Constraint- and Risk-Driven Search for Safety-Critical Scenarios in Autonomous Driving Simulation

Efficient discovery of safety-critical scenarios is a central problem in autonomous-driving simulation because safety-critical events are rare and the scenario parameter space is often high-dimensional. Direct random sampling can spend most of the simulation budget on samples that are weakly relevant, physically unreachable, or behaviorally inconsistent. This study proposes a constraint- and risk-driven framework for safety-critical scenario discovery. The framework first constructs a valid scenario space using physical reachability, behavioral consistency, and traffic-feasibility constraints. It then evaluates valid simulated samples with a hierarchical scenario-value model that combines kinematic criticality, longitudinal controllability risk, and scenario diversity. Finally, a risk-feedback probabilistic search updates the sampling distribution using high-value samples. The method is evaluated primarily in a lead-vehicle hard-braking scenario and additionally in a cut-in scenario. In the lead-vehicle hard-braking scenario, the proposed method improves the discovery rate from 0.68% for Random Sampling to 49.62% under a budget of 1000 evaluations. In the cut-in scenario, the discovery rate improves from 5.28% to 67.68% under the same budget. Ablation results show that constraint-aware construction reduces potentially invalid samples, while the full hierarchical scenario-value model improves coverage of discovered safety-critical samples.