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

MapAgent is an AI framework that automates lane-level map generation for autonomous driving at city scale, combining vision-language models with constraint verification to produce specification-compliant maps. Already deployed by Baidu Maps across 360+ Chinese cities, the system achieves over 95% production automation while reducing manual editing overhead in complex scenarios.

MapAgent addresses a critical infrastructure challenge in autonomous vehicle deployment: the labor-intensive process of constructing and maintaining accurate lane-level maps across multiple cities. Traditional end-to-end vectorized mapping methods struggle with ambiguous visual scenarios—worn road markings, occlusions, and complex intersections—leading to frequent human post-editing and specification violations. The framework innovates by coupling perception with explicit verification, introducing a Judge-Planner-Worker loop where a vision-language model diagnoses errors against specification rules, a planner generates minimal corrective edits, and workers implement validated changes. This architecture ensures compliance with traffic regulations while maintaining production scalability through selective triggering on low-confidence tiles only.

The deployment at Baidu Maps demonstrates immediate real-world validation and commercial viability. Supporting 360+ cities with 95% automation represents significant progress toward reducing mapping costs—a major barrier for autonomous vehicle companies globally. The framework's ability to handle long-tail and complex scenarios particularly matters as AV adoption expands into difficult geographic regions. For the autonomous driving industry, reliable automated mapping directly impacts deployment timelines and operational costs, affecting competitive positioning among vehicle manufacturers and robotaxi operators. The technical approach of combining symbolic reasoning (constraint verification) with neural perception models reflects broader industry trends toward hybrid AI systems that balance flexibility with explainability.

• →MapAgent achieves 95% production automation for lane-level map generation across 360+ Chinese cities through specification-driven verification loops.
• →The framework combines vision-language models with constraint-aware reasoning to handle ambiguous scenarios where visual evidence alone is insufficient.
• →Selective deployment on low-confidence tiles maintains scalability while reducing computational overhead for city-scale production systems.
• →Specification-compliant map production reduces human post-editing labor, a major cost driver in autonomous vehicle infrastructure development.
• →Successful integration with Baidu Maps validates the approach's practical effectiveness for commercial autonomous driving applications.