235 GENERALISED PREDICTIVE CONTROL OF AN INDUSTRIAL MOBILE ROBOT HUOSHENG HU✝ and DONGBING GU✠ ✝ Department of Computer Science, University of Essex, Wivenhoe Park, Colchester CO4 3SQ, England ✠ Department of Electronics, Changchun Institute of Optics and Fine Mechanics, Changchun, Jilin, 130022, China ABSTRACT Most of industrial mobile robots or autonomous guided vehicles have been built on car-like platforms to achieve high loading capacities. However, to plan smooth trajectories for these mobile robots and subsequently tracking them precisely is a difficult task since these robots are subject to non-holonomic constraints. In this paper, we address two fundamental control problems in order to achieve a high performance of real-time operation, namely a motion controller using the Generalised Predictive Control algorithm and the position estimation using Extended Kalman Filter which integrates the data from optical encoders and a rotating laser scanner. Some experimental results are given to show the feasibility of the proposed control algorithm. Keywords: Robotics, Mobile Robots, Position Estimation, Intelligent Systems, Nonlinear Systems 1. INTRODUCTION The deployment of car-like mobile robots in industry needs to address three main issues: (i) how to reduce noise in sensor data; (ii) how to handle nonhlonomic constraints; (iii) how to deal with dynamic changes in the environment. This paper is focused on the first two issues, namely path tracking which is to follow a previously planed path by taking into account the position estimation and the constraints imposed by the vehicle's kinematics. Most path tracking methods are based on the error between the current estimated vehicle position and the desired path. In [14], a linear proportional feedback of errors is used. Its improved version has been realised in [4] with adding error integration to form a PI feedback. Another PI control algorithm is implemented in [10]. The disadvantage of those PID controllers is that its outputs are only based on position and orientation error with little consideration of the vehicle's kinematics and nonlinearity. In [11], a nonlinear control algorithm is presented and the stability is proved under the assumption of perfect velocity tracking. The Generalised Predictive Control (GPC) algorithm developed by Clarke et al. [2][3] appears to overcome • a nonminimum-phase plant, • an open-loop unstable plant or plant with badly- damped poles, • a plant with variable or unknown dead-time, and • a plant with unknown order. The GPC algorithm is based on a receding horizon approach similar to that used by the human operator in many manual control tasks in which the ability of the human operator to "look ahead" or "preview" is a vital strategy. It has been used for rotorcraft terrain following [5] and manipulator control [12]. In [15], GPC is used to track path for CMU mobile robot in which the model is obtained by local linearisation and discretisation of the dead reckoning equations for constant velocities, and considering first order dynamics for both the steering and the drive system. In contrast, the intermediate control variables are adopted for the GPC algorithm in this paper in order to avoid complexity caused by linearization of the vehicle kinematics and reduce the computation time. The position estimation plays a key role in path tracking. Many approaches have been developed so far such as odometry, laser scanner, gyro, GPS, sonar and vision, of which odometry is simple, inexpensive and easy to implement in real time. The measurement based on odometry is called relative positioning [1]. The disadvantage of odometry is that it is inaccurate with an unbounded accumulation of errors. The laser scanner with artificial beacons is a promising absolute positioning technique that scans the artificial beacons in the environment and measures their bearing relative to each other. In [13] and [16], the location of the robot was calculated by the triangulation approach with the identified beacon's bearing relative to each other. The accuracy of the location is however affected by beacon's positions and bearing measurements. When the robot moves, the accuracy will be affected because of the time needed for scanning. In this paper, an Extended Kalman Filter (EKF) is presented to fuse the data from odometry and the observation from the laser scanner [7]. This paper is structured as follows. Section 2 introduces a car-like mobile robot with details of kinematics and motion control architecture. Section 3 describes a predictive path controller that is intended to minimise a cost function with the error between the desired robot position and the Proceedings of the IASTED International Conference, INTELLIGENT SYSTEMS AND CONTROL October 28-30, 1999, Santa Barbara, California, USA