Learning algorithms for manipulator control

A method of robot manipulator control is proposed whereby algorithms are used to learn sum of polynomials representations of manipulator dynamics and kinematics relationships. The learned relationships are utilized to control the manipulator using the technique of Resolved Acceleration Control. Such learning is achieved without recourse to analysis of the manipulator; hence the name Self-Learned Control. Rates of convergence of several learning algorithms are compared when learning estimates of various non-linear, multivariate functions. Interference Minimization is found to be superior to the Gradient Method, Learning Identification and the Klett Cerebellar Model. Simplification of the implementation of Interference Minimization is described. A variant, Pointwise Interference Minimization, is introduced that is suitable for certain applications. Self-Learned Control with path specification in Cartesian coordinates is demonstrated for a simulated two link manipulator. It is shown that sum of polynomials representations of the inverse dynamics, inverse kinematics and direct position kinematics relationships are adequate to achieve control comparable to that achieved using their analytical counterparts and can be learned without analysis of the manipulator. Further research is outlined to achieve automatic adaptation to tool mass, implementation of sum of polynomials estimators and enhancement of the Klett Cerebellar Model.Applied Science, Faculty ofElectrical and Computer Engineering, Department ofGraduat