Safety behavior abstraction and model evolution in autonomous driving

In the autonomous driving systems (ADSs) literature, most existing approaches primarily focus on identifying driving scenarios, which is challenged by the reality that real-world driving scenarios are countless and unpredictable, and it is impossible to have a comprehensive set of driving scenarios to demonstrate the test efficiency in covering all possible situations ADSs might face. To address these challenges, one fundamental step is to abstract complex ADS behaviors, e.g., (semi-)automatically derive a holistic view of how an ADS behaves under its driving environment with high-level representations, such as prior-knowledge-based models. Therefore, in this paper, we propose a novel Risk-basEd Model comprehension and Evolution approach for autonomous Driving sYstems, named REMEDY, which facilitates the development of such models and enables automated model evolution (i.e., discovering and extracting ADS behaviors and their interactions with the environment) via model execution and simulation with the autonomous driving simulator CARLA. To enable efficient model evolution, we also equipped REMEDY with a risk-based strategy using Q-Learning, which is empirically evaluated by comparing it with three baselines (i.e., a random strategy, a coverage-based greedy strategy, and DeepCollision—a state-of-the-art approach). Results show that REMEDY is capable of discovering new and diverse behaviors, and the risk-based strategy is efficient in discovering risky ADS behaviors.