Temperament estimation of toddlers from child–robot interaction with explainable artificial intelligence

Abstract

Personality estimation is a vital ability for communicating with others. It can help robots that interact with humans to model various human behaviors with a few parameters. Numerous studies have proposed models for estimating human personality from human–robot interaction. However, a limited number of methods have focused on the personalities of toddlers, which are dominated by innate temperaments. In this study, we propose a regression model that estimates the toddler temperament from images acquired by a teleoperated childcare robot named ChiCaRo. We gather a dataset from actual interactions between toddlers and ChiCaRo, and extract features from the data to train the regression model. Moreover, an explainable Artificial Intelligence model known as Shapley additive explanations (SHAP) is employed to understand the estimation tendency of the trained model and to compare the tendency with the temperament definition. The proposed model achieved a mean squared error of 0.024 for the average of all temperament factors. The analysis of SHAP confirmed that the model could reasonably learn the tendency compared to the definition in most temperament factors and suggested the possibility of data bias under a specific temperament factor.