Towards development-oriented testing and evaluation of ADS: A review of current practices and future trends

As autonomous driving systems (ADS) evolve, safety of the intended functionality (SOTIF) has become an increasingly critical concern. Testing and evaluation play a fundamental role in the process, primarily serving to expose system failures and verify functional correctness. Current practices, such as data-injection testing and scenario-based validation, are well established for these purposes. However, with the emergence of advanced architectures such as end-to-end (E2E) learning systems and large foundation models, conventional testing methodologies are increasingly insufficient to support the iterative and complex development cycles of modern ADS. In response, this paper focuses on development-oriented testing and evaluation, which emphasizes leveraging intelligent testing methods to guide the adaptive and personalized improvement of the system under test (SUT). From this perspective, we review existing practices and highlight emerging trends across three key dimensions: (1) challenges in black-box testing posed by the evolution of system architectures; (2) system boundary exploration and tracking for efficient scenario generation; and (3) personalized, model-aware testing strategies to support continuous development of ADS. Finally, we discuss open challenges and future directions for building intelligent, development-driven testing frameworks for next-generation autonomous vehicles.