Accelerated Verification of Autonomous Driving Systems based on Adaptive Subset Simulation

Authors: Ye Tian, Aohui Fu, He Zhang, Lanyue Tang, Jian Sun

Abstract:
Safety verification of Autonomous Driving Systems (ADS) is essential to ensure reliable performance in naturalistic traffic environments. Existing test automation methods such as Importance Sampling often suffer from poor efficiency in high-dimensional scenario spaces. This paper proposes Adaptive Subset Simulation (ADSS), an accelerated verification framework for ADS that iteratively identifies more critical scenarios while significantly reducing testing cost. ADSS is evaluated on two systems under test—Baidu Apollo and the Intelligent Driver Model (IDM)—using both 3-dimensional car-following and 6-dimensional cut-in scenarios constructed in LGSVL. Experimental results demonstrate that ADSS achieves substantially higher efficiency in estimating rare-event collision rates compared with conventional approaches, particularly as scenario dimensionality increases.

[IEEE Xplore]

Recommended citation: Tian, Y., Fu, A., Zhang, H., Tang, L., & Sun, J. (2024). Accelerated Verification of Autonomous Driving Systems based on Adaptive Subset Simulation. IEEE Transactions on Intelligent Vehicles.
Download Paper