Machine Learning-Based Deformation Estimation for Grasping 3D Soft Objects

When grasping soft objects, it is of utmost importance to precisely adjust the gripping force and accurately predict deformations. This study was conducted to measure the force and deformation responses at specific graspable points, facilitating the development of a machine learning model capable of predicting force and deformation responses at arbitrary positions within three-dimensional soft objects. To evaluate our proposed methodology, we collected datasets using a sponge and a plastic bottle, independently trained the model, and carried out a comprehensive assessment. The results underscore the model's remarkable accuracy, with the deformation estimation error compared to the actual deformations (with a maximum deformation of 10 mm) falling within the range of ±0.67 mm, affirming its suitability for grasping tasks. Additionally, our findings demonstrate that the model maintains a reasonable level of prediction accuracy even when the number of measurement locations is reduced.