Hypergraph Convolutional Networks Based Spatial Tactile Modeling for Object Geometric Property Recognition

This paper presents the application of Hyper- graph Convolutional Networks (HGCNs) for tactile spatial processing in multifingered robotic hands. Building on prior work employing Graph Convolutional Networks (GCNs) for modeling irregular sensor layouts, we address the architectural complexity introduced by topological segmentation approaches through the use of hypergraphs, which naturally capture higher-order relationships among tactile sensors. We evaluate HGCNs, standard GCNs, and feedforward neural networks (FNNs) on object geometric property recognition using eight objects and multimodal input (touch states, taxel coordinates, and joint angles). Our results demonstrate that HGCNs achieve high recognition rates of 96.61% while reducing model redun- dancy, and that hyperedge structure and types of hypergraph adjacencies significantly influence model performance. These findings suggest HGCNs offer scalable and effective tactile data processing.