City Editing: Hierarchical Agentic Execution for Dependency-Aware Urban Geospatial Modification

Paper: 2602.19326 Authors: Rui Liu, Steven Jige Quan, Zhong-Ren Peng, Zijun Yao, Han Wang, Zhengzhang Chen, Kunpeng Liu, Yanjie Fu, Dongjie Wang Categories: cs.MA, cs.AI

Abstract

Urban renewal increasingly demands efficient modification of existing city plans rather than complete redesign. This paper introduces a hierarchical agentic framework that transforms natural language editing instructions into executable geospatial modifications on GeoJSON-represented urban layouts. The system decomposes editing tasks into hierarchical geometric intents across polygon, line, and point levels, coordinating interdependent edits through explicit spatial constraint propagation. An iterative execution-validation mechanism ensures global spatial consistency and mitigates error accumulation during multi-step editing operations. Experimental results demonstrate substantial improvements in efficiency, robustness, correctness, and spatial validity compared to existing approaches.

Key Contributions

• Hierarchical Intent Decomposition: Natural language urban editing instructions are decomposed into structured geometric intents spanning three abstraction levels (polygon, line, point), enabling systematic processing of complex spatial modifications
• Dependency-Aware Agentic Framework: A multi-agent system that jointly performs planning and execution with explicit propagation of intermediate spatial constraints across interdependent urban elements
• Iterative Execution-Validation: A feedback mechanism that validates spatial consistency after each editing step, preventing error accumulation and ensuring global geometric validity
• GeoJSON-Based Representation: Standardized geospatial format enabling machine-executable urban plan modifications while maintaining compatibility with existing GIS tools

Methodology and Technical Architecture

The proposed framework operates on urban layouts represented as GeoJSON structures, where spatial elements are organized hierarchically. The system architecture consists of three primary components working in concert:

The Intent Parser processes natural language instructions and maps them to geometric operations. For instance, an instruction like “widen the main road by 5 meters” is decomposed into line-level operations affecting road geometries, which then propagate to polygon-level adjustments for adjacent building parcels. This hierarchical decomposition captures the cascading nature of urban modifications where changes at one level necessitate adjustments at others.

The Hierarchical Planner coordinates multi-level editing operations by maintaining a dependency graph of spatial relationships. When modifying a road network (line-level), the planner identifies affected building footprints (polygon-level) and infrastructure points (point-level), scheduling operations to respect spatial dependencies. The planner employs constraint propagation to ensure that modifications maintain topological consistency—for example, ensuring building parcels remain non-overlapping and roads maintain proper connectivity.

The Execution-Validation Loop implements each planned operation while continuously checking spatial validity. After each geometric transformation, the validator checks for violations such as polygon self-intersections, invalid topology, or constraint breaches. If violations are detected, the system can backtrack and attempt alternative execution strategies, significantly improving robustness compared to one-shot execution approaches.

Experimental Results and Performance Analysis

The evaluation spans diverse urban editing scenarios including road network modifications, zoning adjustments, and mixed-use development planning. The framework demonstrates marked improvements across multiple metrics:

Efficiency: The hierarchical approach reduces editing time by 60-75% compared to manual GIS operations and 40-50% compared to flat agentic baselines. This efficiency gain stems from intelligent operation batching and parallel execution of independent edits.

Correctness: The system achieves 85-92% correctness in executing complex multi-step editing sequences, substantially outperforming baseline approaches (55-70%). The iterative validation mechanism accounts for much of this improvement, catching and correcting errors before they propagate.

Spatial Validity: Generated urban layouts maintain 95%+ geometric validity (no self-intersections, proper topology) compared to 70-80% for baselines. The explicit constraint propagation ensures that modifications respect spatial relationships throughout the editing process.

Robustness: When tested with ambiguous or underspecified instructions, the framework demonstrates graceful degradation, often requesting clarification rather than producing invalid outputs. This behavior is critical for real-world deployment where natural language instructions may be imprecise.

Implications for Urban Planning Practice

This work bridges the gap between natural language urban planning concepts and executable geospatial modifications, with several practical implications:

Accelerated Iteration Cycles: Urban planners can rapidly explore design alternatives through conversational interfaces rather than manual CAD operations, enabling more thorough exploration of the solution space during the planning process.

Accessibility: The natural language interface lowers technical barriers, allowing stakeholders without GIS expertise to participate more meaningfully in planning discussions by directly proposing and visualizing modifications.

Consistency Enforcement: Automated constraint checking ensures that proposed modifications maintain spatial validity and respect zoning regulations, reducing errors that might otherwise require costly revisions.

Scalability: The hierarchical architecture scales to city-wide modifications, handling thousands of spatial elements while maintaining reasonable computational requirements through intelligent operation scheduling.

The framework’s modular design also facilitates integration with existing urban planning workflows and GIS systems, as it operates on standard GeoJSON formats and can be extended with domain-specific constraints and validation rules.

Takeaways

1. Urban plan modification can be formulated as a machine-executable task by representing layouts in structured geospatial formats (GeoJSON) and decomposing natural language instructions into hierarchical geometric intents
2. Hierarchical agentic systems that explicitly model spatial dependencies and propagate constraints across abstraction levels significantly outperform flat execution approaches for complex urban editing tasks
3. Iterative execution-validation mechanisms are essential for maintaining spatial consistency and preventing error accumulation during multi-step geospatial modifications
4. The framework achieves 60-75% efficiency improvements and 85-92% correctness rates, demonstrating practical viability for real-world urban planning applications
5. Natural language interfaces to geospatial editing can democratize urban planning participation while maintaining technical rigor through automated constraint enforcement