Learning From Failures: LLM-Guided Safety-Critical Scenario Recommendation for Self-Evolving Autonomous Driving

Autonomous vehicles are expected to operate safely in complex environments, yet failures can still arise under rare and safety-critical conditions. Existing validation pipelines often rely on large-scale simulation and manual scenario design but lack the ability to semantically interpret failure events and respond with adaptive improvements. This article presents a self-evolving framework enhanced by AI capabilities, which leverages large language models (LLMs) to analyze failure cases, recommend semantically relevant test scenarios, and adapt affected modules through targeted few-shot learning. Structured execution logs are converted into semantic representations, which are used to retrieve safety-critical scenarios from a scenario bank. These scenarios enable reliable reproduction of failure behaviors and support few-shot adaptation of the affected module. A closed-loop evaluation confirms behavioral improvement while ensuring system stability. The framework has been extensively evaluated across a diverse set of driving tasks, consistently demonstrating enhanced robustness with minimal human supervision. By integrating semantic failure understanding, scenario-based testing, and data-efficient adaptation, this approach offers a scalable and generalizable solution for failure-aware improvement in safety-critical autonomous driving systems.