Rapid learning using CMAC neural networks: real time control of an unstable system

Abstract

Results of real-time experiments involving the control of an unstable, nonlinear plant using the feedback error learning approach and CMAC (cerebellar model articulation controller) neural networks are presented. The plant comprised a one-wheeled cart pushed from behind by means of a vertical rod in a ball bearing (similar to the caster on the leg of a chair) by a General Electric P-5 industrial robot with a wrist-mounted video camera. The task was to accurately follow a winding track (with a constant cart velocity) drawn on a flat surface. It was found that simple fixed control laws alone were not sufficient to keep the cart from flipping rapidly around the point of contact with the robot (referred to as 'spinning out'). However, similar fixed control laws were sufficient for training control CMAC neural networks during online learning, such that the plant was rapidly stabilized. The controller was able to track the path accurately, at the requested velocity, within three laps around the track. Stable performance could not be achieved in a similar neural network configuration trained by direct inverse modeling.<>