Assessing pedestrian responses to autonomous and personal mobility robots in crowded public spaces

Robots increasingly share spaces with people, supporting delivery services and mobility for the aging populations, yet their ability to share space comfortably lacks understanding and benchmarks for designers and policy-makers. We compared human-robot (HRI) and human-human (HHI) interactions across four real-world crowd datasets spanning Europe, North America, and Asia, using a unified pipeline to detect interactions, stratify by crowd density, and model pedestrian behavior. Local motion patterns (speed, acceleration, and jerk) remained closely matched between HRI and HHI across all densities. In contrast, proxemics diverged, with effects that grew approximately linearly with robot speed and weakened under higher crowding: In the dataset with the faster navigating robot, pedestrians maintained about 0.23 meters more clearance around the robot than around other pedestrians under less crowded conditions and about 0.05 meters more under more crowded conditions, while in the dataset with the predominantly stationary robot, the corresponding differences were small and inconsistent across crowding levels. The main conclusions were robust to parameter variations and remained stable across a broad range of motion-processing and interaction-labeling settings. Our findings provide density- and speed-aware benchmarks for proxemics in social robot navigation and empirically grounded targets for design, evaluation, and modeling across robotics and urban mobility.