Persistent Autoregressive Mapping with Traffic Rules for Autonomous Driving

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.

Figures (5)

• Figure 1: Schematic of Persistent Rule Effectiveness. (a) Without rule persistence: At time t, a bus lane sign is observed. At t+1, without remembering the rule, the vehicle attempts an illegal lane change. (b) With rule persistence: The system retains the bus lane rule, preventing the incorrect maneuver and ensuring continued adherence to traffic rules.
• Figure 2: Comparison between traditional pipeline and PAMR. (a) Traditional methods adopt a multi-stage pipeline: constructing lane vectors, extracting and associating traffic rules. This separated approach fails to maintain vector consecutiveness, rule persistence, and end-to-end learning. (b) Our PAMR framework performs autoregressive co-construction of lanes and rules, enabling sequential reasoning. Through joint construction, PAMR achieves continuous vectors, persistent rule awareness, and end-to-end integration of geometric and semantic information.
• Figure 3: Comparison with MapDR in occluded scenarios. For each example, the left three images show the input PV views, while the right images present the corresponding ground-truth HD maps. Green lines and blue lines represent the dividers and borderlines, respectively. In the case of occlusion and glare, MapDR generates fragmented vector representations. In contrast, MapDRv2 provides more complete results.
• Figure 4: Overview of the PAMR framework. Left: The sequential processing of map-rules, where each segment takes PV frames and trajectory as input, tokenizes them along with cache from previous segment (if any), and feeds them into MLLM. The MLLM outputs are then detokenized into lane vectors with associated rules. Right: Bird's-eye view visualization of the map-rule construction process, showing how information is propagated across consecutive segments through the caching mechanism. The cache ensures continuous rule awareness even as the vehicle moves forward, enabling consistent map-rule generation across the entire trajectory.
• Figure 5: Visualization of map-rule construction. Segment 1-5 demonstrate the sequential processing results within individual segments, while the HD map shows the final integrated output after segments are concatenation. The green lines represent the constructed lane vectors while the black lines indicate border lines, and the red arrow shows the vehicle trajectory within each segment.