A Deep Learning-Based Approach for System Modeling of a Novel Cable-Driven Parallel Robot

Abstract

Abstract. Cable-driven parallel robots (CDPRs) drive the end-effector through cables, offering advantages such as low inertia and high payload-to-weight ratio, which make them highly promising for applications in industrial and construction fields. However, the inherent flexibility of CDPRs introduces pronounced nonlinearities and uncertainties, posing significant challenges for system modeling and precise control. In this study, a novel type of CDPR, rigid-flexible hybrid parallel robot (RFHPR) designed for sorting and palletizing is investigated. To address the accurate system modeling challenges of RFHPR, this article proposes the CaRINet, a network combining gate recurrent (GRU) and transformer architecture for system modeling of cable-driven robots. For training CaRINet, a dataset is constructed from the RFHPR by applying the excitation trajectory generated by combining the finite Fourier series and quintic polynomial interpolation as the input, and collecting the corresponding end-effector position and motor torque as the output. Model performance studies are conducted to optimize CaRINet, and comparative experiments are performed against the conventional system identification method. The experimental results demonstrate the effectiveness of the CaRINet and exhibit higher performance compared to conventional identification methods.