Multi-fidelity Bayesian optimization for simulation based autonomous driving systems testing

Simulation-based testing has become a powerful method for uncovering critical failures in Autonomous Driving Systems (ADS). However, the high computational cost of executing realistic driving simulations, which often takes minutes if not hours, at full fidelity significantly limits the scalability of this approach. In this work, we introduce MFBO-Drive, a novel approach that formulates critical scenario generation as a multi-fidelity optimization problem. By adaptively selecting both candidate test scenarios and their corresponding simulation fidelity levels, MFBO-Drive enhances the efficiency of ADS testing without compromising effectiveness. MFBO-Drive is instantiated using the well-established Multi-Fidelity Bayesian Optimization (MFBO) method and incorporates a fidelity exploration control mechanism to balance simulation cost with predictive reliability. We evaluate the approach on over 100,000 driving scenarios in the MetaDrive simulator, comparing it against strong baselines including single-fidelity Bayesian optimization and random search. Experimental results show that MFBO-Drive significantly enhances cost-effectiveness, achieving a 16.8% improvement over state-of-the-art Bayesian optimization, while maintaining comparable test quality. These results highlight its promise for budget-constrained ADS testing in simulation environments.