计算机工程与应用2025，Vol.61Issue(19)：12-42,31.DOI:10.3778/j.issn.1002-8331.2502-0072
无人驾驶深度强化学习决策模型性能评测方法综述
Review of Performance Evaluation Methods for Deep Reinforcement Learning Decision Models in Autonomous Driving
摘要
Abstract
Recently,end-to-end autonomous driving technology,using deep reinforcement learning(DRL)as the primary decision-making method,has achieved significant progress in complex dynamic environments.However,due to the unique trial-and-error learning mechanism of DRL,it must undergo a strict multi-dimensional evaluation process before being applied to real-world driving environments.Thus,performance evaluation becomes a critical and indispensable step for transferring DRL-based autonomous driving decision models to the real world.Firstly,this paper reviews the mainstream technical implementation methods in the field of autonomous driving.Next,the focus is placed on DRL methods,summa-rizing their research paradigms and the latest advances in autonomous driving decision-making,and discussing the issues and bottlenecks they face when addressing autonomous driving tasks.Following this,a comprehensive review of perfor-mance evaluation methods for end-to-end DRL autonomous driving decision models is provided,covering aspects such as safety,robustness,comfort,efficiency,and reliability,analyzing influencing factors,and summarizing the performance evaluation process.Subsequently,commonly used and open-source virtual simulation platforms for autonomous driving are compared and summarized regarding their features and applicable scenarios.Finally,open issues in performance evalu-ation and prospects for future research on evaluation methods are presented,providing theoretical support and reference for related research and model deployment.关键词
智能交通/无人驾驶/深度强化学习/评测方法/决策性能/端到端控制Key words
smart transportation/autonomous driving/deep reinforcement learning/evaluation methods/decision perfor-mance/end-to-end control分类
航空航天引用本文复制引用
顾同成,徐东伟,孙成巨..无人驾驶深度强化学习决策模型性能评测方法综述[J].计算机工程与应用,2025,61(19):12-42,31.基金项目
车路一体智能交通全国重点实验室开放基金课题(2024-B010) （2024-B010）
国家自然科学基金(62373325,6190334). （62373325,6190334）
