Toward Improved Child–Robot Interaction by Understanding Eye Movements

Abstract

Globally, 1 in 160 children has an autism spectrum disorder (ASD). Problems with joint attention (JA) are core features of ASDs. Here, we investigate how typically developing (TD) children and children with ASD initiate JA with a gaze contingent avatar. Thirty-one participants with ASD and 33 TD matched controls directed an avatar's gaze to a series of referent images. Observing pupil diameter and gaze location data, we explore how distinguishing the two groups as well as their different eye-movement behaviors could be used to improve child-robot interaction. With a sequence to sequence neural network we distinguish if a child is TD or has an ASD, then using K-means clustering, we group pupil diameters and gaze locations independently to determine the child's attention level as well as to refine the classification process. Using these metrics, we could trigger appropriate responses from the robot to increase the level of attention from the child toward the robot. Results show significant differences between the eye behaviors of individuals with ASDs and those without. Further to this, we achieve a 79.76% classification accuracy when using pupil diameter data to distinguish the two groups.