Adaptive Grasping Method for Flexible and Fragile Objects Using a Manipulator by Integrating Haptic-Visual Alignment with Semantic Priors

Deformations and damage frequently occur during the grasping of flexible and fragile objects, rendering traditional force-control and vision-based strategies inadequate for meeting robustness and safety requirements. This paper proposes an adaptive grasping method integrating tactile-visual feature alignment with semantic prior guidance.First, a large model extracts semantic constraints such as "material–deformation threshold–operation region" and aligns them with tactile encodings. Then, adaptive impedance control enables real-time force and pose adjustment. Experiments covering 60 categories of flexible/fragile objects and over 5,000 grasping tests demonstrate: compared to a diffusion policy without haptic feedback, this method achieves a +15.2% increase in grasping success rate, a 33.5% reduction in target damage rate, and an 18.4% decrease in 6D pose error. Ablation analysis indicates that tactile and semantic priors contribute +6.3% and +4.5% performance gains, respectively. These results demonstrate the method's effectiveness for real-time flexible grasping.