Control of a flexible-joint robot using neural networks

Abstract

Traditional robot control strategies assume both joint and link rigidity for the purpose of simplifying the control problem. The demand for greater precision coupled with the increased use of lightweight materials necessitates the inclusion of elastic dynamics in the control strategy. These highly nonlinear dynamics which increase the order of the system are extremely difficult to formulate with sufficient accuracy. The standard form of adaptive control does not appear to be applicable, since the basic assumptions on the system dynamics and nonlinear characteristics are rarely satisfied. For the case of manipulators with flexible joints, we propose an alternate control scheme which does not rely on accurate a priori knowledge of the manipulator dynamics, but instead can "learn" these dynamics by using a neural network.