Near-sensor reservoir computing for Braille recognition via high stability memristors

Abstract

Converting external physical information into tactile sensations for efficient dynamic processing like human beings is crucial for edge applications such as intelligent prosthetics and robotics. Reservoir computing, a bio-inspired computing paradigm, excels at processing temporal signals and offers advantages like low training costs and easy deployment on edge devices. Many applications have been developed for reservoir computing using physical devices. However, there has been a paucity of research using reservoir computing to simulate the human tactile system. Furthermore, the implementation of a reusable physical reservoir computing system is of significant importance. Herein, we implement a near-sensor physical reservoir computing system for haptic simulation, utilizing a simple peripheral circuit design. The reservoir's high-dimensional, nonlinear, and short-term memory requirements are physically realized by a memristor with an integrated lithium polymer electrolyte and polycrystalline tungsten oxide layer, which exhibits good cycle-to-cycle consistency. As a proof of concept, the system completes the learning and classification tasks for Braille numerals and characters, achieving a high recognition accuracy of up to 96% within 400 cycles. This approach offers innovative insights for developing human–machine interaction applications with enhanced intelligent perception capability.