Neural netowrk algorithms for robot control

grantor: University of TorontoA neural network module was developed with the purpose of improving a commercial robot's (namely the CRS Robotics A460) trajectory tracking performance. The A460 is equipped with a factory supplied Proportional, Integral and Derivative (PID) controller. The neural network is trained on-line to minimize the trajectory tracking error of the robot joint actuators and improve trajectory tracking performance. The Error Back-Propagation neural network (Rumelhart et. al. (1986)), Mixture of Experts neural network (Jacobs and Jordan, 1992) and MOVE neural network (Graham and D'Eleuterio (1990)), are implemented into the neural network module. Experiment results are given illustrating a 66% reduction of joint error norm. A state-space model of the robot control system was developed and implemented in a numerical simulation. Sensitivity function calculation capabilities were included in the simulation. Sensitivity calculations from the simulation are verified with results from robot experiments.M.A.Sc