同济大学 汽车学院//新能源汽车工程中心，上海 201804

交互场景因其高维复杂性和高风险性，成为自动驾驶汽车测试与运行中的关键安全挑战之一。近年来，以数据生成、对抗生成、知识生成及大语言模型为代表的学习型场景生成方法凭借较高的真实性与覆盖广度，在提升交互场景生成质量与测试验证效率方面展现出显著优势。为系统性梳理该领域的研究进展、核心技术与瓶颈问题，本文对自动驾驶交互场景的学习型生成方法进行全面综述，旨在为后续研究提供清晰的技术脉络与发展方向。首先，概述交互场景的基本特征，并对主流公开数据集进行对比分析，探讨其对学习型生成方法的支撑作用。同时，总结交互相关的标准规范、场景描述与衡量指标，分析其对交互行为表达能力的影响。其次，从传统学习方法与大语言模型方法两个方向出发，归纳各类方法的技术路线与生成效果。最后，结合现有成果，总结当前学习型生成方法在数据和方法等方面面临的挑战，展望未来的研究方向。研究发现，传统学习方法在真实性与多样性的平衡、安全关键场景的高效发现与生成方面存在固有挑战 。大语言模型方法虽在场景语义理解和复杂逻辑构建上展现出卓越潜力，但普遍面临推理速度慢、信息虚构、生成内容与物理世界对齐困难以及决策过程“黑盒”特性导致的可解释性不足等瓶颈。此外，当前研究仍共同受制于现有数据集缺乏交互关键信息、场景严重性缺少统一衡量标准等基础性问题。本文认为研究表明，尽管已有方法在交互建模上取得了初步成果，但仍需进一步提升生成质量与泛化能力，以满足自动驾驶汽车实际需求。

Interactive traffic scenarios—characterized by high-dimensional complexity and elevated safety stakes—constitute a foremost validation bottleneck for autonomous vehicles (AVs). Recent learning-based scenario-generation paradigms—encompassing data-driven synthesis, adversarial generation, knowledge-guided generation, and large language models (LLMs)—have demonstrated superior fidelity and coverage, substantially improving both the quality of interactive-scenario corpora and the efficiency of safety-critical testing. To systematically chart the advances, core techniques, and remaining impediments in this domain, this paper presents a comprehensive survey of learning-oriented approaches to generating interactive scenarios for AVs, delineating a coherent technical roadmap and future research directions.First, we dissect the fundamental attributes of interactive scenarios and benchmark prevailing open datasets, elucidating their respective capacities to underpin learning-based generators. Concurrently, we review pertinent standards, formal scenario-description languages, and evaluation metrics, analyzing their expressive power over interactive behaviors. Second, we taxonomize existing methods into two strands—conventional learning frameworks and LLM-based paradigms—synthesizing their algorithmic pipelines and empirical generation performance. Finally, we distill cross-cutting challenges residing in data curation and methodological design, and outline prospective research avenues.Our analysis reveals that traditional learning methods struggle to simultaneously guarantee realism and diversity, and remain inefficient in discovering and synthesizing safety-critical edge cases. LLM-based approaches, while exhibiting remarkable proficiency in semantic comprehension and complex logical composition, suffer from sluggish inference, hallucination, misalignment with physical plausibility, and opaque decision-making that undermines interpretability. Moreover, the field at large is constrained by datasets that lack fine-grained interaction-critical annotations and by the absence of unified severity metrics for scenario endangerment. We contend that, despite preliminary successes in interactive modeling, substantial enhancements in generation fidelity and generalizability are imperative to meet the stringent demands of real-world AV deployment.