一句话：提出 ABLE：结合 GFlowNet、动态 STL 鲁棒性目标与主动学习，高效生成多样化的违法驾驶场景以测试自动驾驶系统。
自动驾驶系统（ADS）部署前需满足复杂交通法规。为全面测试 ADS，应系统性地发现违反不同法规的多样场景。挑战在于：可测法规多（中国交通法规 13 条可测条款、81 种违法情形；新加坡 16 条、43 种情形），且许多法规仅在复杂场景中相关；现有方法依赖无碰撞等简单判定，生成违法场景能力有限。本文提出 ABLE，受 GFlowNet 启发，通过生成多样场景来高效违反多条法规。ABLE 以基于信号时序逻辑（STL）鲁棒性语义的动态测试目标和主动学习驱动测试，以探索巨大搜索空间。基于 Apollo 与 LGSVL 的评估显示，ABLE 在测试 Apollo 6.0 和 7.0 时分别比最先进方法多违反 17% 和 25% 的法规，且多为难以违反者。
An automated driving system (ADS), as the brain of an autonomous vehicle (AV), should be tested thoroughly ahead of deployment. ADS must satisfy a complex set of rules to ensure road safety, e.g., the existing traffic laws and possibly future laws that are dedicated to AVs. To comprehensively test an ADS, we would like to systematically discover diverse scenarios in which certain traffic law is violated. The challenge is that (1) there are many traffic laws (e.g., 13 testable articles in Chinese traffic laws and 16 testable articles in Singapore traffic laws, with 81 and 43 violation situations respectively); and (2) many of traffic laws are only relevant in complicated specific scenarios. Existing approaches to testing ADS either focus on simple oracles such as no-collision or have limited capacity in generating diverse law-violating scenarios. In this work, we propose ABLE, a new ADS testing method inspired by the success of GFlowNet, which Aims to Break many Laws Efficiently by generating diverse scenarios. Different from vanilla GFlowNet, ABLE drives the testing process with dynamically updated testing objectives (based on a robustness semantics of signal temporal logic) as well as active learning, so as to effectively explore the vast search space. We evaluate ABLE based on Apollo and LGSVL, and the results show that ABLE outperforms the state-of-the-art by violating 17% and 25% more laws when testing Apollo 6.0 and Apollo 7.0, most of which are hard-to-violate laws, respectively.
@inproceedings{ZhangZS0SD023,
title = {Testing Automated Driving Systems by Breaking Many Laws Efficiently},
author = {Xiaodong Zhang and Wei Zhao and Yang Sun and Jun Sun and Yulong Shen and Xuewen Dong and Zijiang Yang},
booktitle = {ISSTA},
year = {2023},
doi = {10.1145/3597926.3598108}
}