2024 AutoTest Technical Conference

The rapid advancements in automated and connected vehicle technologies requires robust validation frameworks to ensure their safety and efficacy before deployment. This study introduces a novel Deep Reinforcement Learning (DRL) based methodology for the virtual validation of autonomous vehicles within any given virtual environment. Our approach leverages state-of-the-art DRL techniques to systematically generate and evaluate critical driving scenarios that autonomous vehicles may encounter. Techniques such as Deep Q-Networks (DQN), Proximal Policy Optimization (PPO), Soft/Distributional Actor-Critic (XAC), Deep Deterministic Policy Gradient (DDPG) algorithms are incorporated to various selected dynamic traffic participants within the given virtual environment, to develop a scalable testing framework that dynamically adapts to behaviour of the device-under-test (DUT)-vehicle integrated with the test-object, facilitating a comprehensive assessment of autonomous driving (AD) functions. We configure the dynamic participants in the simulation environment to obey certain pre-defined rules and boundary conditions to generate interesting or critical scenarios with a broad spectrum of traffic situations. Our methodology is validated through targeted scenarios, which can demonstrate the ability or disability of AD functions to navigate complex traffic environments safely and efficiently. Our findings indicate that the proposed DRL-based framework not only enhances the precision and depth of virtual validation processes but also reduces the dependency on expensive and time-consuming Naturalistic field operational tests (N-FOT) and repetitive replay or XiL tests in laboratory. Each scenario generated and executed through our AI-Core DRL Framework is stored in standard formats such as OpenSCENARIO DSL, which can be reused in any form of XiL testing for given driving functions. The application of our approach significantly reduces the unnecessary execution of repetitive stored scenarios but generates and executes targeted scenarios for the test object, resulting in significantly faster validation cycles. Additionally this approach can also be extended for testing various software and embedded components in diverse domains.