Reproduce “UncAD: Towards Safe End-to-end Autonomous Driving via Online Map Uncertainty” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Pengxuan Yang, Yupeng Zheng, Qichao Zhang, Kefei Zhu, Zebin Xing, Qiao Lin, Yun-Fu Liu, Zhiguo Su and 1 more
UncAD integrates map-point uncertainty into predictions and candidate ego paths, using it to select a route. It estimates map-point uncertainty with a Laplace-based model and filters high-risk candidates using uncertainty and collision checks. On nuScenes, VAD’s average displacement error falls from 0.72 m to 0.62 m, while collision and drivable-area conflict rates decrease. These benchmark results do not establish real-world safety.
Key insights
- Incorporating uncertainty information into both prediction and planning improves safety outcomes
- Filtering trajectories based on both driving area uncertainty and collision risk leads to safer decisions
- The framework can be effectively integrated into existing state-of-the-art autonomous driving systems
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.