Self-adaptive composite control for flexible joint robot based on RBF neural network

Among numerous control schemes for flexible joint robots, the main problem is that the full state variable of acceleration and jerk must be known, which are difficult to measure, and the noise may be merged in the main signal. To solve this problem, a self adaptive composite control scheme is developed to control the flexible joint robots with modeling errors and subject to uncertain disturbances, which is based on considering the system as a low dimensional nominal rigid and a linear elastic subsystem. Using this approach, the controller consists of a slow and a fast term, the slow control is based on the well-known Computed Torque method and a RBF neural network based compensating controller. The neural network is trained on line based on Lyapunov theory to compensate for the modeling uncertainties, thus its convergence is guaranteed. The fast term is designed to provide stiffness and damping for eliminating elastic deformation. Simulations are presented for a planner manipulator with two flexible joints, the trajectory tracking results are provided to demonstrate performance of the scheme.