自动驾驶汽车的高效对抗性场景测试方法研究
作者:
作者单位:

作者简介:

通讯作者:

基金项目:

深圳市基础研究重点项目(JCYJ20200109115414354, JCYJ20200109115403807);广东省基金项目(2020B515130004,2023A1515011813)

伦理声明:



Efficient Adversarial Scenario Test for Autonomous Vehicles
Author:
Ethical statement:

Affiliation:

Funding:

This work is supported by Shenzhen Basic Key Research Project (JCYJ20200109115414354,JCYJ20200109115403807) and Foundation of Guangdong Province of China (2020B515130004,2023A1515011813)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
    摘要:

    在自动驾驶安全性的研究和应用中,测试里程长、暴露危险场景单一的问题使自动驾驶安全性能的提升受到限制。使用对抗性场景进行测试被认为是解决上述问题的重要手段,然而,现有研究采用通用的优化算法作为框架,将大量计算资源浪费在对参数空间的探索过程中,效率低下。在计算成本的约束下,这些算法甚至无法在更复杂的环境中测试出足够多、足够丰富的失效样本。复杂环境中的对抗性场景测试面临三大挑战:信息匮乏;对抗性样本在庞大的参数空间中稀疏分布;搜索过程中探索与利用难以平衡。该文从这三大挑战出发,提出一种高效的对抗性场景测试框架,通过代理模型来获取更多关于参数空间的信息,精选小样本,以打破庞大空间中稀疏事件的制约,对未知区域和对抗性样本附近的目标进行有针对性的搜索和更新,以实现探索和利用的平衡。实验证明,该文提出方法的搜索效率是随机采样的 4 倍,与通用遗传算法相比,效率提升一倍以上,在有限的仿真测试次数下,生成了更多容易使被测自动驾驶系统失效的对抗性测试用例。特别地,该文提出的方法能够找出许多离群的对抗性样本,揭示出现有算法无法识别的失效模式。此外,该文提出的方法能够快速、全面地定位出被测算法的脆弱场景,为自动驾驶算法的测试验证、迭代升级提供支持。

    Abstract:

    In the field of autonomous driving safety research and application, the limitations of limited testing mileage and exposure to only a single hazardous scenario hinder the improvement of autonomous driving safety performance. To address these issues, testing with adversarial scenarios is considered crucial. However, existing studies utilize generic optimization algorithms as frameworks, resulting in a wastage of computational resources in exploring the parameter space, thereby leading to low efficiency. Moreover, under the constraint of computational cost, these algorithms may not be able to test a sufficient number of diverse failure samples, especially in complex environments. Adversarial scenario testing in complex environments faces three major challenges: information scarcity, sparse distribution of adversarial samples in a vast parameter space, and the difficulty in balancing exploration and exploitation during the search process. To tackle these challenges, this paper proposes an efficient framework for adversarial scenario testing. This framework employs a surrogate model to gather more information about the parameter space, selects small samples to overcome the sparse event constraints in the vast space, and focuses on the unknown regions and adversarial samples for targeted search and update, thereby achieving a balance between exploration and exploitation. Experimental results demonstrate that the proposed method in this paper exhibits a search efficiency four times higher than random sampling and more than double the efficiency compared to general genetic algorithms. Additionally, with a limited number of simulation test runs, it generates a greater number of adversarial test cases that are likely to cause the tested autonomous driving system to fail. Notably, the proposed method can identify many outlier adversarial samples, unveiling failure modes that existing algorithms fail to recognize. Furthermore, the proposed method can swiftly and comprehensively identify the vulnerable scenarios of the tested algorithm, providing support for the testing, validation, and iterative upgrade of autonomous driving algorithms.

    参考文献
    相似文献
    引证文献
引用本文

引文格式
桑 明,蒋拯民,李慧云.自动驾驶汽车的高效对抗性场景测试方法研究 [J].集成技术,2024,13(2):15-28

Citing format
SANG Ming, JIANG Zhengmin, LI Huiyun. Efficient Adversarial Scenario Test for Autonomous Vehicles[J]. Journal of Integration Technology,2024,13(2):15-28

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
历史
  • 收稿日期:2023-07-26
  • 最后修改日期:2023-07-26
  • 录用日期:2023-11-20
  • 在线发布日期: 2023-11-20
  • 出版日期: