Physics-Based Simulation and Machine Learning for the Practical Implementation of EIT-Based Tactile Sensors

This work reports an innovative framework toward the practical implementation of tactile sensors based on electrical impedance tomography. Particularly, we tested our framework on the development of a piezoresistive insole-shaped sensor prototype. Our approach couples the implementation of the forward model through a physics-based general purpose FEM software and an ANN model for the inverse problem solution. First, we developed a FEM forward model in COMSOL Multyphysics, and we optimized the parameters of the model to better resemble the prototype characteristics. Then, we trained an ANN model with an "artificial" dataset generated by feeding the forward model with a large set of different conductivity distribution samples and measuring the voltage at the boundary electrodes. For comparison, we employed the forward model also to compute the sensitivity matrix, which is required to apply standard linear reconstruction methods. We statistically compared the performance of the proposed machine learning approach with those of standard linear reconstruction methods. The results on simulated data highlight the higher accuracy of the ANN with respect to the other methods. In particular, the mean conductivity RMSE is 0.8 S/m. Finally, we tested our approach with the physical prototype to evaluate the performance in touch position detection. Again, the ANN achieves the minimum mean position error (5.74 mm), demonstrating the feasibility of using machine learning trained with artificial datasets for solving the inverse problem in EIT-based tactile sensors.