Experimental implementation of neural network controller for robot undergoing large payload changes

A robot controller based on artificial neural networks (ANNs) is presented which is capable of compensating for changing payload masses. Two different feedforward (multilayer) neural networks are used to generate the inverse dynamics and to estimate the payload mass of a two-link planar manipulator. The inverse dynamics ANN receives the same input signals as a conventional computed torque controller as well as the payload mass estimate. By using a separate ANN to generate the payload mass estimate, both ANNs can be trained off-line. The proposed neural network architecture is implemented on actual hardware using a neurocomputer. Experimental results indicate that the ANN-based controller is able to capture the nonlinear dynamics of the actual manipulator. The ANN mass estimator responds very quickly to changing payloads