Human Perception of Social Robots: Comparing Large Language Models vs. Psychology-Based Approach for Emotion Simulation

Abstract

Social Assistive Robots (SARs) are designed to act as companions in everyday life, relying on social interaction rather than physical assistance. To support this role, equipping SARs with emotional and personality expression capabilities has been shown to enhance human-robot interaction (HRI). While both psychology-based models and Large Language Models (LLMs) have been proposed for emotion generation, their impact on user perception in real HRI scenarios remains largely unexplored. In this work, we integrated two emotion generation approaches, a psychology-based framework (EmoACT) and an LLM-based model (GPT-4o), into a cognitive architecture for robotic personality, and conducted a betweensubjects study involving 56 participants to evaluate their impact on the perception of robot personality, cognitive and social abilities, and companionship-related traits. Results showed no substantial differences across most dimensions, suggesting that both approaches produce comparable interaction quality. Nevertheless, EmoACT enhanced the perception of Agreeableness, consistently with its theoretical grounding, while offering greater transparency and control over the emotion generation process. These findings suggest that both approaches are viable for emotionally capable SARs, with complementary advantages depending on the specific requirements of the application.