Scholarlink

<p>High-quality benchmarks are essential for revealing potential faults in Autonomous Driving Systems (ADS). However, most existing datasets adopt coarse-grained scenario definitions and lack structured categorization, limiting their applicability for targeted behavioral analysis under varying road structures and environmental conditions. To address this, we introduce <span>Muses</span>, a fine-grained benchmark that organizes 1,218 driving scenarios across 12 road structure types, 9 roadside element categories, and default weather conditions. <span>Muses</span> incorporates an automated monitoring tool that detects eight categories of anomalous driving behaviors and records frame level annotations, replay logs, historical trajectory data, actor details, and video evidence for reproducible evaluation. We evaluate <span>Muses</span> across systems, including the industrial Apollo ADS and CARLA built-in autonomous agent. Our empirical results show that anomaly frequency increases consistently with structural complexity, and that <span>Muses</span> exposes consistent behavioral anomaly categories across ADSs, demonstrating strong generalizability. By improving scenario granularity, environmental diversity, and reproducibility infrastructure, <span>Muses</span> enables more precise and systematic assessment of ADS robustness and safety vulnerabilities. Data access: <a href="https://github.com/hjlhhh-eng/Muses">https://github.com/hjlhhh-eng/Muses</a>.</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

Multi-scenario benchmark for autonomous driving systems: Exposing diverse behavioral anomalies

  • 期刊论文

  • 发表日期：

    2026年5月30日

查看全文
  • Jialing Huang,
  • Tingting Wu,
  • Zuohua Ding,
  • Yongkui Xu,
  • Yunwei Dong

摘要

High-quality benchmarks are essential for revealing potential faults in Autonomous Driving Systems (ADS). However, most existing datasets adopt coarse-grained scenario definitions and lack structured categorization, limiting their applicability for targeted behavioral analysis under varying road structures and environmental conditions. To address this, we introduce Muses, a fine-grained benchmark that organizes 1,218 driving scenarios across 12 road structure types, 9 roadside element categories, and default weather conditions. Muses incorporates an automated monitoring tool that detects eight categories of anomalous driving behaviors and records frame level annotations, replay logs, historical trajectory data, actor details, and video evidence for reproducible evaluation. We evaluate Muses across systems, including the industrial Apollo ADS and CARLA built-in autonomous agent. Our empirical results show that anomaly frequency increases consistently with structural complexity, and that Muses exposes consistent behavioral anomaly categories across ADSs, demonstrating strong generalizability. By improving scenario granularity, environmental diversity, and reproducibility infrastructure, Muses enables more precise and systematic assessment of ADS robustness and safety vulnerabilities. Data access: https://github.com/hjlhhh-eng/Muses.