SaFeR: Safety-Critical Scenario Generation for Autonomous Driving Test via Feasibility-Constrained Token Resampling

Researchers propose SaFeR, a new AI system for generating safety-critical scenarios to test autonomous driving systems. The approach uses transformer-based models with a novel resampling strategy to balance adversarial testing, physical feasibility, and realistic behavior in autonomous vehicle simulations.

• →SaFeR addresses three conflicting objectives in autonomous driving testing: adversarial criticality, physical feasibility, and behavioral realism.
• →The system uses a transformer-based model with differential attention mechanism to capture naturalistic driving distributions.
• →A novel resampling strategy maintains naturalistic behavior while generating adversarial scenarios within feasible parameters.
• →The approach prevents generation of theoretically inevitable collisions through offline reinforcement learning approximation.
• →Testing on Waymo and nuPlan datasets shows superior performance compared to existing baseline methods.