Enhancing Robotic Tactile Exploration With Multireceptive Graph Convolutional Networks

While robotic tactile sensors have been developed to help robots to perceive and interact effectively with their surrounding environment by mimicking the structure and function of human skin, most of them overlook the role of near-contact behavior and data structure modeling in robotic perception, which limits robotic exploration capabilities. To address this problem, this article presents a novel proximity-tactile fingertip (PT-TIP) sensor, and a new multireceptive graph convolutional network (MR-GCN) that seamlessly integrates near-contact behavior and tactile perception in rich sensory data. Moreover, MR-GCN utilizes two graph structures, including topology graph and affinity graph, to capture temporal and spatial connections and differences among sensing units on PT-TIP, and it learns a robust feature representation from different receptive fields with attention mechanisms. The performance of MR-GCN was evaluated in two common robotic tasks, namely, object recognition and grasp stability detection, and the results show that the presented method outperforms state-of-the-art work in both tasks.