Urban Generative Intelligence (UGI): A Foundational Platform for Agents in Embodied City Environment

1 Introduction

• Urban challenges span interconnected physical, social, economic, and environmental systems, including congestion, pollution, inequality, and housing issues.Source: §1 · p. 1
• Despite advances from big data and AI to urban computing and digital twins, practical implementation for systemic urban challenges remains a gap.The introduction frames this as a shortfall in addressing the complexity of urban issues.Source: §1 · pp. 1–2
• UGI combines UrbanKG, a city simulator, and a natural-language interface to support embodied agents that interact with an urban environment.The infrastructure provides simulated interactions and embodied feedback to agents.Source: §1 · p. 2
• CityGPT is trained with general and city-specific data and serves as the foundation model for agents across urban tasks.The proposed agents include mobility, economic activity, and social interaction simulations, as well as location recommendation and schedule-planning assistants.Source: §1 · p. 2
• The proposed evaluation framework assesses foundation models and agents across knowledge mastery, simple reasoning, planning, and decision-making.Source: §1 · p. 2

2.1 Complex Urban System

• Earlier generative agents were studied in simple virtual settings, leaving their ability to interact with complex urban environments unclear.The paper frames exposure to rich urban information and realistic simulators as an open research direction.Source: §2.1 · p. 3

2.2 Urban Computing and Intelligence

• Urban computing advanced from data analytics and prediction toward behavior simulation and decision-making, but existing methods often simplify real-life problems.The paper argues that LLM integration may support more systematic pattern discovery, dynamics prediction, and decision-making.Source: §2.2 · pp. 3–4

2.3 Digital Twin City

• Digital Twin City research uses data-driven digital emulations for sensing, analysis, and optimization, yet faces challenges in multi-source data realism, real-time updates, and human-friendly interaction.These constraints matter for scenario testing and decision-making by policymakers and urban planners.Source: §2.3 · p. 4

2.4 Large Language Models and Agents

• Prior LLM-based urban work often focuses on city-related concepts or tools, while embodied agents have generally been tested in simple environments such as rooms or virtual games.The paper proposes a more complex city environment for agents with a broad range of human-like embodied behaviors.Source: §2.4 · pp. 4–5

2.5 Agent-based Modeling and Simulation

• Conventional agent-based simulations progressed from rules to neural networks and reinforcement learning, but agents still require human-defined goals or rules.UGI proposes LLM-driven agents for simulating city dynamics and supporting applications such as decision-making.Source: §2.5 · p. 5

2.6 Metaverse

• The paper positions its foundational platform as a basis for a future city metaverse with embodied agents and urban generative intelligence.Source: §2.6 · p. 6

Architecture overview

• The architecture assembles a city simulator, UrbanKG, and diverse data streams as open digital infrastructure for generative models and agents.Models can access computation and factual knowledge, test strategies in simulated scenarios, and evolve based on feedback.Source: §3 · pp. 6–7

Open Digital Infrastructure

• The infrastructure collects multi-source spatial-temporal urban data, uses Mirage to simulate human-thing-space interactions, and uses UrbanKG to fuse data and provide factual knowledge.Source: §3 · p. 7

Language Interface

• A standardized natural-language protocol lets models and agents access simulator computation and UrbanKG knowledge without customized integration procedures.The interface is intended to reduce development barriers for language-model-driven agents.Source: §3 · p. 7

Generative Intelligence

• CityGPT encodes local urban knowledge, and the proposed agents target mobility, economy, community, and society tasks including decision-making and agent-based simulation.Source: §3 · p. 7

• The infrastructure plan progresses from urban modeling and multi-source real-data streams to a simulator and interfaces for agents and human users.The simulator is designed to support mobility, social, and economic simulations.Source: §4 · p. 7

• Areas-of-interest (AOIs) represent city areas such as commercial, industrial, residential, and public-service land, connected by a road network.Source: §4.1 · p. 7
• Points-of-interest (POIs) represent urban functions, with examples including restaurants, hotels, and auto repair stores.Source: §4.1 · p. 7
• The urban model combines spatial structure and function through roads, junctions, AOIs, and POIs.Source: §4.1 · p. 8
• Human activity is modeled as movement across roads and AOIs, visits to POIs over time, and peer-to-peer or peer-to-cluster messaging.Social relationships may be predefined through online networks or obtained from spatial proximity.Source: §4.1 · p. 8
• Infrastructure networks are modeled as AOI-vertex topologies, while economic systems include companies, individuals, governments, and banks engaging in consumption, wages, taxes, and interest.Source: §4.1 · p. 8
• The combined modeling describes urban mobility, socialization, and economic behavior to guide real-data-stream construction and city-simulator implementation.Source: §4.1 · p. 8

