Neural networks in robotics and control: some perspectives

In recent years, ANNs have been successfully used in various applications such as system identification and control, robotics, pattern recognition and vision. One important application of ANNs is in the area of robotics. In particular, ANNs have been used to compute inverse kinematic transformations of multi-link robot manipulators. The advantage of using neural approach over the conventional inverse kinematic algorithms, is that ANNs can avoid time consuming calculations. Furthermore, in a manner typical of ANNs, it would be very easy to modify the learned associations upon changes in the structure of robot manipulators. With reference to the control paradigm, ANNs have the ability to approximate arbitrary nonlinear functions which is an essential requirement in the design of controllers for nonlinear dynamic systems. Because of their learning and adaptive features, ANNs can be trained to adaptively control various nonlinear systems. On the other hand, the conventional controllers are system specific in other words, a controller designed for a class of nonlinear systems may completely fail to control another class of nonlinear systems. The purpose of this paper is to provide an overview of different neural structures employed in the robotics and control paradigms. The paper describes the biological neuron and the associated terminology. The following neural structures are discussed: static neural networks, recurrent and dynamic neural networks, and fuzzy neural networks