GATSim: Urban Mobility Simulation with Generative Agents

TL;DR

GATSim presents a generative-agent framework for urban mobility that integrates an urban mobility foundation model with a cognitive architecture featuring hierarchical memory and planning–reaction–reflection loops. It addresses limitations of rule-based ABMs by modeling diverse population profiles, evolving preferences, and socially-informed decision making, achieving realistic microscale behaviors and emergent macroscale traffic patterns. The work demonstrates competitive realism against human judges in role-playing scenarios and shows plausible traffic evolution and incident adaptation through multi-day learning. Its modular design and public prototype enable flexible scenario analysis for planning, policy testing, and crisis management in urban transportation contexts.

Abstract

Traditional agent-based urban mobility simulations often rely on rigid rulebased systems that struggle to capture the complexity, adaptability, and behavioral diversity inherent in human travel decision making. Inspired by recent advancements in large language models and AI agent technologies, we introduce GATSim, a novel framework that leverages these advancements to simulate urban mobility using generative agents with dedicated cognitive structures. GATSim agents are characterized by diverse socioeconomic profiles, individual lifestyles, and evolving preferences shaped through psychologically informed memory systems and lifelong learning. The main contributions of this work are: 1) a comprehensive architecture that integrates urban mobility foundation model with agent cognitive systems and transport simulation environment; 2) a hierarchical memory designed for efficient retrieval of contextually relevant information, incorporating spatial and temporal associations; 3) planning and reactive mechanisms for modeling adaptive mobility behaviors which integrate a multi-scale reflection process to transform specific travel experiences into generalized behavioral insights. Experiments indicate that generative agents perform competitively with human annotators in role-playing scenarios, while naturally producing realistic macroscopic traffic patterns. The code for the prototype implementation is publicly available at https://github.com/qiliuchn/gatsim.

Figures (11)

• Figure 1: GATSim prototype implementation. (a) Generative agent activity revision and en-route decision making. It demonstrates how agents dynamically adapt their activity plans and make real-time travel decisions based on current conditions, retrieved memories, and contextual information. (b) GATSim Prototype GUI. Users input the fork name, run name, and command, then click "submit" and "start" to initiate simulation. The interface displays network information, demographic data, and real-time agent plans and actions.
• Figure 2: Overall architecture of the GATSim framework, showing the interaction between the Urban Mobility Simulation Environment and the Generative Agents Module. The framework integrates three core components: the Urban Mobility Foundation Model for population synthesis and behavior generation, generative agents with cognitive architectures designed for urban mobility, and the simulation environment for system dynamics and user interaction management.
• Figure 3: Transportation network representation in the simulation environment. The hierarchical system uses graph-structured data for LLM-based scenario generation, tilemaps for map editing and animation, and bitmap visualization for end-user interaction. Transit networks are modeled following Spiess and Florian's approach with artificial boarding and alighting links.
• Figure 4: Generative agent activity planning process. Agents synthesize multiple information sources including the previous day's long-term reflection, initial activity plan, and daily reflection, combined with current broadcast events, retrieved relevant memories, and recent conversation summaries to generate today's updated long-term reflection, comprehensive activity plan and concepts to keep in mind.
• Figure 5: The agent action generation process, showing the flow from perception through reasoning to action execution. Plan revision occurs at strategically selected moments through four sequential stages: perception, memory retrieval, interaction, and reasoning with decision-making.