关键词: 自动驾驶汽车, 交互场景, 交互数据集, 交互场景描述, 数据生成, 对抗生成, 知识生成, 大语言模型

Abstract:

Interactive traffic scenarios, characterized by high-dimensional complexity and inherent risk, constitute a critical safety challenge in the testing and operation of autonomous vehicles (AVs). In recent years, learning-based generative method—exemplified by data generation, adversarial generation knowledge-driven generation, and large language models (LLMs)—has demonstrated significant advantages in improving the quality of interactive scenario generation and the efficiency of testing and validation, owing to their superior realism and coverage breadth. To systematically chart the advances, core techniques, and bottleneck issues in this domain, this paper presents a comprehensive review of learning-based generative methods for autonomous vehicle interactive scenarios, aiming to offer a clear technical roadmap and development direction for subsequent research. First, the fundamental attributes of interactive scenarios are outlined, and mainstream public datasets are comparatively analyzed to discuss their supportive role for learning-based generative methods. Concurrently, relevant standards, scenario description frameworks, and evaluation metrics for interaction are summarized, with an analysis of their impact on the expressive capacity of intertive behaviors. Second, the technical approaches and generation outcomes of various methods are categorized along two main directions: traditional learning-based methods and large language model-based me-thods. Finally, building upon existing findings, the article summarizes the current challenges faced by learning-based generative methods, both in terms of data and methodology, and proposes future research directions. The study finds that traditional learning-based methods face inherent challenges in balancing realism and diversity, as well as in the efficient discovery and generation of safety-critical scenarios. While LLM-based approaches show exceptional potential in semantic understanding and constructing complex logical narratives, they commonly encounter bottlenecks such as slow inference speeds, information hallucination, difficulties in aligning generated content with physical world constraints, and a lack of interpretability due to the “black-box” nature of their decision processes. Moreover, current research is still constrained by fundamental issues, including the scarcity of interaction-critical information in existing datasets and the absence of a unified metric for quantifying scenario criticality. This article concludes that although existing methods have achieved preliminary success in interaction modeling, significant improvements in generation quality and generalization capability are still required to meet the practical demands of autonomous vehicle development.

Key words: autonomous vehicles, interactive scenario, interactive dataset, interactive scenario description, data generation, adversarial generation, knowledge generation, large language model