Simulation-Guided Autonomous Data Collection for Social Navigation in Multi-Story Buildings

Abstract

Autonomous robots deployed in buildings must be able to navigate through or around environmental features such as narrow/glass doors, elevators, and crowded passages. Failure to handle these situations can prevent a robot from completing its route, damage facilities, or endanger people. Additionally, directly testing navigational behaviors is risky because of unexpected perception, planning, or control errors that may cause unsafe robot behavior. To address these problems, we developed and tested a simulated Scout Mini robot capable of navigating complex buildings spanning multiple floors in the Gazebo simulator. We configured the simulated robot, sensors, localization, mapping, and navigation stack, then implemented behaviors based on detected door and elevator states, which we deployed onto the real Scout Mini. This simulation-based testing prepares the robot for a planned deployment in George Mason University’s FUSE building, where it will provide autonomous escorts and patrols to collect data including sensor readings, robot trajectories, and navigation states during interactions with humans. This work lays the foundation for collecting realistic autonomous social navigation data and developing robots capable of navigating complex, multi-story buildings safely and efficiently. By reducing reliance on manually driven data collection, autonomous deployment could enable social navigation datasets to be collected at a greater scale, allowing more data to be gathered with a wider variety of people, interactions, and environments.