Data-Driven Investigation on Anisotropic Electrical Impedance Tomography for Robotic Shear Tactile Sensing

Abstract

Electrical Impedance Tomography (EIT)-based robotic tactile sensing holds great promise for future robot technology by geometrical scalability, mechanical durability, and ease of fabrication. The current EIT-based approach demonstrates wide success in measuring distributed normal tactile stimuli by reconstructing the scalar of the electrical conductivity. As an advancement, we propose a theoretical investigation on measuring conductivity tensor distribution using EIT, empowered by the deep neural network, that would facilitate a multi-modal tactile measurement on both the normal and shear inputs. An architecture based on Convolutional neural network and spatial sensitivity awareness on the loss function demonstrates the high-performing nonlinear tensor reconstruction that yet has not been proposed to date. Along with a sim-to-real approach for extensive investigation on conductivity tensor combinations, computationally efficient and robust (well-posed) model training was achieved, which results in a generalized reconstruction capability. Extensive simulation studies on reconstruction accuracy, noise robustness, and generalized reconstruction capability are presented. This preliminary investigation would be a substantial basis for constructing a new type of tactile sensor measuring distributed multi-modal tactile sensing.