Using CMAC neural networks and optimal control

This paper explores the real-time control of an industrial robotic arm to balance a mass at the end of a pole. The 3D inverted pendulum is a MIMO nonlinear inherently unstable system. The control system uses combined optimal and neural network techniques. To provide stable control, an optimal, linear quadratic regulator controller was developed from the linearized system model. When applied to the robotic system, this controller produced a relatively large limit cycle, due primarily to unmodelled system nonlinearities. The CMAC neural network was then introduced into the controller to implement a technique referred to as prediction feedback. The purpose of this adaptive feedback controller was to learn system nonlinearities, reject any residual noise, and reduce the system limit cycle. When applied to the robotic pole-balancer, the addition of adaptive prediction feedback helped to significantly decrease the magnitude and frequency of oscillation. This experiment is a primary example of how an intelligent controller can be developed by combining the strengths of different control techniques