AI Generates Driving Scenarios from Real-World Failure Records.

Researchers have developed a novel pipeline for generating testing scenarios for Autonomous Driving Systems (ADS) by drawing insights from real-world failure records. Current simulation methods often rely on fixed scenario representations or extensive manual effort to design test templates. This new approach utilizes Large Language Models (LLMs) to synthesize diverse and accurate scenarios from natural language descriptions of historical crashes, such as those found in NHTSA records. The pipeline is modular and generates synthetic scenarios compatible with specific testing constraints. It was successfully applied to create a varied set of scenarios for autonomous navigation testing on the Metadrive simulator. These generated scenarios, which included combinations of road types, non-ego vehicle movements, and on-road anomalies like work zones, aligned well with specified testing conditions. Crucially, the method revealed interesting system failures within a small budget of only 20 scenarios, demonstrating an efficient way to uncover latent vulnerabilities.