Anthropomorphic Physical Property Comprehension of Robots: Leveraging VLM and Tactile Perception for Robotic Grasping

In the field of robotic grasping, robots typically employ uniform grasping forces for object manipulation. However, when dealing with soft objects like plastic cups, this consistent force may cause deformation due to their compliant nature. Therefore, adapting grasping forces to different object properties is crucial for maintaining object integrity. With the rapid advancement of vision-language models (VLMs), the ability to deeply understand the characteristics of grasped objects has been significantly enhanced. Building upon this, this paper proposes an Anthropomorphic Physical Property Comprehension Grasping (APPCG) framework that integrates visual physical-property priors with tactile verification. Our approach begins with VLM-based inference of target object properties to generate structured physical descriptors. Initial grasping forces are then computed using mechanical equilibrium equations. During preliminary grasping execution, GelSight sensors capture contact deformation to estimate real-time contact forces via Hooke's law and detecting deformation through spatiotemporal criteria ($r_{d}^{t}$, $MSE^{t}$). Finally, distinguishing rigid or flexible objects by stiffness, we dynamically determine final grasping forces, enabling robotic adaptation to object-specific characteristics. Experiments across daily objects demonstrate the framework's effectiveness: rigid objects achieved 97.7% success (127/130) while flexible objects attained 98.5% success (197/200).