High-dimensional functional boundaries search for deviation-robust testing of autonomous driving system

Abstract

Testing and evaluation are essential for verifying the safety of the intended function (SOTIF) of autonomous driving systems (ADS), which focuses on estimating the system’s functional boundaries through a limited set of tests to assess its safe operational range. To achieve this, a series of valuable safety-margin scenarios must be designed as test cases. However, scenario testing faces the dilemma of the curse of dimensionality and the requirements for test coverage. Consequently, the construction and selection of test cases become significant challenges. Moreover, due to the black-box nature of the system under test (SUT), surrogate models are often introduced during the scenario generation process, which can introduce model deviation relative to the actual system and potentially lead to ineffective test scenarios as well as incorrect estimation of system functional boundaries (SFB). To address these challenges, an efficient framework for generating high-dimensional safety-margin scenarios and tracking SFB of SUT is proposed, which utilizes a baseline surrogate model to generate a diverse and comprehensive library of safety-margin test scenarios through a multi-population genetic algorithm (MPGA). Additionally, a System Functional Boundary Tracking (SFBT) module is employed to compensate for the deviation between the baseline surrogate model and the actual SUT, thereby adaptively generalizing the library of critical scenarios to estimate its high-dimensional functional boundaries. This framework will potentially assist in the testing and validation of the Operational Design Domain (ODD) for ADS.