Gaze estimation for social robotics: Bridging the gap: appearance-based gaze estimation models for human-robot interaction

A single glance can say a lot. It can show interest, attract attention, hint at what someone wants, or regulate turn-taking. Humans use gaze naturally in social life, and as social robots enter hospitals, classrooms, and public spaces, they need to understand this silent language too. A robot with eyes and a face that ignores gaze, or responds at the wrong moment, can feel strange and lose people’s trust. Accurate gaze estimation has traditionally relied on specialized hardware such as dedi- cated eye trackers, which are often expensive, intrusive, and difficult to deploy in natural human-robot interaction. This thesis asks whether appearance-based gaze estimation models, which infer gaze from ordinary camera images of faces, can offer a practical alternative for social robots. I evaluated appearance-based gaze estimation models against eye trackers and manual annotation, and tested them in human-robot interaction settings. The results show that within 2 meters, these models are already useful when model choice, camera resolution, and calibration are handled well. They can partly replace manual annotation and, with simple correction, match eye trackers for average gaze patterns, though fast attention shifts remain difficult to capture. I also found that tablets can increase task engagement but may also distract users. Finally, I built a real-time task-completion detection system, showing that effective gaze interaction is possible on standard robots using low-cost devices. Overall, this thesis demonstrates that AI-based gaze estimation cannot yet read minds and remains far from perfect. Still, it can serve as a practical component of human-robot interaction when gaze signals are interpreted in relation to task-specific social meaning, and when experimental design and parameter selection are carefully tailored. Future research should extend these capabilities to real-world and multi-user environments and integrate gaze with speech, gesture, and vision-language models, bringing gaze-aware robots closer to seamless social interaction in everyday settings.