A robot control application with neural networks

Abstract

An artificial neural network's inherent nonlinearity gives some advantages to their use on different kinds of problems including those in the control engineering arena. The learning of inverse dynamics with neural networks is an example of robot control applications. The dynamics of nonlinear systems vary with their parameters, and, in some cases, determining a single global model of the plant dynamics can be a very difficult problem. Designing piecewise control laws are useful methods to overcome this problem. In robotics, increasing the degree of freedom and working range of each link directly creates more complex dynamics. The structure of a multilayer perceptron is dependent on the controlled plant and, for more complex systems, large networks are required and this increases the real time calculations of robot control. For the proposed scheme, in order to decrease the real time calculations, the working range of the robot is divided into several regions and, for every region, a separate neural network is used. Instead of learning whole dynamics with one large network, using this kind of strategy, one divides the complexity of the dynamics to small networks. In real time control, this piecewise or regional neural network structure is used together with a PD controller.