A Preliminary Study on Softness Optimization Using Machine Learning for the Digital Twin of Soft Robots

초록

This study presents a machine learning-based approach for optimizing Young’s modulus, a critical physical parameter of soft robots. Instead of directly utilizing conventional material property data, the method predicts Young's modulus based on positional coordinate data measured from key points on the deformed soft robot. The research consists of simulation and experimental phases. In the simulation phase, the convergence of the Young’s modulus estimation framework is first validated through gradient descent optimization. Subsequently, random forest and neural network models are trained using coordinate data collected over a Young’s modulus range of 10²–10¹⁰ Pa. The random forest model exhibits the lowest RMSE for predicting specific Young’s modulus values (10⁶ and 10⁸ Pa), demonstrating optimal performance. In the experimental phase, deformation data from a TPU-based 3D-printed soft robot are applied to the optimized random forest model to predict Young’s modulus in real-world conditions. The proposed method provides realistic predictions compared to publicly available modulus values. These findings confirm that simulation-trained machine learning models can be effectively applied to optimize soft robot design and control, enhancing the reliability of digital twins and soft robot engineering.

키워드