Towards a Comparative Study of Neural Networks in Inverse Model Learning and Compensation Applied to Dynamic Robot Control

This report deals with the applications of neural networks in inverse model learning and compensation to the mobile manipulator dynamic trajectory tracking and control. The mobile base is subject to a non-holonomic constraint and the base and onboard manipulator case disturbances to each other. Compensational neural network controllers are proposed to learn to reach a sequence of targets with given times, to track dynamic trajectories under a non-holonomic constraint and torque limit constraint and to compensate for uncertainties in the non-holonomic base and the manipulator and the disturbances between the base and the manipulator. Both multi-layered perceptron networks and radial basis function networks are considered in the report. Comparison was made between neural network controllers with and without model information. It is shown through various simulations the proposed neural network compensation schemes perform better than conditional controllers.