DriveE2E: An Infrastructure-Grounded Ego-Closed-Loop Replay Benchmark for End-to-End Autonomous Driving

TMLR Paper9286 Authors

Abstract: Closed-loop evaluation is important for end-to-end autonomous driving, but existing CARLA-based benchmarks often rely on manually designed scenarios whose traffic patterns may differ from real-world urban driving. We present DriveE2E, an infrastructure-grounded ego-closed-loop replay benchmark for evaluating end-to-end autonomous driving models in reconstructed real-world intersection scenarios. DriveE2E uses high-mounted infrastructure sensors to extract traffic trajectories from 100 hours of urban intersection data, constructs CARLA-compatible digital twins for 15 real intersections, and imports 800 curated traffic scenarios into simulation. In DriveE2E, the tested model controls the ego vehicle and receives simulation-generated observations from its current simulated state, while non-ego agents replay trajectories extracted from real-world traffic. This protocol does not model fully reactive multi-agent behavior; instead, it provides a reproducible intermediate regime between open-loop log replay and fully reactive simulation. We instantiate the benchmark with representative E2EAD baselines and analyze their open-loop and ego-closed-loop replay performance across behavior categories. The results suggest that DriveE2E can expose differences between open-loop trajectory accuracy and policy behavior under ego feedback in dense intersection scenarios. Code is included in the supplementary material, and will be publicly released upon acceptance.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Amrit_Bedi1

Submission Number: 9286