A neural network based learning controller for robot manipulators

An iterative learning control using neural network design is presented for robot manipulators with input disturbance and re-initialization uncertainty. A sampled-data feedforward learning algorithm is designed under a feedback configuration and a rigorous proof via a discrete approach is given to study the learning performance. It is shown that under a sufficient condition on the learning gain, convergence and robustness of tracking error in the iteration domain can be guaranteed at each sampling instant if the sampling period is small enough. Since the implementation of learning gain depends on the information of the input-output coupling matrix of robot manipulator, a neural network is proposed to solve the implementation problem. A training procedure is applied to estimate the robot manipulator by using only input-output data. The neurons, equivalent to the premise and consequent parameters of a fuzzy system, are tuned by gradient descent and least squares estimate. This will give an initial setting of the neural-network based iterative learning controller. During the control iterations, the neural network can still be tuned for each iteration in order to improve the approximation accuracy and increase the tracking speed.