New Varying-Parameter Recursive Neural Networks for Model-Free Kinematic Control of Redundant Manipulators With Limited Measurements

Abstract

Taking advantage of the computational efficacy, the recursive neural network (RNN) has been successfully applied to the tracking control of redundant manipulators. However, the existing control system based on the fixed-parameter RNN (FP-RNN) model is found to be deficient in terms of convergence, robustness, and practicability, and the measurements of end-effector position, velocity, and acceleration are necessary. To overcome the limitations of the existing FP-RNN-based method, this article investigates and discretizes three kinds of existing varying-parameter RNN (VP-RNN) models to achieve the model-free control of redundant manipulators. More importantly, considering the drawbacks of these VP-RNN models, this article proposes a new continuous-time VP-RNN (CVP-RNN) model and a model-free control system with finite-time convergence and noise-rejection capability, using limited measurements. The control system includes two CVP-RNN models, one of which is utilized to estimate the Jacobian matrix of redundant manipulators and another model to solve the inverse kinematics of manipulators. Moreover, the proposed CVP-RNN model and control system are improved to discrete form to facilitate the deployment of CVP-RNN in digital circuits and numerical algorithm development. Based on the measurable end-effector position, low-pass filters are exploited to deal with the problem of immeasurable end-effector velocity and acceleration, thereby implementing the proposed algorithm without the measurements of velocity and acceleration. Finally, both simulation and experiment based on the KINOVA Gen2 manipulator substantiate the efficacy and merits of the proposed control method.