Exploring Real World Map Change Generalization of Prior-Informed HD Map Prediction Models

TL;DR

This work interrogates whether prior-informed online HD map prediction models, trained with synthetically perturbed map priors, can generalize to real-world map changes. It presents a large-scale experimental study comparing diverse perturbation types and evaluating on multi-year autonomous driving data to quantify transferability. The findings reveal a substantial sim2real gap: synthetic perturbations largely capture only simple real-world changes, and excessive prior noise can degrade real-world performance, though certain discrete mutations like feature duplication can be beneficial. Overall, the results motivate developing more robust generalization strategies for HD map prediction and addressing the sim2real discrepancy to enable safer, lifelong deployment.

Abstract

Building and maintaining High-Definition (HD) maps represents a large barrier to autonomous vehicle deployment. This, along with advances in modern online map detection models, has sparked renewed interest in the online mapping problem. However, effectively predicting online maps at a high enough quality to enable safe, driverless deployments remains a significant challenge. Recent work on these models proposes training robust online mapping systems using low quality map priors with synthetic perturbations in an attempt to simulate out-of-date HD map priors. In this paper, we investigate how models trained on these synthetically perturbed map priors generalize to performance on deployment-scale, real world map changes. We present a large-scale experimental study to determine which synthetic perturbations are most useful in generalizing to real world HD map changes, evaluated using multiple years of real-world autonomous driving data. We show there is still a substantial sim2real gap between synthetic prior perturbations and observed real-world changes, which limits the utility of current prior-informed HD map prediction models.

Figures (4)

• Figure 1: A real-world map change from an autonomous vehicle dataset. In this paper we investigate which synthetic perturbations applied to a simulated prior map at training time best model a real prior map (left) for training a prior-informed online mapping model to produce the updated map (right), evaluated using a vast collection of real-world changes gathered over multiple years of autonomous vehicle operation.
• Figure 2: Overall architecture of our model, heavily influenced by liao_maptr_2023sun_mind_2023. See these references for more details on implementation.
• Figure 3: Types of perturbations. Note that \ref{['fig:duplication']} looks the same as \ref{['fig:no_perturb']} since features are exactly duplicated.
• Figure 4: Qualitative results handling real world changes. For minor real-world changes, e.g. driveway geometry (a--d) and curb geometry (e--h), a model trained with prior perturbations correctly predicts changes to many of the real-world features. However, for substantial changes in road layout, e.g. additional medians (i--l) and new road construction (m--p), the model fails to meaningfully deviate from the prior to account for the new intersection geometry. Note that the topdown map is centered on the vehicle in all figures.