Real-Time Social Robot’s Responses to Undesired Interactions Between Children and their Surroundings

Abstract

Aggression in children is frequent during the early years of childhood. Among children with psychiatric disorders in general, and autism in particular, challenging behaviours and aggression rates are higher. These can take on different forms, such as hitting, kicking, and throwing objects. Social robots that are able to detect undesirable interactions within its surroundings can be used to target such behaviours. In this study, we evaluate the performance of five machine learning techniques in characterizing five possible undesired interactions between a child and a social robot. We examine the effects of adding different combinations of raw data and extracted features acquired from two sensors on the performance and speed of prediction. Additionally, we evaluate the performance of the best developed model with children. Machine learning algorithms experiments showed that XGBoost achieved the best performance across all metrics (e.g., accuracy of 90%) and provided fast predictions (i.e., 0.004 s) for the test samples. Experiments with features showed that acceleration data were the most contributing factor on the prediction compared to gyroscope data and that combined data of raw and extracted features provided a better overall performance. Testing the best model with data acquired from children performing interactions with toys produced a promising performance for the shake and throw behaviours. The findings of this work can be used by social robot developers to address undesirable interactions in their robotic designs.