Saccadic undershooting in gaze generation for virtual characters

Saccades are the primary eye movements used to shift gaze and play an important role in generating realistic gaze behavior for virtual characters. While many gaze generation studies focus on high-level attention or target selection, finer characteristics of human eye movements–such as saccadic undershooting–have received relatively little attention. In this paper, we propose a psychologically plausible gaze generation model that explicitly incorporates saccadic undershooting to produce more human-like gaze behavior. The model parameters were derived from eye-tracking data collected in a VR environment and integrated into a gaze generation framework for virtual characters. Quantitative evaluation using leave-one-participant-out cross-validation shows that the proposed model reproduces the undershooting patterns observed in human data and achieves lower absolute errors compared with a representative existing model and a random baseline. A subjective user study further indicates that participants can perceive differences between gaze behaviors generated by the models. Although no significant differences were found in median ratings of human-likeness, roboticness, or head–eye coordination, the bimodal rating patterns suggest that subtle variations in gaze behavior may influence users’ perceptions of virtual characters.