Agentic Urban Science
智能体城市科学
Research on agent-based and AI-driven approaches to urban planning and design.
研究面向城市规划与设计的智能体与 AI 驱动方法。
2025 - present
AI has become strong at urban analysis—learning patterns from data to forecast traffic, land use, or environmental conditions. The next frontier is AI-assisted decision-making: systems that recommend sites, allocate resources, and evaluate trade-offs while reasoning transparently about constraints, regulations, and stakeholder values.
Agentic Urban Science explores autonomous agents and multi-agent architectures for urban planning, simulation, and design support. The central claim is that planning decisions require capabilities beyond statistical prediction alone: value-based reasoning, rule-grounded verification, and explainable justification when exploring alternative futures.
We articulate an Agentic Urban Planning AI Framework organised into three cognitive layers—Perception, Foundation, and Reasoning—and six logic components spanning analysis, generation, verification, evaluation, collaboration, and decision. Reasoning patterns from chain-of-thought prompting, ReAct, and multi-agent collaboration provide the technical substrate; the aim is to augment human planners, not replace them.
A complementary line of work applies retrieval-augmented generation to urban street networks: indexing spatial concepts over network structure so that language models can ground planning queries in the geometry and connectivity of real streets—bridging symbolic reasoning with the spatial fabric of the city.
Related empirical work is listed below.
AI 已擅长城市分析——从数据中学习模式以预测交通、用地或环境。下一前沿是AI 辅助决策：推荐选址、分配资源、评估权衡，同时透明地推理约束、法规与利益相关方价值。
智能体城市科学 探索面向城市规划、模拟与设计支持的自主智能体与多智能体架构。核心论点是：规划决策需要超越统计预测的能力——基于价值的推理、规则约束的验证，以及在探索替代方案时可解释的论证。
我们提出 Agentic Urban Planning AI Framework，组织为感知、基础与推理三层认知结构，以及涵盖分析、生成、验证、评估、协作与决策的六个逻辑组件。链式思考、ReAct 与多智能体协作等推理模式提供技术基底；目标是增强规划师，而非取代他们。
一条互补工作线将检索增强生成应用于城市街道网络：在网络结构之上索引空间概念，使语言模型能将规划查询锚定于真实街道的几何与连通性——连接符号推理与城市的空间肌理。
相关实证工作见下文。
