Adaptive Gaze Modulation in Social Robots: A Reinforcement Learning Approach to Attention Regulation

Attention serves as a critical antecedent to social presence, which fundamentally influences acceptance, trust, and overall interaction quality in human-robot interaction (HRI). This paper investigates the development of a gaze modulation framework that enables robots to strategically influence human attention through two complementary Q-learning-based modules: Gaze-Garnering Modulation (GGM) and Gaze-Avoidance Modulation (GAM). To measure gaze feedback, we introduce a novel metric—the Dynamic Gaze Engagement Index (DGEI)—that integrates attention ratio with stationary gaze entropy (SGE) to evaluate not just the quantity but also the quality of visual attention. This feedback allows the system to continuously adapt to each individual’s unique attentional patterns and thresholds, providing personalised interaction. In two experiments, 20 participants interacted with a Pepper robot that dynamically adjusted its behaviours (lights, movements, and voice volume) based on real-time gaze feedback. Results demonstrated that GGM significantly enhanced gaze engagement, fostering strong mutual interaction, while GAM effectively redirected attention when appropriate, with participants reporting lower perceived gaze engagement in this condition. Post-experiment questionnaires using the "Psycho-behavioural Interaction - Perceived Attentional Engagement" section of the Networked Minds Social Presence Inventory (NMSPI) revealed significant differences between conditions (t(18)=2.47, p=0.0238), validating the attention modulation by each module and corroborating the behavioural observations. These findings underscore the importance of adaptive robotic behaviours in facilitating dynamic and unobtrusive interactions.