Comparison of different neural approximation approaches in the path tracking problem

The problem of robot path tracking is defined as a search for motor torques that will drive the robot arm along a desired trajectory at every instant of time. Therefore the additional use of a neural controller in connection with a conventional linear controller was already shown to be very powerful. By theoretical study and extensive simulations the authors make detailed comparisons between two neural controller approaches which represent completely different approximation natures. The first approach is based on the well known backpropagation. In the second approach the authors use the cerebellar model articulation controller (CMAC) developed by Albus. To demonstrate the capabilities of the BP- and the CMAC-based neural controller for path tracking systems, simulation results are obtained for a mathematical model of 3-joints of the MANUTEC R3 industrial robot. The authors´ attention is especially focused on path tracking accuracy in general, time complexity, storage capacity, learning speed generalization capabilities and interference characteristics varying some relevant parameters in both networks.