Title: MapAgent: An Industrial-Grade Agentic Framework for City-scale Lane-level Map Generation

Abstract

Standardized lane networks are essential infrastructure for autonomous driving and precise, lane-level navigation. However, the manual effort required to build and maintain these networks across hundreds of cities is substantial. While recent end-to-end vectorized mapping approaches can directly derive lane geometry and topology from sensor inputs, they often rely on implicit, dataset-specific supervision for mapping specifications and traffic rules. Consequently, in complex environments characterized by worn markings, missing lines, or occlusions, visual data alone frequently fails to uniquely determine correct lane configurations. This ambiguity leads to specification violations, which constitute a significant portion of the workload for human post-editing.

To address these challenges, we introduce MapAgent, an industrial-strength agentic framework designed to enhance a vectorization backbone for the production of specification-compliant lane maps. Unlike approaches that simply append an agent loop to prediction models, MapAgent integrates the backbone’s perception capabilities with explicit specification verification, constraint-aware reasoning, and deterministic map editing. This process operates within a bounded, verification-driven "Judge-Planner-Worker" loop. Within this architecture, a vision-language Judge identifies errors by analyzing both visual evidence and draft vector outputs. Simultaneously, a tool-calling Planner formulates minimal corrective edits, which are subsequently re-validated.

To ensure scalability for city-wide production, MapAgent is deployed selectively, activating only on tiles where the backbone’s confidence is low. This targeted approach incurs minimal overhead while maintaining high throughput. Evaluations on real-world datasets demonstrate that MapAgent consistently outperforms robust production baselines, with particularly notable improvements in complex and long-tail scenarios. Furthermore, MapAgent has been successfully integrated into Baidu Maps, facilitating lane-level map generation for more than 360 cities across the country. This deployment has raised the overall production automation rate to over 95%, underscoring the framework’s practicality and efficacy for large-scale lane-level map generation.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC