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

Area of Science:

• Computer Science
• Robotics
• Artificial Intelligence

Background:

• Autonomous driving systems face challenges in covering rare, high-risk scenarios in real-world data.
• Current scenario generation methods lack realistic perception and focus on simple trajectory perturbations.

Purpose of the Study:

• To introduce a proactive framework for generating hazardous driving scenarios to enhance autonomous driving safety.
• To address the scarcity of critical data for testing autonomous vehicles in complex environments.

Main Methods:

• Developed the Hazardous Scenario Proactive Generation (HSPG) framework using naturalistic driving data.
• Utilized a sliding-window risk index and high-risk vehicle detection to identify critical interactions.
• Employed Linear Quadratic Regulator (LQR) with Recurrent Posterior Policy Optimization (RPPO) and adversarial strategies for trajectory generation.
• Integrated image synthesis with real-world map data to create realistic street scenarios.

Main Results:

• HSPG effectively identifies latent risks and amplifies collision likelihood by over an order of magnitude in test models.
• The framework demonstrates generalization across diverse traffic scenarios.
• A comprehensive dataset with 150 scenarios, 6019 samples, and multi-perspective views was created.

Conclusions:

• The HSPG framework provides a robust method for generating safety-critical test data for autonomous driving.
• The created dataset serves as a valuable benchmark for evaluating and improving autonomous driving system safety.