Development of a facial expression database covering diverse emotional states using large language models and an android robot

Facial expression databases are fundamental resources in many research areas. However, conventional databases built from human images are often limited in emotional diversity, with additional challenges such as high creation costs and difficulty in reproducing the expressions. In this study, we present a novel approach for constructing a wide-range facial expression database using large language models and an android robot. By leveraging the generative capabilities of large language models, our approach links semantic emotion descriptions to controllable robotic facial actuation, enabling the systematic exploration of diverse affective states. In addition, human-in-the-loop evaluation is incorporated to improve the alignment between generated facial expressions and emotion labels. The resulting database consists of 672 facial expressions corresponding to 75 emotion labels. Analyses of the database characteristics showed that the database captures a wider range of affective states that are broadly consistent with general emotional tendencies. It is expected to serve as a useful resource for future research in many fields, including information science and psychology.