自动驾驶汽车场景测试与评估体系:研究现状、挑战及趋势OACSCD

Autonomous driving technology is accelerating toward large-scale testing and commercial application.Consequently,constructing systematic scenario frameworks for testing and robust evaluation metrics is crucial for safe deployment.This paper reviews the research status,challenges,and future trends of these systems.The study analyzes complexities introduced by vehicle-road-cloud integration and dynamic mixed traffic.It finds that tradition-al"mileage-failure"statistical models are insufficient for end-to-end performance assessment.Regarding test sce-narios,the paper outlines the evolution toward scenario-driven paradigms.It summarizes semantic description meth-ods based on the ISO 34501 standard and the PEGASUS six-layer model.Mainstream scenario generation technolo-gies are also reviewed.Current frameworks show insufficient coverage of long-tail and edge scenarios.Standards are highly fragmented.Furthermore,existing frameworks often under-represent vehicle-to-everything(V2X)collab-orative elements due to an excessive focus on single-vehicle intelligence.Regarding evaluation metrics,existing methodologies are categorized into three dimensions,including competition-based,closed-track/simulation hybrid,and theory-oriented approaches.The review identifies several deficiencies in current systems.Specifically,current metrics insufficiently assess the use of V2X collaborative information.Evaluation dimensions and workflows are fragmented,and objective quantitative metrics for interactive experience are lacking.To address these challenges,next-generation testing systems should focus on four research paths.①Unified scenario description languages and data-sharing frameworks are needed to establish benchmarks for measuring scenario risk criticality and realism.②Hierarchical scenario systems should be built to cover nominal conditions as well as long-tail boundaries for full-do-main coverage.③Comprehensive metrics should integrate communication latency,system resilience,and social eth-ics.④World models and generative AI,combined with causal inference,can simulate extreme conditions and ex-plore unknown failure modes to validate the system's generalization capability.