浙江大学学报（理学版）2026，Vol.53Issue(2)：131-147,160,18.DOI:10.3785/1008-9497.25106
基于CARLA仿真的端到端自动驾驶算法综述
A survey of end-to-end autonomous driving algorithms based on CARLA simulation
摘要
Abstract
With recent advances in deep learning and hardware computing power,end-to-end autonomous driving models have attracted significant attention.The core idea is to use neural networks to directly map sensory inputs to vehicle control commands,optimizing the entire driving process as a single task.However,testing autonomous driving models in the real world involves substantial costs in terms of manpower and resources,along with considerable safety risks-particularly covering rare or extreme scenarios,such as sudden pedestrian crossings.As a result,it has become a common practice to train models in simulators and subsequently transfer them to real-world applications.In this context,this paper first surveys the major autonomous driving simulators currently in use and then focuses on the development of end-to-end driving algorithms based on the CARLA simulation platform.Existing algorithms are categorized according to their level of reliance on simulation data,and we review their strengths and differences in terms of model input and output design,data augmentation techniques,and training strategy optimization,helping to clarify their interrelationships.We further discuss simulation scene synthesis methods based on game engines and neural radiance fields.Finally,we analyze the advantages and limitations of current approaches and highlight potential future research directions.关键词
自动驾驶/仿真器/CARLA/端到端/模仿学习/场景合成Key words
autonomous driving/simulator/CARLA/end-to-end/imitation learning/scene synthesis分类
信息技术与安全科学引用本文复制引用
付光明,卢子奥,马雷,王贝贝..基于CARLA仿真的端到端自动驾驶算法综述[J].浙江大学学报（理学版）,2026,53(2):131-147,160,18.基金项目
科技创新2030-"新一代人工智能"重大项目(2022ZD0116305). （2022ZD0116305）
