Fuzzy-based reinforcement learning of a robot force control skill

Abstract

Humans perform many tasks with relative ease. In spite of this, many tasks are difficult to model explicitly and it is difficult to design and program automatic control algorithms for them. The development, improvement, and application of learning techniques taking advantage of sensory information would enable the acquisition of new robot skills and avoid some of the difficulties of explicit programming. This paper describes an approach for the generation of skills for the control of multidegree of freedom robotic systems. In the method, the acquisition of skills is done online by self-learning. Instead of generating skills by explicit programming of a perception to action mapping, they are generated by trial and error learning, guided by a performance evaluation feedback function. The structure of the controller consists of two fuzzy subsystems both implemented by feedforward multilayer neural networks. The action fuzzy subsystem has the purpose of generating command actions for the system under control. The evaluation-prediction fuzzy subsystem predicts the future value, a function that evaluates the performance of the controller. Simulation results concerning the application of the approach to learning a robot manipulator force control skill are presented.