23 - APPLICATION OF NEURAL NETWORKS TO ROBOTICS

Abstract

This chapter focuses on the application of neural networks in robotics. Many researchers and developers expect that true robotic autonomy will be achieved through neural network technology. It seems reasonable that modeling the human nervous system, which functions autonomously, would provide the technology to build independent functional robots. Three of the most useful capabilities of neural networks are (1) the ability to nonlinearly transform information, (2) generalize to novel conditions, and (3) reduce the dependency for massive computations on serial computers. When compared to traditional kinematic or geometric calculations required by the present robotic control systems, neural networks use few computations. This reduction of computational workload reduces the computational resources required to control robots. If fewer computational resources are needed, then smaller and less expensive computers can be used for robots. If smaller computers are used, then robots are more likely to carry their computers on board and need less space to carry the necessary electronics. Additionally, neural network algorithms are completely compatible with parallel processing. Neural networks can learn to map nonlinear relationships. This capability is one of the most important advantages available from neural network technology. In case of robotics, neural networks are expected to compensate for the nonlinear dynamics necessary to control robotic hardware.