Generative AI Driven Simulation for
Virtual Testing of Autonomous Vehicles

Systematic safety validation of autonomous vehicles (AVs) demands exposure to diverse, realistic, and safety-critical scenarios—particularly rare long-tail events that are prohibitively difficult and costly to collect at sufficient scale through real-world operations. This tutorial provides a comprehensive introduction to generative AI–driven simulation for virtual AV testing, tracing the full pipeline from large-scale naturalistic environment modeling to photorealistic sensor synthesis.

The tutorial begins with TeraSim, an integrated simulation platform encompassing three tightly coupled capabilities: city-scale naturalistic driving environment modeling with statistical realism, safety-critical scenario synthesis informed by real-world crash data and large language models, and closed-loop virtual testing infrastructure. The tutorial then covers state-of-the-art generative models from NVIDIA—NuRec, which enables photorealistic digital twin construction via 3D Gaussian reconstruction from real-world sensor data, and Cosmos, a world foundation model for high-fidelity multi-view driving video generation—that together push the realism frontier of AV simulation.

Through invited talks from leading researchers across academia and industry, this tutorial equips participants with both a conceptual framework and practical knowledge of how generative AI is reshaping AV simulation: from rule-based, graphics-driven environments to data-grounded, continuously evolving digital twins that support rigorous and scalable safety research.