SoftGraspNet for Integrating Deformation Aware Neural Perception with Compliant Gripper Control in Fragile Object Handling

SoftGraspNet addresses the challenge of safely grasping fragile objects whose tolerance to force is both low and uncertain. The system combines deformation aware visual perception with compliant gripper control by predicting object compliance and safe force limits from RGB D input before contact, then refining these estimates during grasping using high resolution tactile feedback. A compliance aware perception module provides spatial compliance maps and probabilistic force envelopes, which are fused with real time tactile sensing through a Bayesian force regulation loop to adapt grip forces safely during contact. Evaluated across a diverse benchmark of fragile objects, the approach demonstrates reliable grasping with low damage rates, showing the value of integrating anticipatory perception, tactile sensing, and uncertainty aware control in a unified robotic manipulation framework.