw.wtb.tue.nl, and 4 M.Steinbuch@tue.nl Abstract: A technique to identify parameters of a robot dynamic model is presented in this paper. It is based on a batch adaptive control algorithm that, using a model of the robot dynamics, realizes a repetitive robot trajectory. The tracking error decreases due to a feedforward control input generated from the dynamic model. This feedforward input is computed after adaptation of the model parameters at the end of each trial. As the algorithm is effective, even if the model parameters are all initially set to zero, it can be used to recover their physical values. For that purpose, an identification experiment is carried out during which the robot is excited persistently. The estimation technique admits an online implementation without a delay between trials and is quite appealing for use in practice. Its merits are experimentally demonstrated on a spatial direct-drive robotic manipulator with 3 rotational joints. Copyright © 2002 IFAC Keywords: Robotics, Identification, Dynamics, Application, Adaptive, Model-based control 1. INTRODUCTION Growing interest in model-based robot control is boosted by the computational power of modern digital processors and by advances in the theory on robot modelling and identification (Armstrong, 1989; Gautier and Khalil, 1992; Kozlowski, 1998; Slotine and Li, 1991; Swevers, et al., 1997; Calafiore, et al., 2001; Olsen and Petersen, 2001). A dynamic model simplifies analysis of the control problem at hand and facilitates design of a solution to that problem. The model may compensate for nonlinear dynamic couplings between robot axes and enables robust robot operation of high performance, even if linear feedback control designs are used (Kostić, et al., 2002a). Adaptive, sliding-mode and other nonlinear control strategies also make use of a dynamic model. To use a dynamic model for control of a robotic system, one must be sure that the model closely matches the real dynamics. To enable a close match, the structure of the model should be capable of describing the relevant aspects of the physics and accurate values for the model parameters are needed. So far, a number of theoretical and experimental studies on estimating model parameters have been reported. Each exploits the well-known property that a model of the robot dynamics can be represented linearly in a minimum set of identifiable parameters, called the base parameter set, BPS, (Mayeda, et al., 1990). The elements of the BPS are nonlinear combinations of robot inertial parameters, such as mass and moments of inertia of the robot links, as well as the Cartesian coordinates of the center of mass. Friction effects should be taken into account as well if model-based control of high quality is desired. Friction may cause problems, e.g., steady state errors and limit cycles. To avoid these, the friction force is counteracted using model-based or non-model based techniques (Ray, 2001). If model-based techniques are used, then relevant values of the friction parameters are needed in addition to the BPS. In general, we can distinguish between the least- squares-like (Kozlowski, 1998; Swevers, et al., 1997; Calafiore, et al., 2001; Olsen and Petersen, 2001) and adaptive techniques (Slotine and Li, 1991) to estimate elements of the BPS and the friction parameters.