A Data-Driven Distributed Recurrent Neural Network for a Collaborative System of Multiple Redundant Manipulators With Unknown Structure

Abstract

This article proposes a novel data-driven distributed recurrent neural network (DDD-RNN) based on neurodynamics principles to address the challenge of precise collaborative motion generation in multimanipulator systems (MMCs) with unknown structural parameters. Unlike traditional methods that rely on precise models and existing data-driven methods with single-order Jacobian estimation, this article designs an improved Jacobian matrix estimation law (IJM). For the first time, it synchronously estimates the first-order and second-order Jacobian matrices online, effectively capturing the time-varying characteristics of robotic manipulators. Furthermore, a recurrent neural network solver is designed based on the neurodynamics criterion, which enables it to take into account the time-varying information of robotic manipulators, thus yielding more accurate motion generation results. Simulations conducted on multiple multimanipulator collaborative systems (MMCs) and experiments performed on the Ufactory XArm6 robots have verified the feasibility of the DDD-RNN method in generating collaborative motions of multiple robotic arms, even when the models of the robotic arms are unknown. Comparisons confirm the superiority of the DDD-RNN in terms of end-effector accuracy and applicability.