UncAD: Towards Safe End-to-end Autonomous Driving via Online Map Uncertainty

TL;DR

This work addresses safety in end-to-end autonomous driving by explicitly modeling online map uncertainty. It introduces Map Uncertainty Estimation (MUE), Uncertainty-Guided Prediction and Planning (UGPnP), and Uncertainty-Collision-Aware Selection (UCAS) to produce and select multi-modal trajectories that avoid high-uncertainty regions and potential collisions. The approach, integrated with state-of-the-art end-to-end methods, yields significant reductions in collision rate and drivable-area conflicts on nuScenes and introduces the DACR metric to better gauge planning safety. Overall, UncAD demonstrates that incorporating map uncertainty into perception, prediction, and planning leads to safer and more robust autonomous driving in complex scenarios.

Abstract

End-to-end autonomous driving aims to produce planning trajectories from raw sensors directly. Currently, most approaches integrate perception, prediction, and planning modules into a fully differentiable network, promising great scalability. However, these methods typically rely on deterministic modeling of online maps in the perception module for guiding or constraining vehicle planning, which may incorporate erroneous perception information and further compromise planning safety. To address this issue, we delve into the importance of online map uncertainty for enhancing autonomous driving safety and propose a novel paradigm named UncAD. Specifically, UncAD first estimates the uncertainty of the online map in the perception module. It then leverages the uncertainty to guide motion prediction and planning modules to produce multi-modal trajectories. Finally, to achieve safer autonomous driving, UncAD proposes an uncertainty-collision-aware planning selection strategy according to the online map uncertainty to evaluate and select the best trajectory. In this study, we incorporate UncAD into various state-of-the-art (SOTA) end-to-end methods. Experiments on the nuScenes dataset show that integrating UncAD, with only a 1.9% increase in parameters, can reduce collision rates by up to 26% and drivable area conflict rate by up to 42%. Codes, pre-trained models, and demo videos can be accessed at https://github.com/pengxuanyang/UncAD.

Figures (5)

• Figure 1: The panoptic images are shown on the left, and the bird's eye view (BEV) map with the ego vehicle’s planning trajectory is shown on the right. In the BEV map, green, pink, and blue ellipses represent the uncertainty of boundaries, pedestrians, and lane dividers, respectively. In this challenging (rainy and sharp turning) scenario, VAD suffers from the inaccuracy online map, leading to collide with map boundaries (the blue trajectory in both sides). Our method leverages map uncertainty to guide the ego vehicle to avoid high-uncertainty areas and drive within confirmed drivable area, resulting in safer trajectory (the red trajectory in both sides).
• Figure 2: The overall architecture of UncAD consists of four core modules. First, the Feature Extractor uses an encoder to project image inputs into BEV or sparse features. The Map Uncertainty Estimation (MUE) module encodes scene information into map queries and estimates map uncertainty. In the Uncertainty-Guided Prediction and Planning (UGPnP) module, agent and ego queries incorporate map uncertainty through query interaction, generating prediction and planning trajectories. Finally, the Uncertainty-Collision-Aware Selection (UCAS) strategy selects the optimal trajectory based on driving area uncertainty and collision risk, ensuring the safest path.
• Figure 3: The DACR result of our method and VAD on nuScenes dataset in turn and straight driving scenarios
• Figure 4: The efficiency of UCAS strategy in avoiding collision.
• Figure 5: The efficiency of our UncAD method in complex scenarios, such as sharp turn and low texturless night scenarios.