Detection of Error-Related Potentials during the Robot Navigation Task by Humans

Abstract

We have developed a system in which humans and autonomous robots can collaborate with each other. In the system, robots often exhibit behaviors not intended by the humans. To avoid this situation, it is necessary to convey the humans' will to the robots. To do this, we have focused on electroencephalogram (EEG) error-related Potential (ErrP), using which we can detect the ErrP when a person observes an error by a robot. In our previous study, we recorded the ErrPs from subjects in a maze task when a robot moved in directions that the subjects did not intend. However, the mean epoch number of the ErrP per subject was small. It is necessary to collect a large number of data using a deep neural network. Generally, medical data and physiological data recorded from people are small. Few Shot Learning is necessary for a small number of data. Thus, Siamese neural networks have been proposed. In this study, we combined the Siamese deep neural network with a support vector machine to discriminate between EEG data with an error (ErrP) and that without an error. Consequently, we could obtain >70% of the maximum classification accuracy among subjects and 0.60 ± 0.22 of the area under curve.