Control Optimization of Robotic Manipulators Using Reinforcement Learning

Abstract

Conventional robot force control schemes are basically model-based methods. However, it is very difficult to obtain exact modeling of robot dynamics and there are various uncertainties in the contact environment. In this paper, we propose a reinforcement learning control approach to overcome these drawbacks. We used artificial neural network (ANN) as the learning structure, and applied the stochastic real-valued (SRV) unit as the learning method. Firstly, simulations of force tracking control of a two-link robot arm were carried out to verify the control design. The results show that even without the information of the robot dynamic model and environment states, operation rules for simultaneous controlling force and velocity can be achieved via repetitive exploration. However, the learning speed was quite slow. Therefore in practical control applications, our approach is to enhance the performance optimization. We propose to combine a conventional controller with the reinforcement learning strategy. Finally, experimental results are presented to demonstrate the proposed method.