EEG-Based Machine Learning Models for Emotion Recognition in HRI

This work attempts to provide a measure of the human affective state employing the analysis of electroencephalographic (EEG) activity to accurately identify the psychological state of the human during the interaction with a humanoid robot. Here, it’s presented the implementation and validation of different optimized classifiers such as Support Vector Machine, Decision Tree and Deep Neural Networks for estimating human affective state using the humanoid robot Pepper, equipped with two opposite personality configurations (positive and negative) to observe if and how a particular robot personality can affect the users’ affective response during the interaction. Affective state was estimated from physiological signals extracted from EEG relying on the two dimensional valence-arousal representation of emotions. The results show how feature selection improves the classification of the affective state of the individual as well as a fine process of hyperparameters optimization. Finally, a proof of the positive effect of a positive attitude of the robot in the interaction between human and robot, in terms of lower levels of negative emotion (such as Stress) recorded, is given.