Reproduce “Data-Driven Risk Fields for Safer End-to-End Autonomous Driving” 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
Yuanxin Tian, Zhiyuan Liu, Jinhao Li, Liangfan Zhu, Shuai Wang, Heye Huang, Qingwen Meng, Fang Zhang and 4 more
DRiF introduces an ego-conditioned BEV risk representation for end-to-end autonomous driving, separating static road constraints from dynamic interaction hazards. Rather than regressing arbitrary absolute risk scores, it learns pairwise risk preferences derived from overlap, future-corridor conflict, and occupancy rules; on the Bench2Drive simulator benchmark, the four-frame model reaches an 88.78 Driving Score and reduces collision-route rate to 15.91%, compared with 85.65 and 22.73% for the TF++ baseline.
Key insights
- Relative ordering is a more effective supervision target than absolute handcrafted risk: replacing ranking with scalar regression lowers Driving Score from 88.78 to 85.41 and multi-ability performance from 72.57 to 65.38.
- Static road structure and dynamic interaction risk are complementary; removing static map supervision reduces the score to 83.23, whereas removing the dynamic risk branch causes a much larger drop to 62.40.
- Combining occupancy, overlap, and future-corridor risk outperforms using only one or two of these signals, indicating that different hazard types contribute distinct planning information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.