一句话：提出 record-modify-replay 框架，解耦自车与记录场景，实现确定性 ADS 评估、自车注入与反事实回放。
高保真仿真器（如 CARLA）在测试自动驾驶系统（ADS）时存在非确定性，导致仿真结果不稳定（flaky），既难以公平比较不同 ADS agent，也阻碍定向测试与调试。本文提出 DT-Drive，一个 record-modify-replay 框架：先录制驾驶场景，再将 ego vehicle 与记录的场景解耦，由此支持“ego-injection”（在同一场景中注入不同自车策略以公平对比多个 ADS）与“counterfactual replay”（通过修改天气、交通等环境条件做反事实重放），从而支撑定向测试与调试。作者在 128 个 CARLA 基准场景上验证，DT-Drive 取得 100% 确定性的评估结果，有效消除了仿真器引入的不稳定性；并使用该工具在相同及修改环境下对 TransFuser 与 TransFuser++ 进行受控比较，展示了其测试与调试价值。工具及样例数据已开源，并提供演示视频。
While simulation is essential for testing autonomous driving systems (ADS), high-fidelity simulators like CARLA often exhibit non-determinism, leading to flaky simulation results. We present DT-Drive, a record-modify-replay framework that ensures deterministic ADS evaluation by decoupling the ego vehicle from the recorded driving scenario. This approach enables “ego-injection” for fair multi-ADS comparisons and “counterfactual replay” via environmental modifications (e.g., weather and traffic), facilitating targeted testing and debugging. Validated on 128 CARLA benchmark scenarios, DT-Drive achieved 100% deterministic evaluation results, effectively eliminating simulator-induced flakiness. We also demonstrate its utility for testing and debugging by conducting a controlled comparison of the TransFuser and TransFuser++ agents under identical and modified environmental conditions. DT-Drive and the sample data are available at https://github.com/SanjeethaPennada/DT-Drive. The tool demo video is available at https://youtu.be/HB1DIZSF5E0.
@inproceedings{paperbot12246,
title = {DT-Drive: A Tool for Deterministic Replay-Based Testing and Debugging of Autonomous Driving Systems},
author = {Sanjeetha Pennada and Matthew Leach and Olek Osikowicz and Donghwan Shin},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3834601}
}