摘要
Understanding the physical properties of object surfaces is crucial for robotic systems to perform delicate operations and perceive the physical environment. Recent advances in optical sensor-based tactile imaging enable the capture of high-resolution contact information, demonstrating significant potential for multimodal robotic perception. Despite this progress, visual-tactile image representation learning remains under-studied. Our analysis explores two key challenges: visual and tactile images exhibit significant spatial inconsistencies, and tactile data shows more severe inter-class similarity and intra-class variation than common modalities. These factors limit the effectiveness of existing approaches based on direct instance-level alignment. To address these challenges, we propose a geometry-aware visual-tactile representation learning framework on a hyperspherical space. Specifically, we introduce a set of modality-agnostic class proxies that serve as shared semantic anchors to guide cross-modal alignment. Instead of enforcing direct pairwise correspondence, both modalities are organized around these proxies, enabling more stable and consistent semantic grouping. Furthermore, we impose a uniformity constraint on the proxies to explicitly regularize the global structure of the hyperspherical space, improving class separability. Extensive experiments on multiple visual-tactile benchmarks demonstrate that the proposed method consistently outperforms existing approaches and leads to well-structured and discriminative hyperspherical representations.