一句话：提出基于真实轨迹挖掘的影响行为模式，自动生成自动驾驶系统关键测试场景的 CRISCO 技术。
自动驾驶系统安全攸关，部署前需充分测试，但现有方法或在高维输入空间中搜索，或依赖简单预定义场景，效率与充分性不足。本文从真实交通轨迹中挖掘具有影响的行为模式，并据此提出场景生成技术 CRISCO。CRISCO 让交通参与者执行影响行为来挑战 ADS，通过求解轨迹约束生成不同测试场景，并逐步加入影响行为模式中的参与者行为，提升非关键场景的挑战性。作者在工业级平台百度 Apollo 上评估，12 小时内暴露 13 类不同安全违规，证明其能高效生成致撞关键场景，并在相同道路上比两种先进 ADS 测试技术多暴露 5 类安全违规。
Autonomous Driving Systems (ADSs) are safety-critical, and must be fully tested before being deployed on real-world roads. To comprehensively evaluate the performance of ADSs, it is essential to generate various safety-critical scenarios. Most of existing studies assess ADSs either by searching high-dimensional input space, or using simple and pre-defined test scenarios, which are not efficient or not adequate. To better test ADSs, this paper proposes to automatically generate safety-critical test scenarios for ADSs by influential behavior patterns, which are mined from real traffic trajectories. Based on influential behavior patterns, a novel scenario generation technique, CRISCO, is presented to generate safety-critical scenarios for ADSs testing. CRISCO assigns participants to perform influential behaviors to challenge the ADS. It generates different test scenarios by solving trajectory constraints, and improves the challenge of those non-critical scenarios by adding participants’ behavior from influential behavior patterns incrementally. We demonstrate CRISCO on an industrial-grade ADS platform, Baidu Apollo. The experiment results show that our approach can effectively and efficiently generate critical scenarios to crash ADS, and it exposes 13 distinct types of safety violations in 12 hours. It also outperforms two state-of-art ADS testing techniques by exposing more 5 distinct types of safety violations on the same roads.
@inproceedings{paperbot7330,
title = {Generating Critical Test Scenarios for Autonomous Driving Systems via Influential Behavior Patterns},
author = {Haoxiang Tian and Guoquan Wu and Jiren Yan and Yan Jiang and Jun Wei and Wei Chen and Shuo Li and Dan Ye},
booktitle = {ASE},
year = {2022},
doi = {10.1145/3551349.3560430}
}