Abstract This article presents a transfer learning method to address the kinematics problem of a planar continuum parallel manipulator with large-deflection links. This method is motivated by the need for lightweight, space-efficient mechanisms in small spaces or human–machine collaborative environments. The manipulator has two independent branch chains with highly flexible panels as links, and the moving platform is driven by the bending deflections of these links. For kinematic analysis, sensitivity analysis is used to identify key parameters influencing the motion. Neural networks are then constructed for forward and inverse kinematics. Simulation data, such as end-effector pose and actuation lengths, are collected for preliminary network optimization. Considering the differences between actual and simulation platforms, transfer learning is applied to further optimize the network parameters. The proposed method leverages the nonlinear prediction capability of neural networks to handle complex large-deformation link modeling. Transfer learning significantly reduces training data and time while enhancing prediction accuracy. Experiments validate the method by comparing it with results without transfer learning. The maximum and average position errors of the end-effector are reduced from over 10.46 mm and 3.56 mm to approximately 1.21 mm and 0.19 mm, respectively.
Liu et al. (Wed,) studied this question.