高等级自动驾驶汽车虚拟测试：研究进展与前沿

安全性测试是高等级自动驾驶汽车（Highly Automated Vehicles，HAV，指具备L3级及以上能力的自动驾驶汽车）规模化应用的基本保障。鉴于HAV测试对象与测试标准的变革，传统基于里程的车辆测试方法论不再适用，场景化虚拟测试正成为验证HAV安全性的核心方法。基于与国内外多家HAV研发机构开展虚拟测试合作的基础上，针对测试场景、测试工具和测试方法等方面的技术难点和学术问题进行汇总、归纳和分析。测试场景方面，围绕场景覆盖度的要求，需重点关注测试场景自主划分、自动化仿真生成和未知高风险场景搜寻等理论方法。测试工具方面，在构建HAV自动驾驶系统“环境感知-规划决策-运动执行”一体化仿真工具的基础上，需研究支撑测试场景生成、驾驶行为双向交互和多传感器物理模型融合的高可信仿真技术。测试方法方面，针对海量测试场景、HAV驾驶能力非单调变化等特征，亟待开展覆盖度驱动型测试方法、加速测试方法和多目标测试与评估等的研究。此外，在上述研究挑战的基础上，面向HAV虚拟测试自动化、快速化、一体化和协同化的应用需求，提出HAV虚拟测试仿真即服务（Simulation as a Service，SAAS）的系统架构，并进一步明确HAV安全性诊断分析、系统自主优化训练和面向系统快速迭代升级的测试方法等SAAS重点研究需求。

Highly Automated Vehicle Virtual Testing: A Review of Recent Developments and Research Frontiers

Safety testing and validation are required for large-scale utilization of highly automated vehicles (HAVs). HAVs contain driving automation above level 3. However, owing to the substantial variations in the testing objects and safety requirements, traditional distance-based vehicle testing methods are unsuitable for HAVs. Scenario-based testing, using simulation, is the currently acknowledged approach. Here, based on the theoretical and application-oriented investigations from the HAV research and development institutions, which are focused on virtual tests, the authors comprehensively reviewed the unresolved technical concerns and research questions. The principal research areas were the testing scenarios, simulation tools, and testing methods. For efficient scenario coverage in the testing scenarios, methods for automatic functional scenario partition, simulation-based scenario generation, and unknown unsafe scenario identification are required. Further,for the virtual test tools, establishing integrated tools that have simulation capability are required for environmental sensing, motion planning, and vehicle kinematic movement. Therefore, future research needs to develop: a) simulation models that support scenario generation and bilateral interaction of driving behavior, and b) physical simulation models with fused-in multi environmental sensors. Moreover, for efficient bulk-testing scenarios and non-monotonic capable driving,developing coverage-driven and accelerated testing methods, and establishing multi-objectives evaluation schemes are required. Further, to address these requirements, this study proposes a simulation as a service (SAAS) HAV testing system, which provides automatic, fast, integrated, and coordinated HAV simulation test service. Under the HAV SAAS frame-work, this study discusses the principal research areas such as HAV safety diagnostic analysis technique, automatic evaluation of HAV algorithm optimization,and testing methods for HAV algorithm updates.