TPAM: A Tactile-Prior Adaptation Module for Multimodal Perception with a Low-Cost Visuotactile Sensor Prototype

Accurate tactile perception is crucial in human–robot interaction, rehabilitation therapy, and intelligent manipulation. While vision-based tactile sensors have advanced high-resolution surface imaging, they often struggle with pressure estimation and exhibit limited robustness under ambiguous contact conditions. To address these limitations, we propose a novel multimodal visuotactile sensing system that integrates gel-based deformation imaging with pressure measurement. The sensor features a low-cost, replaceable elastomer structure and provides real-time force data. On the algorithmic side, we introduce a human-inspired deep fusion framework that combines a learnable fuzzy membership mapping layer with a tactile-prior adaptation module (TPAM), enabling channel-wise modulation of visual features based on pressure semantics. This design emulates the human ability to interpret ambiguous contact states through adaptive multisensory integration. Experimental results based on multi-participant tactile data validate the system’s ability to accurately distinguish subtle variations in both contact geometry and indentation pressure, thereby demonstrating the effectiveness of the proposed human-inspired visuotactile fusion strategy for reliable, real-time tactile perception.