Abstract
Tactile sensors provide useful contact data during the interaction with anobject which can be used to accurately learn to determine the stability of agrasp. Most of the works in the literature represented tactile readings asplain feature vectors or matrix-like tactile images, using them to trainmachine learning models. In this work, we explore an alternative way ofexploiting tactile information to predict grasp stability by leveraginggraph-like representations of tactile data, which preserve the actual spatialarrangement of the sensor's taxels and their locality. In experimentation, wetrained a Graph Neural Network to binary classify grasps as stable or slipperyones. To train such network and prove its predictive capabilities for theproblem at hand, we captured a novel dataset of approximately 5000three-fingered grasps across 41 objects for training and 1000 grasps with 10unknown objects for testing. Our experiments prove that this novel approach canbe effectively used to predict grasp stability.