Scholarlink

<p>Autonomous Driving Systems (ADS) necessitate meticulous testing procedures to ensure their safety before deployment. While current fuzz testing methods primarily focus on varying dynamic object behaviors, they often overlook the critical influence of sensing capability on ADS planning algorithms. Addressing this gap, this paper presents BlindSpotFuzz, a comprehensive approach. BlindSpotFuzz introduces a scenario safety model predicated on detecting blind spots, offering a novel perspective. Leveraging this safety model, we introduce four innovative testing metrics: (1) severity of occlusion, (2) evolutionary coverage, (3) scenario change detection and (4) coverage of blind spots. These metrics serve as feedback to guide fuzz testing in systematically uncovering deficiencies in planning algorithms caused by blind spots. Furthermore, we developed a seed generation module that leverages a large language model (LLM) to extract key road structure features from crash reports and construct the road network. The generated map, combined with crash report insights, guides the LLM to form the testing seeds. The BlindSpotFuzz algorithm was evaluated against Baidu Apollo. Our experimental results demonstrate the capability of our approach to generate unique test cases, explicitly capturing scenarios relevant to sensing blind spots. Furthermore, the proposed metrics exhibit efficiency in covering blind spot scenarios and uncovering potentially dangerous test cases compared to existing baseline testing approaches.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

BlindSpotFuzz: testing autonomous driving systems through blind-spot-guided fuzzing

  • 期刊论文

  • 发表日期：

    2026年2月7日

查看全文
  • Sali Moussa,
  • Junyan Ma,
  • Muhammad Umer Abbasi

摘要

Autonomous Driving Systems (ADS) necessitate meticulous testing procedures to ensure their safety before deployment. While current fuzz testing methods primarily focus on varying dynamic object behaviors, they often overlook the critical influence of sensing capability on ADS planning algorithms. Addressing this gap, this paper presents BlindSpotFuzz, a comprehensive approach. BlindSpotFuzz introduces a scenario safety model predicated on detecting blind spots, offering a novel perspective. Leveraging this safety model, we introduce four innovative testing metrics: (1) severity of occlusion, (2) evolutionary coverage, (3) scenario change detection and (4) coverage of blind spots. These metrics serve as feedback to guide fuzz testing in systematically uncovering deficiencies in planning algorithms caused by blind spots. Furthermore, we developed a seed generation module that leverages a large language model (LLM) to extract key road structure features from crash reports and construct the road network. The generated map, combined with crash report insights, guides the LLM to form the testing seeds. The BlindSpotFuzz algorithm was evaluated against Baidu Apollo. Our experimental results demonstrate the capability of our approach to generate unique test cases, explicitly capturing scenarios relevant to sensing blind spots. Furthermore, the proposed metrics exhibit efficiency in covering blind spot scenarios and uncovering potentially dangerous test cases compared to existing baseline testing approaches.