AN OPTIMAL ADAPTIVE CONTROLLER BASED ON ONLINE ACTOR-CRITIC LEARNING FOR A ROBOT MANIPULATOR

The uncertainties in the parameters of a robot manipulator can significantly affect the robot performance, causing steady-state and trajectory following errors. Adaptive controllers are a good alternative for these systems, since their main feature is the capability to learn online using real-time parameter estimation. Nevertheless, adaptive controllers are not usually designed to be optimal to a prescribed performance index, and thus, are not suitable to applications in which optimal use of resources is highly desirable, for instance humanoid and service robots. This paper presents the design and performance study of a controller that combine features of adaptive control and optimal control applied to a robot manipulator. Specifically, the proposed control scheme is implemented as an actor-critic structure, which is in the reinforcement learning context, characterizing this design as a model-free approach. In contrast to others actor-critic systems in which two independent neural networks are used, one for approximating the value function and another for learning the control actions, in this scheme, a single neural network is defined, reducing the number of parameters to be estimated. The simulation results validate the desired performance of the proposed controller applied in a two-link robot manipulator with revolute joints.