Dynamic Tactile Sensor (DTS) With Data-Driven Super-Resolution for Edge Applications

Abstract

Tactile perception is of paramount importance for natural creatures and robotics. Challenges remain in realizing high-resolution and fast-response tactile perception with limited sensing units and computational resources. Here, we propose a data-driven super-resolution (SR) pipeline to increase the spatial resolution with limited discrete components, and reduce the data transmission load by developing a dynamic tactile sensor (DTS) architecture at the edge [specifically, on the sensor’s embedded microcontroller (MCU)]. Specifically, we proposed a tactile sensor based on micro-electro-mechanical systems (MEMS) barometer with soft contact interface, achieving a force sensing resolution of 0.01 N. A small amount of experimental data was collected from a taxel to refine a finite element method (FEM) model of the sensor. Enhanced by deep learning trained with simulated data only from the model, our sensor array achieved a 464-fold super-resolved accuracy. The fitting MSEs of the contact normal force and position in the polar coordinate system are 0.05 N and (0.11 mm, 0.09°), respectively. We further propose a DTS architecture for sensor array with multiple taxels, which achieved a data transmission reduction by 90%. With an entire model size of only 210 KB, an intelligent edge achieved an online recognition accuracy of 97.7% with a prediction time of 16 ms for finger interactions.