CityPlanner is a new framework for executable urban planning. It introduces UrbanSandbox, a file-based environment for agents to inspect tasks, generate plans, and revise decisions based on feedback. The framework uses atomic-task reinforcement learning, decomposing trajectories into BuildPlan for initial construction and ImprovePlan for iterative refinement. Experiments on a real-world benchmark show CityPlanner outperforms heuristic methods, task-specific RL, and general LLM-agent baselines. Ablation studies confirm the contributions of UrbanSandbox, atomic-task RL, and iterative deployment.
Source: https://arxiv.org/abs/2609.09578