Efficient data-driven approach for bending deformation capture of fiber-reinforced soft actuators

To address the inherent nonlinear response and the complexity of structural and material parameters that hinder the customization of fiber-reinforced soft actuators (FRSAs), this study proposes a robust data-driven method for accurately predicting their bending deformation. The approach creatively integrates finite element analysis (FEA) models, multiple advanced machine learning algorithms, and the concept of transfer learning, leveraging both simulation and experimental data for modeling. Five key geometric parameters of the FRSA and the actuation pressure were selected as inputs, with the bending angle as the output. An accurate FEA model was first developed to generate a comprehensive simulation dataset for initial training. A set of machine learning models, including nonlinear normalized multilayer perceptron regression (NMLPR), XGBoost, and LightGBM, were rigorously trained and optimized using this dataset. To further enhance prediction accuracy and bridge the gap between simulation and physical reality, these models were refined via transfer learning. Rigorous evaluation on an independent test set demonstrated that the transfer-learning-enhanced NMLPR achieved the best performance, with an average root mean square error (RMSE) of 9.28 $$^{\circ }$$ and mean absolute percentage error (MAPE) of 7.69%, outperforming other baseline algorithms. This work presents a robust data-driven methodology for FRSA bending deformation modeling, provides a foundation for precise actuator design and control, and highlights the efficiency of transfer learning in soft robotics applications.