Understanding Collision Avoidance Cues in Virtual Humans

Abstract

In the real world, humans can navigate through crowded spaces without collision because they are able to predict others’ intentions and avoidance strategies. However, navigating among autonomous virtual humans in Virtual Reality (VR) can be challenging due to the absence of anticipatory cues that communicate turning intention. In this paper, we study which body-part kinematic cues are relevant using both a human perceptual experiment in Virtual Reality (VR) and a Reinforcement Learning (RL) analysis sharing the same task formulation. First, we examine participants’ gaze behavior while predicting the avoidance direction of an approaching virtual pedestrian animated using full-body motion capture and gaze-tracking data from a real person. Second, we train RL agents on the same task and systematically ablate input features to quantify the contribution of each kinematic feature. Our gaze analysis shows that participants predominantly attend to the head and upper body, while lower-body fixations are associated with prediction errors. Our RL ablation identifies hip position as the dominant cue for accurate prediction, with head rotation providing a secondary but significant contribution to early decisions. Integrating both analyses, we identify a two-phase cue structure: an early orientation phase driven by head rotation, followed by a displacement phase driven by lateral hip movement. Our results show that realistic movement and gaze cues enable virtual humans to convey turning intentions, highlighting the need to incorporate these cues into simulation models. Moreover, the strong correlation between RL agents and perceptual results suggests that RL can help validate collision-avoidance models and support more natural human–agent interactions in virtual environments.