Neural Networks and the Applications for Robot Control
Chief of Bio-dynamics Division Industrial Products Research Institute
It is well known that robots are being skillfully applied and with favorable performance in a variety of fields, for use in the Japanese manufacturing industry in particular, thanks to progress in robot technology. Today, robots are expected to accommodate men and in the near future be utilized in the field of home life in compliance with human beings. Pessimistically speaking, however, it is impossible to deny that conventional robots, such as teaching playback robots (which men must operate directly), are not able to play roles in the future as expected, so that the development of a new control system which is able to overcome conventional systems in performance ability is indispensable. In other words, flexible control systems by which robots are able to behave autonomously, with minimum human interference is urgently required. We believe that the following three concepts are indispensable for a robot to be equipped with flexibility. a) Manipulators/hands and lggs / wheek with human flexibility. b) Control of flexible and intelligent motions for control in manipulation/handling and locomotion; c) Flexible intelligence and a sense of judgement which permits the robot to execute motions autonomously, adapting itself to the requirements of the human environment. Solving these problems will require investigation into information processing, a study into the function of the brain and central nervous system of human and other living bodies. Thus the information processing theory about neural networks which simulate the functions of the brain has progressed rapidly to activate R & D on the application of motion control and speech processing which have made use of the conventional Neumann computer difficult to handle. Neural networks have the capacity of parallel distributed processing and self-organization as well as learning capacity. Its theory has provided an effective basis for materialization of flexible robots. In the field of level b. and c. mentioned earlier, the neural network theory comprises a large potential to be applied to robots, so that attention is being focused on it. Nevertheless, information processing by neural network is not omnipotent for solving such problems. Presently, it is difficult for a neural network to solve problems which require complex calculations in robot control; for instance, such controls that take force and acceleration into account. Control of flexible robots which mobilize whole arms will require parallel processing of data obtained from many sensors and to control numerous degrees of motion. Therefore, it has become increasingly important for problem solving to combine such problems inherent to robots with parallel processing, self-organization and learning ability of neural networks. From this point of view, therefore, further promotion of R & D on the application technology of neural network for robots is important. These efforts will produce a new neural network-theory for robots and eventually permit autonomous motion. This special issue compilied articles related to applications of neural network to robots, which were produced in the above mentioned environment, from a review on neuromorfhic control, through dynamic system control, optimal trajectory, planning of motion for handling, manipulator locomotion and travelling, to problems in application systems. We hope these articles help our readers understand the present state of Japanese R & D and the application of neural network for robots, as well as new subjects possible for progress in the future. Finally, we gratefully acknowledge Prof. Toshio Fukuda (who contributed a review) and other contributors on their latest achievements.