A Centipede-Inspired Flexible Capacitive Sensor Integrated with Deep Learning for Real-Time Gesture Translation via Bimodal Time-Series Imaging

Abstract

The demand for intelligent wearable human–machine interaction (HMI) systems is rising with the rapid advancement of flexible sensors and artificial intelligence. However, flexible capacitive sensors face challenges such as long fabrication cycles, high costs, and insufficient stability. To address these limitations, this study proposes a low-cost, scalable fabrication method inspired by the multilegged structure of centipedes. The sensor was fabricated using commercially available, inexpensive modified polymer materials through a simple assembly process, and exhibits high reliability over 10,000 cycles, sensitivity (1.69% kPa^–1, 0–20 kPa), fast response (37 ms), low hysteresis (7.02%), and robust performance under varying conditions. A real-time gesture translation system based on a smart glove was developed, which employs an improved Gramian angular field (GAF) method to convert gesture signals into dual-modality images. Integrated with MobileNetV2 and EfficientNetB1 deep learning models, the system achieves 99.73% average recognition accuracy for 25 sign language gestures with a 54.29 ms delay. The smart glove also enables wireless control of a bionic robot hand. This study provides a practical approach for fabricating flexible capacitive sensors and integrating them into real-time gesture recognition systems, offering significant value for hearing-impaired communication and potential applications in motion monitoring, underwater communication and sensing, and HMI.