Robot arm fuzzy control by a neuro-genetic algorithm

Abstract

Robot arm control is a difficult problem. Fuzzy controllers have been applied succesfully to this control task. However, the definition of the rule base and the membership functions is itself a big problem. In this paper, an extension of a previously proposed algorithm based on neuro-genetic techniques is introduced and evaluated in a robot arm control problem. The extended algorithm can be used to generate a complete fuzzy rule base from scratch, and to define the number and shape of the membership functions of the output variables. However, in most control tasks, there are some rules and some membership functions that are obvious and can be defined manually. The algorithm can be used to extend this minimal set of fuzzy rules and membership functions, by adding new rules and new membership functions as needed. A neural network based algorithm can then be used to enhance the quality of the fuzzy controllers, by fine tuning the membership functions.

The approach was evaluated in control tasks by using a robot emulator of a Philips Puma like robot called OSCAR. The fuzzy controllers generated showed to be very effective to control the arm. A complete graphical development system, together with the emulator and examples is available in Internet.