On the reduction of costs for robot controller synthesis ?

The demand for smarter and more adaptive robots, mainly due to the increasing availability of sophisticated sensor systems for both manipulation and mobile robots and the recently emerging market for intelligent service systems, has a direct impact on the costs of robot programming. Especially the development of control software for industrial robots is becoming more and more expensive. This paper investigates the possibility of automatizing the cycle of designing and reening the control programs. In particular, two function approximators, Time Delay Neural Networks and Fuzzy Controllers, are evaluated showing that both can be induced from examples of control actions supplied by a good teacher. However, due to the diiculty of nding good training sets, pure approximation is not suucient to solve the task in its generality. Since a human teacher performs not at all optimally in executing control actions and usually uses sensorial inputs that are not directly compatible with the sensor input available to the robot, the oo-line generated controller must be reened on-line. Therefore, Reinforcement Learning is discussed and proposed as a methodology to perform the on-line reenement.