2020/10/16 by Claudia Pérez-D’Arpino, Pérez-D'Arpino, Claudia, Can Liu +7 · 10 citations
Engineering · #Artificial Intelligence (cs.AI) #Evacuation and Crowd Dynamics #FOS: Computer and information sciences #Robotics (cs.RO)
paper · pdf · doi:10.48550/arxiv.2010.08600
openalex publication_date 2020/10/16 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Navigating fluently around pedestrians is a necessary capability for mobile\nrobots deployed in human environments, such as buildings and homes. While\nresearch on social navigation has focused mainly on the scalability with the\nnumber of pedestrians in open spaces, typical indoor environments present the\nadditional challenge of constrained spaces such as corridors and doorways that\nlimit maneuverability and influence patterns of pedestrian interaction. We\npresent an approach based on reinforcement learning (RL) to learn policies\ncapable of dynamic adaptation to the presence of moving pedestrians while\nnavigating between desired locations in constrained environments. The policy\nnetwork receives guidance from a motion planner that provides waypoints to\nfollow a globally planned trajectory, whereas RL handles the local\ninteractions. We explore a compositional principle for multi-layout training\nand find that policies trained in a small set of geometrically simple layouts\nsuccessfully generalize to more complex unseen layouts that exhibit composition\nof the structural elements available during training. Going beyond walls-world\nlike domains, we show transfer of the learned policy to unseen 3D\nreconstructions of two real environments. These results support the\napplicability of the compositional principle to navigation in real-world\nbuildings and indicate promising usage of multi-agent simulation within\nreconstructed environments for tasks that involve interaction.\n