一句话：提出 CCTest 框架，将复杂场景分解为简单场景以生成关键测试数据，高效暴露自动驾驶系统的真实缺陷。
基于场景的仿真测试是验证自动驾驶系统（ADS）的主流手段，但生成既能有效暴露真实缺陷、又能控制生成复杂度的关键场景仍很困难。本文提出 CCTest 测试框架：它把复杂场景分解为简单场景，并据此生成关键测试数据，同时对场景给出定性分析。为保证结果可归因，CCTest 构造的情境中每辆车都由被测 autopilot 控制且各自拥有可行的安全策略，因此任何观察到的安全违规都只能由被测 autopilot 的缺陷造成，而非场景本身的不可行性。评估表明，CCTest 在 Apollo、Autoware 以及 Carla 与 LGSVL 模拟器内置 autopilot 上成功揭示出重大缺陷，并且在高效发现真实缺陷方面优于现有工具。
Scenario-based testing is the dominant approach for validating autonomous driving systems (ADS) through simulation. However, generating critical scenarios that effectively uncover real defects in ADS while managing the complexity of the generation remains a challenge. This paper introduces CCTest, a test framework that generates critical test data using simple scenarios decomposed from complex ones and provides qualitative analysis of the scenarios. CCTest ensures that any safety violation observed indicates a defect in the tested autopilot by creating situations where each vehicle is controlled by the autopilot under test and has a feasible safety policy. Evaluations have shown that CCTest effectively revealed major defects in Apollo, Autoware, and the built-in autopilots in Carla and LGSVL simulators. CCTest also outperforms existing tools in efficiently uncovering real defects in autopilots.