An Adaptive-Gain Recurrent Neural Network for Solving the Joint-Angle Drift Issues of Redundant Manipulators

This paper proposes a new adaptive-gain recurrent neural network (AG-RNN) to effectively cope with the joint-angle drift issues in redundant manipulators. Specifically, a joint-angle drift-free with the feedback control of the velocity layer motion equation (JADF-FC) is proposed via an optimization criterion for synchronously optimizing linear terms and quadratic. Then, the JADF-FC is reasonably formulated into a standard quadratic programming (QP) issue. Different from the previous recurrent neural networks (RNNs), the AG-RNN proposed in this paper constructs an error-based differential equation with a new adaptive-gain. It should be noted that the proposed adaptive-gain does not gradually approach infinity as time increases, which is more in line with actual hardware implementation requirements than the existing time-variant-gain. The adaptive-gain can reduce the joint-angle drift errors of the redundant manipulator. Therefore, the proposed AG-RNN can solve the QP problem of the manipulator more effectively and quickly. To validate the performance of the proposed AG-RNN, it is compared with representative RNNs. The experimental results indicate that smaller joint-angle drift errors can be get by the proposed AG-RNN solving JADF-FC scheme than the other solutions when solving the joint-angle drift issues.