An Adaptive Control System for Biological and Robotic Simulations.

Abstract

An adaptive control algorithm based on Albus' CMAC (Cerebellar Model Articulation Controller) was studied with emphasis on how to build a Multilayered CMAC Control System. This concept has been devised to circumvent the excessive memory requirements of CMAC controllers for complex control systems with many inputs. The Neighborhood Sequential Training technique was devised as a general CMAC training technique. This training technique is straightforward to implement and well matched to CMAC's memory generalization. A two-layered CMAC control module was simulated for a six dimensional CMAC problem of trajectory control for a six degree of freedom manipulator. Layering was accomplished by the decomposition of direct movements of the manipulator end-effector in Cartesian space into three sequential orthogonal sub-movements. The neighborhood sequential training was used to train individual CMACs in the CMAC control module. The resulting system reduced the memory requirement by almost two orders of magnitude. The manipulator tracked a straight line path with average deviation error of less than 0.17 cm for a gross end-effector movement of 22.650 cm.