In order to meet the verification requirements of L4/L5 autonomous driving algorithm for extreme scenes and fast iterations, this paper proposes an integrated framework of high-fidelity virtual test scene generation, smart car behavior modeling and closed-loop verification. Firstly, the C-LSTM-GAN model is constructed. Under the condition of real traffic flow, LSTM is used to maintain the consistency of time series, and kinematics-collision physical verification is embedded to realize the automatic generation of dynamic scenes with high fidelity of statistics and physics. Second, we designed a hierarchical reinforcement learning (HRL) architecture: the upper-layer proximal policy optimization (PPO) algorithm performs global path planning within a discrete action space, with a reward function that integrates safety, efficiency, comfort, and compliance. Lower Model Predictive Control (MPC) solves the optimal control based on the bicycle model, ensures the smooth and feasible local trajectory, and introduces attention mechanism into state coding to improve the interpretability of decision. Thirdly, an automatic scene generation tool based on combinatorial testing is developed, and quantitative evaluation indicators are established from multiple dimensions such as function, comfort and personification, and a Carla-to-real vehicle comparison channel is built to reproduce key scenes such as "sudden braking in front of the vehicle" in a closed site to verify the simulation fidelity. Experimental results demonstrate that the proposed method achieves a Frechet Inception Distance (FID) of 15.8 and 22.1 in highway and urban scenarios, respectively, significantly outperforming rule-based generation and standard GAN. In cut-in tests, it achieves zero collisions, a minimum time-to-collision (TTC) of 3.8 seconds, an acceleration root mean square (RMS) of 1.08 m/s², and a 100% task success rate. The braking trajectory error between the virtual vehicle and the real vehicle is less than 0.5 m, and the deceleration error is less than 0.15 m/s². The framework provides a low-cost, reproducible and continuously evolving virtual test closed loop for autonomous driving algorithm, which effectively makes up for the shortcomings of real vehicle test in long tail, danger and iteration speed.
Zhou et al. (Sun,) studied this question.