CityPlanner improves urban planning with interactive sandbox agent
CityPlanner: A Sandbox Agent for Executable Urban Planning
Artificial IntelligenceComputation and Language
Summary
Urban planning involves choosing the best ways to design cities by balancing costs and services, which is very complex. The authors created CityPlanner, a tool where an agent uses a sandbox environment to try out city plans, check how good they are, and then improve them step-by-step. To make this easier to learn, the system breaks the planning into two parts: building an initial plan and then refining it based on feedback. Tested on real city data, CityPlanner performs better than other existing methods for urban planning.
What this means in practice
- •For urban planners: Generate and iteratively refine city development plans using executable feedback to balance cost and service quality efficiently.
- •For software engineers: Build interactive planning tools that integrate task files and feedback loops for flexible urban design workflows.
Authors
Wentao Zhang, Jingyuan Wang, Zetong Zhou, Yifan Yang, Wenrui Wang
Abstract
Urban planning is a real-world spatial optimization problem that requires selecting feasible actions from large candidate spaces under practical objectives such as cost and service quality. Existing optimization and reinforcement learning methods are effective for fixed formulations, but often depend on task-specific representations and constraint handling. We propose \emph{CityPlanner}, a sandbox-agent framework for executable urban planning. CityPlanner introduces \emph{UrbanSandbox}, a unified file-based environment where agents inspect task files, generate plans, run evaluators, and revise decisions based on executable feedback. To make learning tractable, we further propose atomic-task reinforcement learning, which decomposes long sandbox trajectories into \emph{BuildPlan} for initial construction and \emph{ImprovePlan} for feedback-based refinement. Experiments on a real-world benchmark show that CityPlanner consistently outperforms heuristic, task-specific RL, and general LLM-agent baselines. Ablations verify the contributions of UrbanSandbox, atomic-task RL, and iterative deployment. We release the code and dataset at https://anonymous.4open.science/r/co-agent-C1C8