Selection of optimal hyperparameter for detecting multiple contacts from large-area tactile sensors based on electrical impedance tomography

Abstract

Large-area tactile sensors based on the technique of electrical impedance tomography (EIT) has drawn considerable interest in human-robot interactions. However, due to the ill-posed condition, it is challenging to differentiate between the real contacts and the artifacts from the reconstructed image. To address this issue, a new method to select an optimal hyperparameter that tunes the amount of regularization is developed in the context of tactile sensing. The optimal hyperparameter is determined to be the minimum value to obtain a stabilized number of sub-regions in the reconstructed image. The proposed method not only guarantees a correct detection on the number of multiple contacts at the minimum amount of regularization, but also provides a proper range of hyperparameters. The optimal hyperparameter is found in a chair-shape relation with the boundary signal-to-noise ratio (SNR), by varying the noise level of the hardware in simulation. The optimal hyperparameter decreases significantly when the boundary SNR increases between 5 ∼ 10 dB and 25 ∼ 35 dB, and keeps almost unchanged when SNR is between 10 ∼ 25 dB. The chair-shape relation also holds for contact conditions with varied intensities and sizes. Experimental validations on the proposed method are conducted on a compliant piezoresistive tactile sensor made of exfoliated graphite polymer composites. By varying the number of contacts in experiments, the relation between the optimal hyperparameter and the boundary SNR is consistent with the chair-shape curve. The investigation made in this work helps improve the performance of identifying multiple contacts from tactile sensors based on electrical impedance tomography.