@inproceedings {pub6212,
	title = {Overcoming Social Perception Challenges in Nonverbal Human-Robot Communication},
	author = {Yu Fang AND Matti Kr{\"u}ger},
	year = {2025},
	month = {August},
	abstract = {Human eyes serve both perceptual and expressive functions. Their orientation towards regions of interest underlies selective perception and thus simultaneously provides insights into spatial attention and intention in social perception. Eye contact is a fundamental cue in social perception, enabling attention exchange. Thus, the use of eye contact through artificial eye models is being explored as a communication tool in human-robot interaction. However, screen-based eye models, in particular, can be ambiguous as spatial pointers, making nonverbal communication that relies on spatial referencing, such as eye contact, challenging. This challenge can affect the clarity and effectiveness of robotic eye gaze in social communication.

We propose two approaches to improve robotic eye gaze spatial referencing:  (1) attention model-driven eye-head gaze movements and (2) a reflection-like feature for screen-based eye models. In the first approach, a forehead-mounted camera synchronizes the robot{\textquoteright}s visual field with head movements, ensuring stable subjective vision. An attention-retina model generates eye movements by integrating visual attention with retinal structure, while head movements replicate human-like eye-head coordination. The second approach extends a screen-based eye model with a dynamic reflection-like overlay that is contingent on eye movements. This overlay complements pupil-based gaze direction displays with contextual visual cues to improve spatial reference identification. 

User studies combining subjective feedback and objective measurements about the use of referential gaze indicate that both approaches significantly enhance the robot{\textquoteright}s ability to convey referential eye gaze. Participants reported improved clarity in perceiving the robot{\textquoteright}s gaze direction and intent (F(3,80)=197.68, p<0.001 for approach 1 and F(2,87)=9.93, p<0.001 for approach 2). At the same time, quantitative analysis demonstrated increased accuracy in spatial reference identification (F(2,168)=14.99, p<0.001). 

These findings highlight the prospect of integrating visuomotor behavior with interactive context in social nonverbal communication. They contribute to advancing socially intelligent robots with enhanced social perception, strengthening their ability to engage effectively in human-robot interaction.},
	publisher = {SAGE},
	booktitle = {European Conference on Visual Perception}
}