DeformGrasp Deep Learning Driven Robotic System for Real Time Grasping of Deformable Objects Using Visual and Tactile Fusion

Deformable object grasping is challenging because soft items change shape during contact and their material properties are hard to infer from vision alone. This paper presents DeformGrasp, an end to end deep learning system that fuses RGB D vision with high resolution tactile sensing to estimate deformation state and generate adaptive grasp pose and grip force commands in real time. The approach uses a Deformation State Estimator and a Grasp Policy Network with a synchronised sensor fusion pipeline running at deployment speed, and it is evaluated on DeformBench, a dataset of annotated grasp trials across multiple deformable object categories, where it achieves strong grasp success and improves handling quality through better force selection and reduced deformation.