Learning Safety-Critical Scenarios from Real-World Pre-Crash Data for Autonomous Driving Safety Validation

A major challenge in Autonomous Vehicle (AV) safety validation is the scarcity of safety-critical interactions and the difficulty of modeling their underlying decision dynamics. We propose CrashGAIL, a conditional adversarial imitation learning framework that learns interaction-conditioned policies from real-world pre-crash trajectories and synthesizes high-risk two-vehicle scenarios under four interpretable motion-direction conditions. CrashGAIL combines (i) a dual-agent conditional policy architecture, (ii) a VAE warm-up for stabilizing latent representations, and (iii) an auxiliary condition-classification head to improve condition-behavior consistency, followed by PPO-based adversarial policy optimization. Experiments on 816 in-depth crash cases show improved kinematic fidelity while maintaining realistic geometric diversity compared with a strong time-series generative baseline. In counterfactual simulation with the Baidu Apollo stack, CrashGAIL-generated scenarios yield a higher overall crash involvement rate and expose failure modes that are less visible under replay-based testing.