4.2.1 Databases

• Road and area-of-interest data are cleaned and enriched with road attributes, land use, population, and road connections.Road topology and junction types are derived from OpenStreetMap; AOI polygons are linked to nearby roads.Source: §4.2.1 · p. 9
• POIs are matched to AOIs, connecting urban functions to city space, while UrbanKG supplies spatio-temporal correlations and data fusion.The linked structure can represent activities such as going for leisure and driving to a park.Source: §4.2.1 · p. 9
• Sampled travel records support models that generate citywide human mobility patterns, while person and trip records represent individual movement.A trip records start and end positions, start time, and mobility mode.Source: §4.2.1 · p. 9
• Public online sources provide enterprise, industry, wage, consumption, and rent information that is added to AOIs.Enterprise records include category, registered capital, employee count, and average wage.Source: §4.2.1 · pp. 9–10
• Satellite and street-view images extend the city database with multimodal data, and processed modeling data is stored for simulator input or agent API access.Street-view coordinates are spaced at 100-meter intervals.Source: §4.2.1 · pp. 9–10

4.2.2 Knowledge Graphs

• UrbanKG organizes city-entity relations across spatial structure, urban functions, and image links to support agent information inputs.Its relations include nearBy, locateAt, competitive, coCheckin, satelliteImageOf, and streetViewOf.Source: §4.2.2 · p. 10

• The simulator lets agents move between roads and AOIs, sense current conditions, query places, and control behavior by editing trip lists.Agents can use these interfaces to locate a restaurant-type POI and travel to its AOI.Source: §4.3 · pp. 10–11
• It supports driving, walking, biking, and public transportation, using IDM and MOBIL for driving and PCS for walking.The simulation also integrates infrastructure-network extensions for power, water, drainage, and mobile communications.Source: §4.3 · pp. 10–11
• More than 10 times wall-clock acceleration is reported for nearly one million agents at urban scale.The simulator attributes this to control-flow and data-flow design and an indexing subsystem.Source: §4.3 · p. 10
• Runtime data retrieval supports pull and push access to spatial structure and agent status, while a unified control interface modifies trips.The simulator synchronizes its clock with accessed agents.Source: §4.3 · pp. 10–11
• Social messaging and financial flows extend mobility simulation to urban life activities and provide an environment for agent exploration, learning, and evaluation.Financial flows include consumption, income, and taxes, triggered primarily by visits to POIs.Source: §4.3 · p. 11

4.4.1 Application Programming Interface

• The API offers Protobuf and gRPC-compatible communication alongside a higher-level Python interface for data access, sensing, and agent control.The Python API returns familiar structures such as dictionaries and lists.Source: §4.4.1 · p. 11
• Clients can retrieve AOI, road, and person status and control agents by changing trip lists.SetTrips changes a person’s moving target, departure time, and mobility mode.Source: §4.4.1 · p. 11
• Push notifications report events such as entering or leaving an AOI or road and starting or finishing a trip, avoiding excessive polling.A client can use an AOI-entry notification to control a person’s subsequent behavior.Source: §4.4.1 · p. 12

4.4.2 Natural Language Interface

• The natural-language interface wraps the APIs so users can request data retrieval and agent control with standardized instructions.Examples include querying an AOI’s attributes and setting an agent’s destinations, times, and travel modes.Source: §4.4.2 · p. 12

5.1 CityGPT: Large Language Model for Urban Generative Intelligence

• CityGPT is built through continued pre-training on urban knowledge followed by supervised fine-tuning and DPO alignment for urban tasks and instruction-following.The alignment data include general chat, domain-specific chat, and task-solving examples such as schedule planning and navigation.Source: §5.1 · pp. 13–14

5.2 General Framework for Generative City Agents

• The embodied-agent framework combines memory, persona, and preference with perceive, act, and communicate modules around a city foundation model.Agents receive embodied feedback from the city simulator and UrbanKG infrastructure.Source: §5.2 · p. 15

5.2.1 Simulation Agent: Generating Individual and Collective Behaviour

• Mobility agents use historical movement, demographic profiles, and perceived context to generate behavior, with routine anchors helping preserve daily schedules.Context can include current time and road traffic; generated locations are registered in the simulator.Source: §5.2.1 · p. 15
• Economic simulation models labor, consumption, financial markets, and taxation through household, government, and bank agent decisions.The system observes overall labor and consumption market dynamics.Source: §5.2.1 · p. 16
• Social agents model online and offline interaction, information propagation, and individual attitudes and emotions; S3 reports simulations of propagation on nuclear-energy and gender-discrimination events.Evaluation can consider both individual-level and population-level behavior.Source: §5.2.1 · p. 16

