Application of foundation models for autonomous driving: a survey of data synthesis

With the evolution of data-driven autonomous driving technology, transferring driving responsibility from humans to machines is now feasible. Addressing the long-tail distribution problem in autonomous driving data is crucial for enhancing safety. Recent advances in foundational models, including large language models (LLMs) and generative AI models, have significantly improved data synthesis capabilities. This paper reviews the application of foundation models in three key areas: sensor data synthesis, traffic flow synthesis, and world models.