Pressure-labelled tactile perception in robotics enabled by a skin-inspired dual-mode triboelectric sensor and neural networks

—Artificial intelligence (AI) enabled biomedical robots rely on tactile perception, but tactile sensing is limited by force induced variability that mixes material signals with contact pressure. Here, we introduce a pressure-labelled tactile intelligence framework built on a bioinspired tactile sensor (BT sensor) with a self-powered triboelectric element and fingertip like architecture, delivering decoupled dynamic tactile and static pressure information. The sensor integrates fast-adapting (FA) and slow-adapting (SA) elements based on contact electrification and electrostatic induction, respectively, for synchronous acquisition of transient responses and sustained pressure signals. A cross-talk-suppressed decoupling circuit separates triboelectric signals into dynamic and static components, enabling the SA output to serve as a reliable pressure reference for pressure labelled learning. Within a feature augmented framework combining physics informed descriptors with a one dimensional convolutional neural network (1D CNN), the pressure labelled strategy achieves 99.2% recognition accuracy across ten materials and elevates aggregate cross pressure recognition from 50.0% for dynamic-only sensing to near-perfect robustness, demonstrating suppression of force induced variability. Integrated with a robotic manipulator, it identifies three dimensional (3D) printed organ model materials with 100% accuracy. These results establish a controlled proof-of-concept for pressure-labelled tactile material recognition in anatomically relevant biomedical model systems.