A recurrent neural network-based adaptive variable structure model following control of multijointed robotic manipulators

A scheme that uses neural networks for an adaptive implementation of variable structure control for multijointed robotic manipulators in complex task executions is presented. The control strategy is developed within the general framework of nonlinear model-following control, and within attempts to minimize the total time for nullifying the deviations from the desired model behavior while ensuring a specified percentage of time on the sliding manifolds in order to exploit the disturbance attenuation features present during the sliding motions. These objectives are realized by tailoring an adaptation process that consists of appropriately adjusting the controller gains to keep the motion on the sliding manifolds, and of progressively updating the sliding manifold parameters. A rapid execution of the adaptation process is facilitated by a multilayer recurrent neural network with a supervised training algorithm. The resulting control scheme is decentralized and permits the design of independent joint controls. A quantitative performance evaluation of the neural network-based adaptive controller is given in various task scenarios such as regulation, trajectory tracking, and model following