Abstract
Public urban spaces like streetscapes and plazas serve residents andaccommodate social life in all its vibrant variations. Recent advances inRobotics and Embodied AI make public urban spaces no longer exclusive tohumans. Food delivery bots and electric wheelchairs have started sharingsidewalks with pedestrians, while diverse robot dogs and humanoids haverecently emerged in the street. Ensuring the generalizability and safety ofthese forthcoming mobile machines is crucial when navigating through thebustling streets in urban spaces. In this work, we present MetaUrban, acompositional simulation platform for Embodied AI research in urban spaces.MetaUrban can construct an infinite number of interactive urban scenes fromcompositional elements, covering a vast array of ground plans, objectplacements, pedestrians, vulnerable road users, and other mobile agents'appearances and dynamics. We design point navigation and social navigationtasks as the pilot study using MetaUrban for embodied AI research and establishvarious baselines of Reinforcement Learning and Imitation Learning. Experimentsdemonstrate that the compositional nature of the simulated environments cansubstantially improve the generalizability and safety of the trained mobileagents. MetaUrban will be made publicly available to provide more researchopportunities and foster safe and trustworthy embodied AI in urban spaces.