CityPlanner Introduces a Sandbox Agent for Executable Urban Planning

Summary

Urban planning requires choosing feasible actions from large candidate spaces while balancing objectives such as cost and service quality. The CityPlanner study presents a sandbox-agent framework designed to make these decisions executable and revisable. Its UrbanSandbox is a unified file-based environment: agents inspect task files, generate plans, run evaluators, and revise their decisions using executable feedback. The authors also introduce atomic-task reinforcement learning, which breaks long sandbox trajectories into two stages: BuildPlan for initial plan construction and ImprovePlan for feedback-based refinement. On a real-world benchmark, CityPlanner consistently outperformed heuristic methods, task-specific reinforcement learning, and general LLM-agent baselines. Ablation studies supported the contributions of UrbanSandbox, atomic-task reinforcement learning, and iterative deployment. The authors state that they release the code and dataset.