Neural-adaptive backstepping for flexible-joint robots with neither extra parameters, extra networks, nor robust terms

Abstract

A typical direct neural-adaptive control of an n-link rigid manipulator uses n neural network outputs in tandem with n proportional-derivative controls. However, previous approaches to rigid-link, flexible-joint robot control using 3 steps of backstepping have required knowledge of the spring constants, extra adaptive parameters, extra neural network outputs (beyond 3n), and/or robust control terms. This paper presents a control analogous to those for rigid robots at each step of backstepping, using exactly n neural network outputs at each step, no extra adaptive parameters, no extra neural networks, and no extra robust control terms. Simulation results demonstrate the performance with end-effector trajectory tracking of a two-link planar flexible-joint robot.