MambaMap: Online Vectorized HD Map Construction using State Space Model

Ruizi Yang, Xiaolu Liu, Junbo Chen, Jianke Zhu

Abstract

MambaMap is a novel framework for real-time, vectorized High-Definition (HD) map construction that leverages State Space Models (SSMs) to efficiently fuse temporal information. By replacing computationally expensive Transformers with Mamba-inspired linear-complexity modules, it achieves state-of-the-art accuracy in handling occlusions and long-range perception without the typical memory overhead associated with quadratic attention mechanisms. Evaluated on industry-standard benchmarks, MambaMap demonstrates superior performance in both accuracy and inference speed, making it a viable solution for onboard autonomous driving systems.

Introduction

High-definition (HD) maps serve as the fundamental "source of truth" for autonomous vehicles, providing the centimeter-level geometric and semantic details necessary for safe path planning. Traditionally, these maps are generated offline through labor-intensive SLAM pipelines and manual annotation. However, the industry is increasingly shifting toward online vectorized HD map construction, where the map is built dynamically from onboard camera feeds.

This real-time construction faces two primary challenges: occlusions (e.g., large vehicles blocking lane dividers) and long-range perception. To address these, models must maintain a "temporal memory" to integrate information across frames. Existing solutions often face a dilemma: recurrent neural networks (like GRUs) suffer from vanishing gradients and forget distant information, while Transformers exhibit quadratic computational complexity ($O(n^2)$), which limits their scalability for long sequences in real-time environments. MambaMap addresses this gap by introducing State Space Models to the HD mapping domain.

Methodology

MambaMap introduces a specialized architecture that utilizes State Space Models (SSMs)—specifically a gated version inspired by the Mamba architecture—to achieve linear complexity while maintaining long-range dependency modeling. The framework is built upon several key innovations:

1. Memory Bank and Gated State Space Block

The system maintains a sliding-window buffer that stores historical Bird’s Eye View (BEV) features and instance queries. This ensures the model has access to relevant past data without being overwhelmed by redundancy. A Gated State Space Block acts as a selective mechanism, allowing the model to prioritize critical historical information while discarding sensor noise and artifacts.

2. BEV Mamba Fusion (BMF)

Processing 2D BEV grids using inherently 1D SSMs requires a specialized approach. MambaMap employs a Multi-Directional Scanning strategy, flattening BEV features in four directions (left, right, up, down). This ensures the model captures spatial relationships and geometric structures from all orientations, effectively translating 2D road topology into a format the SSM can process.

3. Instance Mamba Fusion (IMF)

To ensure temporal consistency of map elements (such as lane boundaries or pedestrian crossings), the IMF module refines predictions over time. It utilizes the Hungarian algorithm for instance matching across frames and a Spatial-Temporal Scanning strategy. This allows the model to learn both the intra-frame appearance of an object and its evolution across the temporal sequence, leading to more stable and coherent map outputs.

Experimental Evaluation

The authors evaluated MambaMap on two large-scale, industry-standard datasets:

  • nuScenes: Comprising 1,000 scenes from Boston and Singapore, utilizing a six-camera surround-view setup.
  • Argoverse 2: A dataset featuring seven-camera setups across six distinct cities, used to test the model's generalization capabilities across diverse urban environments.

Results and Discussion

MambaMap establishes a new state-of-the-art (SOTA) across multiple benchmarks:

  • Accuracy: On the nuScenes dataset (60x30m range), MambaMap achieved 67.3 mAP (mean Average Precision), significantly outperforming previous leading models such as StreamMapNet (63.4 mAP) and SQD-MapNet (65.8 mAP).
  • Long-Range Robustness: In extended perception ranges (100x50m), the model maintained a high 66.5 mAP, demonstrating its ability to handle distant features where other models typically degrade.
  • Generalization: On the geographically disjoint split of Argoverse 2, MambaMap outperformed competitors by at least 3.1 mAP, suggesting superior adaptation to unseen urban layouts.
  • Efficiency: Despite the sophisticated dual-fusion architecture, the model maintains an inference speed of 12.7 to 13.6 FPS on a single NVIDIA A30 GPU, meeting the requirements for real-time deployment.

Conclusion

MambaMap demonstrates that linear-complexity State Space Models are a highly effective alternative to Transformers for temporal fusion in autonomous driving tasks. By successfully applying multi-directional scanning to BEV features and ensuring instance consistency through spatial-temporal scanning, the framework provides a robust and efficient solution for online HD map construction. Future work may involve extending this Mamba-based architecture to other BEV-centric tasks, such as 3D object detection and motion forecasting, to create a unified and efficient world model for autonomous navigation.

Summary updated

BibTeX
@inproceedings{yang2025mambamap,
  author = {Ruizi Yang and Xiaolu Liu and Junbo Chen and Jianke Zhu},
  title = {{MambaMap: Online Vectorized HD Map Construction using State Space Model}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2025},
  doi = {10.1109/iros60139.2025.11247426},
}
Figure