Persistent Autoregressive Mapping with Traffic Rules for Autonomous Driving

Shiyi Liang12*, Xinyuan Chang2* Changjie Wu2*, Huiyuan Yan1, Yifan Bai1,
Xinran Liu2, Hang Zhang2, Yujian Yuan2, Shuang Zeng12, Mu Xu2, Xing Wei1‡
1 Xi’an Jiaotong University     2 Amap, Alibaba Group
*Equal contribution    Work done during internship    Corresponding author
Schematic of Persistent Rule Effectiveness

PAMR ensures persistent rule effectiveness. Even after a traffic sign (e.g., a bus lane sign at time t) is no longer visible, the system retains the rule in its cache, preventing illegal maneuvers at time t+1.

Abstract

Safe autonomous driving requires both accurate HD map construction and persistent awareness of traffic rules, even when their associated signs are no longer visible. However, existing methods either focus solely on geometric elements or treat rules as temporary classifications, failing to capture their persistent effectiveness across extended driving sequences. In this paper, we present PAMR (Persistent Autoregressive Mapping with Traffic Rules), a novel framework that performs autoregressive co-construction of lane vectors and traffic rules from visual observations. Our approach introduces two key mechanisms: Map-Rule Co-Construction for processing driving scenes in temporal segments, and Map-Rule Cache for maintaining rule consistency across these segments. To properly evaluate continuous and consistent map generation, we develop MapDRv2, featuring improved lane geometry annotations. Extensive experiments demonstrate that PAMR achieves superior performance in joint vector-rule mapping tasks, while maintaining persistent rule effectiveness throughout extended driving sequences.

Method Overview

Overview of the PAMR framework

Overview of the PAMR framework. The model processes sequential PV frames and trajectory data. Each segment is tokenized and fed into an MLLM, along with a cache from the previous segment. The MLLM outputs detokenized lane vectors and their associated rules. This information is propagated across segments via the caching mechanism, enabling continuous and consistent map-rule generation.

MapDRv2: A New Benchmark

Comparison between MapDR and MapDRv2 annotations

To evaluate persistent map generation, we introduce MapDRv2. The original MapDR dataset (a) suffers from fragmented lane annotations in occluded scenarios. MapDRv2 (b) provides continuous and smooth lane geometries, offering a more suitable benchmark for evaluating models that generate consistent HD maps.

Qualitative Results

Visualization of sequential map-rule construction

Visualization of sequential map-rule construction. Segments 1-5 show the processing results within individual segments, while the "Global Map-Rule" shows the final integrated output. PAMR maintains rule awareness and generates smooth, accurate lane vectors even across lane changes and when signs are out of view.

Main Results

Main results table on MapDRv2 test set

PAMR significantly outperforms baselines in the comprehensive task of joint rule extraction and association (HMA F-score). This demonstrates the effectiveness of our co-construction approach for deep semantic understanding and end-to-end vector-rule mapping.

BibTeX

If you find our work useful in your research, please consider citing:

@article{liang2025persistent,
  title={Persistent Autoregressive Mapping with Traffic Rules for Autonomous Driving},
  author={Liang, Shiyi and Chang, Xinyuan and Wu, Changjie and Yan, Huiyuan and Bai, Yifan and Liu, Xinran and Zhang, Hang and Yuan, Yujian and Zeng, Shuang and Xu, Mu and others},
  journal={arXiv preprint arXiv:2509.22756},
  year={2025}
}