Targeted testing scenario generation for autonomous vehicles based on quantitative evaluation feedback

Scenario generation for Autonomous Vehicle (AV) testing has advanced, yet existing methods often have limited capability to steer scenario generation toward predefined safety objectives and are difficult to adapt when evaluation requirements change. This study proposes a feedback-driven framework for targeted AV testing scenario generation, in which scenario synthesis is guided by the Comprehensive Evaluation Index (CEI). The framework integrates a Continuous Conditional Generative Adversarial Network (CcGAN) with an iterative feedback mechanism, enabling scenario generation to be adaptively steered toward predefined evaluation objectives. By using CEI as a continuous conditioning signal that jointly characterises scenario risk, complexity, and rarity, the proposed method generates scenarios with controllable safety-critical characteristics. The iterative feedback loop further adjusts the target-label sampling distribution according to the deviation between target and realised scenario criticality. Counterfactual simulation results show that the feedback-driven scenarios exhibit higher CEI levels and crash rates than the non-feedback baseline, indicating that the proposed framework can more effectively concentrate scenario generation in predefined high-criticality regions for AV safety testing.