5.2.2 Decision Making Agent: Task Solving and Personal Assistance

• Location-recommendation agents interpret user preferences, organize search results, adapt recommendations, and explain them using historical behavior, demands, or spatial-temporal context.They are also described as identifying improper user demands relative to the real-world city environment.Source: §5.2.2 · pp. 16–17
• Schedule-planning agents use CityGPT and city feedback to form personalized plans that respect fixed commitments, preferences, and time constraints.The agent evaluates feasibility using travel time and proximity and can adjust plans as new information arrives.Source: §5.2.2 · p. 17

Level 1: Automated domain-specific question answer evaluation

• Level 1 tests CityGPT’s learned urban knowledge with multiple-choice question answering, using accuracy as the metric.Questions come from general evaluation datasets and ChatGPT-generated pairs based on the pre-training corpus.Source: §6 · pp. 17–18

Level 2: Human labeled domain-specific question answer evaluation

• In human-labeled urban questions, ChatGPT cannot solve the presented problems, while domain-enhanced CityGPT solves them easily.The questions cover commonsense knowledge and route planning.Source: §6 · p. 18, Figure 8 · p. 18, Figure 9 · p. 19

Level 3: Problem solving of various urban tasks

• Level 3 evaluates agents on four urban tasks, measuring overall success across hundreds of samples per task.Tasks are next-location prediction, PoI navigation without a map, daily schedule planning, and society simulation; success criteria vary by task.Source: §6 · p. 18

7.1 Transportation System

• A generative mobility agent aims to produce realistic, intention-aware travel behavior as an efficient alternative to costly, infrequent travel surveys and rule-based generators.The cited rule-based approach models movement among frequent locations but lacks an in-depth understanding of mobility intent and user profiles.Source: §7.1 · p. 19

7.2 Business Intelligence

• The site-selection model combines knowledge graphs with CityGPT’s reasoning and common-sense capabilities to produce precise, interpretable results with improved efficiency and breadth.It addresses limitations in feature representation and reasoning in data-driven models, as well as knowledge graphs’ challenges with varied urban data.Source: §7.2 · pp. 19–20

7.3 Urban Economy System

• The economic simulation is presented for exploring macroeconomic behavior and activities and testing policy scenarios and their potential effects on economic indicators.It is framed as a controlled environment for assessing trade-offs and anticipating unintended consequences before policy implementation.Source: §7.3 · p. 20

7.4 Urban Society

• Social simulation supports studying individual behavior and population dynamics, testing interventions, and exploring emergency and risk-management scenarios when real-world emergency data is sparse.Simulated settings let researchers and practitioners examine potential effects of interventions and policies before real-world deployment.Source: §7.4 · pp. 20–21

8.1 Dive into Complicated Urban Issues

• UGI’s real-urban environment and embodied agents are intended to provide context-aware insights, but deployment, adaptation to changing urban conditions, and integrated understanding of urban systems remain challenges.The discussion also calls for solutions that address rapid urbanization and climate change.Source: §8.1 · p. 21

8.2 Scale Up to Large City

• Large-scale LLM-agent societies face substantial computational challenges, and full-process acceleration remains unresolved.The discussion notes techniques such as batch prompting, which it reports can improve inference time and cost by up to 5x.Source: §8.2 · pp. 21–22

8.3 Openness of the Environment

• Users can modify or create city environments by changing spatial data, networks, infrastructure, images, and human-activity inputs or algorithms.The discussion presents this openness as supporting work on agents as well as planning, design, and governance.Source: §8.3 · p. 22

8.4 Developer Community

• UGI’s development calls for a multidisciplinary community in which data, simulation, language-model, and urban-domain researchers contribute distinct resources and agents.Urban researchers can build domain-specific agents through programming or a natural-language interface.Source: §8.4 · p. 22

9 Conclusion

• UGI integrates LLMs with urban data and digital twins in a platform for developing and deploying embodied agents across social, economic, and environmental challenges.The agents are empowered by CityGPT, and the platform is presented as bridging technological capabilities and practical urban-system applications.Source: §9 · p. 22
• The conclusion positions UGI as a new paradigm for generative intelligence in urban space and for modeling complex urban systems.It links this approach to intelligent, sustainable, and resilient urban development.Source: §9 · p. 22
• The platform’s stated aim is to support future cities that are more adaptive and responsive to inhabitants’ evolving needs.Source: §9 · p. 22