Shape-Based Extrapolation of Contact Patterns to Support Tactile Regrasping

A method is proposed to support robotic regrasping by leveraging tactile data to extrapolate unseen contact regions. Initial tactile feedback from multimodal capacitive sensors mounted on robotic fingers was used to classify the object shape into prototypical categories. Based on this classification, shape-specific extrapolation strategies extend the tactile map beyond the initial contact area, providing a computationally efficient estimate of potential contact without requiring complex physical simulations. The extrapolated regions were evaluated against measured contact data collected via a systematic grid-based scan using three metrics: the tactile centroid deviation, defined as the Euclidean distance between the geometric centers of binary contact regions; the grasp success rate estimated by a pretrained grasp assessment network; and the structural similarity index to assess local structural fidelity. Experiments on cuboidal, spherical, and cylindrical objects demonstrated the effectiveness of the approach in predicting unseen contact and identifying safe zones where extrapolated and real contacts align. The results showed reliable performance for small rigid objects, with the shape classifier achieving 94.3% accuracy on a held-out test set. However, a reduced accuracy was observed for larger, highly curved geometries, likely due to the limited curvature resolution of the tactile sensors. Large-diameter cylinders are occasionally misclassified as cuboids. Potential improvements include enlarging the dataset, refining the classifier, and integrating high-resolution sensors to enhance adaptability and precision.