Enhancing semantic and risk controllability in safety-critical scenario generation: An LLM-guided conditional...

Safety-critical scenario generation is essential for autonomous vehicle validation. However, existing methods often struggle to simultaneously achieve semantic controllability, risk-level controllability, and trajectory realism. To address this limitation, we propose a Large Language Model (LLM)-guided conditional diffusion framework for controllable safety-critical scenario generation. Specifically, we first use an LLM to convert natural-language descriptions into structured generation conditions. These conditions are then incorporated into a conditional latent diffusion model to guide scenario generation with controllable semantic and risk features. The generated candidate scenarios are further filtered to retain those with high semantic consistency. Experiments are conducted on the highD dataset for car-following and cut-in scenarios. The results show that the proposed method can generate realistic and risk-controllable scenarios. Compared with representative baseline methods, it achieves a better balance among criticality, realism, and controllability. Ablation experiments further demonstrate that LLM-based semantic conditioning improves both scenario-type and risk-level accuracy. The proposed method provides a flexible and controllable tool to generate safety-critical scenarios for autonomous vehicle testing and safety evaluation.