Deep Neural Network-Based Jacobian Control of Robot Manipulators: Offline Regression and Online Adaptation

Abstract

Deep neural networks (DNNs) are powerful tools with exceptional approximation capabilities across various applications. However, there is still a lack of substantial progress in attaining stable DNN-based robot feedback control. Current neural network-based Jacobian feedback control methods typically update weights for each specific robot task, emphasizing local task learning over general kinematics learning. There is currently no systematic way of learning the Jacobian matrix using DNNs and deploy it online for task space control in a stable manner without additional offline training for each task. This article introduces a DNN control system comprising a deep regression module and an online adaptation module. The regression module is trained offline with shuffled precollected data, and the online adaptation module is updated online with real-time data for specific tasks. This system enables DNNs to learn general robot kinematics using the regression module and adapt to new tasks online. The performance of the proposed method is demonstrated through experiments on tracking tasks performed by a UR5e robot manipulator.