Classifying Engagement in School Age Children during storytelling activities with NAO robot

Modelling effective human-robot interaction requires the robot to detect and respond to engagement signals, which include attention, interest, and empathy. This study explores the application of automated gaze-labelling techniques, among the use of other measures connected to the emotional, cognitive and behavioural engagement constructs, to classify the engagement in child-robot interactions using the NAO robot in a storytelling paradigm. A total of 102 children, aged 7 to 9, participated in structured individual sessions recorded for gaze analysis and engagement scoring. Engagement measures were manually annotated by observers using the Engagement Observation Scale. Gaze-360 and K-means clustering were used for automated gaze labelling. According to the assigned engagement score, 38 children were included in the study and divided into low/high engagement group. Fifteen features were extracted and organized in 4 datasets considering the engagement constructs (i.e. affective, behavioural and cognitive). These datasets were classified with 4 machine learning techniques namely Support Vector Machine (linear and quadratic kernel), Decision Tree and K-Nearest Neighbour according to the high/low engagement rate. Results indicate an engagement recognition accuracy >89% for the best configuration that includes features related to the engagement of all constructs. This result suggests that the set of extracted features can be used to classify the level of engagement.