Efficient Precision-Driven Scenario Design: Tailoring Collision Type Probabilities for Richer Autonomous Testing

Multi-objective reinforcement learning strategy employs vehicle reachable set optimization are proposed to address two prevalent problems in autonomous vehicle testing: the lack of critical scenarios and the sameness in scenarios generated by traditional methods.The framework prioritizes vehicle dynamics to construct relevant and varied testing scenarios, with a focus on risk triggering states. To streamline the navigation of safety states, we implement a distance-based reward function. Simultaneously, other reward functions balance the frequency and distribution of critical events, drawing on historical trends and Kullback-Leibler divergence for fine-tuning. Our model's efficacy is underpinned by stringent evaluation, showcasing a synergy between training efficiency and scenario variety. Further validation is provided through advanced hardware-in-the-loop simulations, confirming the robustness of our scenario design components in real-world conditions. The proposed method exhibits strong adaptability compared to Reinforcement Learning and Diversity-Driven Exploration, ensuring that the actual probability distribution of accident occurrences closely aligns with the expected distribution $(KL< 0.5)$. Furthermore, it achieves a coverage rate of 98.77% for the environmental states that may lead to accident scenarios, effectively preventing the occurrence of a single dominant critical scenario type in testing.