Safe2Drive: Evaluating Safe Driving Behaviors of E2E Autonomous Driving Models
Researchers have introduced Safe2Drive (S2D), a new benchmark designed to evaluate the safety of end-to-end autonomous driving models. S2D includes 100 challenging scenarios, such as work zones and pedestrian jaywalking, and introduces a SafeDriving Score (SDS) to measure safety-critical behaviors. When tested on S2D, two leading models, LEAD and SimLingo, showed significantly reduced driving scores compared to their performance on existing benchmarks, indicating brittle safe-driving capabilities and a lack of robust behavioral reasoning. AI
IMPACT Highlights critical safety gaps in current end-to-end autonomous driving models, necessitating further research into robust behavioral reasoning.