HSPG: An open-loop testing framework for autonomous driving based on proactive generation of hazardous scenario

Autonomous driving algorithms struggle to achieve sufficient coverage of long-tail scenarios in complex traffic environments, primarily due to the scarcity of high-risk samples in real-world data. Existing scenario generation methods also have limitations, as they mostly rely on trajectory perturbation without realistic perception support. To address this issue, we propose the HSPG (Hazardous Scenario Proactive Generation) framework, a proactive hazardous scenario generation approach based on naturalistic driving data. HSPG systematically amplifies potential risks through structural perturbations of original traffic scenarios. A sliding-window-based risk index is introduced to automatically identify interaction-intensive periods and extract candidate scenarios. A high-risk vehicle detection mechanism then selects critical surrounding vehicles as interaction agents. By integrating a Linear Quadratic Regulator (LQR) with Recurrent Posterior Policy Optimization (RPPO) and adversarial strategies, high-risk trajectories are generated. These trajectories are further transformed into realistic street scenarios via an image synthesis module coupled with real-world map data, forming a comprehensive safety-critical test dataset. Experimental results demonstrate that HSPG effectively identifies latent risks, enhances collision likelihood by at least an order of magnitude under autonomous driving test models, and generalizes across diverse scenarios. A dataset comprising 150 scenarios, 6019 samples, and six multi-perspective camera views has been constructed, providing a valuable benchmark for safety evaluation in autonomous driving. Our dataset can be found at https://huggingface.co/datasets/gitchee/nuScenes-Atk.