1 Neural Predictive Control for a Car-like Mobile Robot Dongbing Gu and Huosheng Hu Department of Computer Science, University of Essex Wivenhoe Park, Colchester CO4 3SQ, UK Fax: +44-1206-872788 Email: dgu@essex.ac.uk, hhu@essex.ac.uk Abstract: This paper presents a new path-tracking scheme for a car-like mobile robot based on neural predictive control. A multi-layer back-propagation neural network is employed to model non-linear kinematics of the robot instead of a linear regression estimator in order to adapt the robot to a large operating range. The neural predictive control for path tracking is a model-based predictive control based on neural network modelling, which can generate its output in term of the robot kinematics and a desired path. The desired path for the robot is produced by a polar polynomial with a simple closed form. The multi-layer back-propagation neural network is constructed by a wavelet orthogonal decomposition to form a wavelet neural network that can overcome the problem caused by the local minima when training the neural network. The wavelet neural network has the advantage of using an explicit way to determine the number of the hidden nodes and initial value of weights. Simulation results for the modelling and control are provided to justify the proposed scheme. Key words: Model predictive control, Wavelet neural network, Path tracking 1. INTRODUCTION Car-like mobile robots are mainly used in industry, ports, planet exploration, nuclear waste cleanup, agriculture and mining since they have the necessary loading capability. The control scheme for such mobile robots is traditionally decomposed into three subtasks: trajectory generation, position estimation and path tracking. Although behaviour-based approaches have been very popular since Brooks' invention in 1980s[2], their sole implementation on these mobile robots for many complex applications causes some difficulty. Instead, a behaviour-based approach can be incorporated with a traditional planning-based system to form a hybrid system to deal with unexpected situations [1]. The aim of trajectory generation is to generate a feasible path that allows the robots to move smoothly. A path consisting of straight lines and circular arc segments is easy to generate, but may not be suitable for a car-like mobile robot which has the non-holonomic constraint and its steering angles at the line-arc transition points have discontinuous curvature. Although a Clothoid's trajectory [10] can provide smooth transitions with a continuous curvature, its disadvantage is lack of the closed-form expression. Nelson proposed a continuous-curvature polar polynomial for path generation [14], which has a closed form expression and is easy to calculate. Pinchard improved Nelson's polar polynomial with variable speeds, leading to a fifth-order polynomial for a continuous-curvature path [16]. Position estimation plays a key role in path tracking. Many approaches have been proposed in terms of the different sensors on robots. Odometry is simple, inexpensive, and easy to implement in real time, but suffers from the unbounded accumulation of errors. The triangulation approach is appropriate for initialising position, but is too sensitive to angle measurements when robots move around. Grid-based position [12] is a successful method in dealing with sensory uncertainty, but is not suitable for model- based predictive control due to the time it takes. A Kalman filter is a feasible estimation approach that can fuse multiple sensory measurements to provide relatively accurate results. Path tracking for a car-like robot is in fact a feedback control problem to which model-based predictive control (MPC) has been an effective mechanism [4][19]. Most MPC control applications are based on linear models of dynamic systems to predict outputs over a certain horizon. Future sequences International Journal of Robotics and Autonomous Systems, Vol. 39, No. 2-3, May, 2002