Assessing Last-Mile Space–Time Accessibility with Barriers via Generative Navigational Agents

Abstract

Conventional last-mile accessibility assessments often assume isotropic motion, neglecting the impact of physical barriers and cognitive pathfinding on accessibility. This research investigates the last-mile mobility by assessing how urban barriers and frictions disrupt last-mile accessibility. The study presents a hybrid framework integrating high-resolution LiDAR data with large language modeling (LLM). LiDAR data locates barriers and measures barrier impedance, while LLMs serve as generative navigational agents to simulate barrier-adaptive decision-making. This two-tiered methodology transforms unprocessed spatial limitations into accessibility metrics within space-time prisms. Findings based on two comparative case studies reveal that conventional models omitting physical barriers have significant accessibility overestimations, which advocates for human-centered designs in non-idealized urban areas.