How UniMapGen Revolutionizes Large‑Scale Lane‑Level Map Generation with Generative AI

Abstract

UniMapGen is a generative framework for constructing global lane‑level vector maps with high efficiency and precision. It replaces traditional perception pipelines that rely on expensive survey vehicles or coarse satellite‑image‑based segmentation with a token‑based autoregressive generation process.

Introduction

High‑definition (HD) maps are essential for autonomous driving but their production is limited by the cost of specialized mapping fleets and the incompleteness of satellite imagery (occlusion, latency, annotation gaps). Existing pipelines treat map creation as a segmentation or detection problem, which yields fragmented road vectors and poor cross‑patch continuity. UniMapGen reframes map construction as a generative task: lane lines are encoded as discrete token sequences and produced by a large‑scale autoregressive model that can ingest multiple modalities.

Method

UniMapGen consists of three tightly coupled modules.

Multimodal input architecture The model accepts bird’s‑eye‑view (BEV) satellite images, low‑cost perspective‑view (PV) image sequences, optional natural‑language prompts, and contextual map data. Each modality can be omitted at training or inference, enabling operation with only satellite data when necessary.

Map vector tokenization Road polylines are uniformly sampled to a fixed spatial resolution (e.g., 20 cm per point) and then reordered into a canonical sequence. Each sampled point is quantized into a discrete token from a learned vocabulary, turning the vector prediction problem into a sequence‑to‑sequence generation task.

State‑update mechanism Instead of processing independent image patches, UniMapGen maintains a global state that records three semantic anchors for every vector: Start , End , and Cut . During inference the model iteratively expands a generated region (“growth” process), using the Cut anchors of previously generated patches as seeds for the next patch. This guarantees topological continuity and eliminates “broken road” artifacts across large areas.

Experiment

Evaluation was performed on the OpenSatMap dataset (20 cm resolution) covering extensive urban and suburban road networks. UniMapGen achieved state‑of‑the‑art completeness and continuity scores, outperforming detection‑based MapTR and segmentation‑based SegNeXt. Quantitative results show a significant increase in vector completeness and a reduction in fragmented edges, while qualitative visualizations demonstrate smooth lane curves and seamless stitching across patch boundaries.

Multimodal ablation confirms that adding PV sequences improves temporal freshness: real‑time PV frames correct outdated satellite BEV imagery, leading to more accurate current‑road representations.

Conclusion

UniMapGen demonstrates that large‑scale lane‑level map generation can be achieved by converting lane detection into discrete token generation and by enforcing a state‑update framework for global consistency. The flexible multimodal design (BEV, PV, text prompts) allows the system to adapt to varying data availability and to incorporate interactive updates. Future work will extend the token vocabulary to traffic signs, right‑of‑way rules, and other map elements, and will deepen text‑prompt interaction through synthetic data generation.

Paper homepage: https://amap-cvlab.github.io/UniMapGen/

Paper link: https://arxiv.org/abs/2509.22262