Explainable Temperament Estimation of Toddlers by a Childcare Robot

Abstract

Personality estimation of others is a critical ability to communicate with each other. It enables robots to interact with humans and provides the former the ability to predict the intentions of the latter. Many researchers have developed personality estimation mechanisms. However, the estimation method for toddlers’ personality, such as the dominance of their innate temperament, has not been proposed yet. In this paper, we proposed an estimation model of toddlers’ temperament based on interaction data with a teleoperated childcare robot, ChiCaRo. The proposed method utilized the feature selection algorithm to increase estimation accuracy. Additionally, we employed an explainable AI model called Shapley additive explanations (SHAP) to understand which features from the interaction were important in terms of temperament estimation. The proposed estimation model demonstrated over 85% estimation accuracy for the average of all temperament factors. The experimental results of SHAP provided an understandable relation between the features and temperament factors and indicated that similar feature values from interaction videos used in child personality estimation could also be used for the temperament estimation of toddlers.