Automatically generating complex driving scenarios for testing autonomous driving systems

Abstract

Testing Autonomous Driving Systems (ADS) is crucial for ensuring safety and reliability, as driving errors, such as collisions impacting human safety or violations of traffic regulations, can be highly critical. Due to the critical importance of safety issues during testing, Software-in-the-Loop Simulation (SILS) has emerged as a valuable and crucial tool for efficient and safe ADS testing. SILS facilitates virtual testing of ADS, offering efficiency in a simulated environment without the need for physical hardware. Most of the existing work are based on the two predominant approaches for seed driving scenario set determination in SILS testing: Manual Crafting and Generation-based Fuzzing. Manual Crafting allows targeted scenarios; however this approach is highly labor-intensive and does not effectively discover corner cases. Generation-based Fuzzing is used to automatically generate driving scenarios, but it faces challenges in route generation and dynamic entity handling. This paper introduces a method to construct Complex Scenarios by combining all layers of the scenario layer model, addressing three challenges in existing approaches. The contributions include a Coverage-based Route Generation approach, ensuring comprehensive coverage of lanes in HD-Map (High-Definition Map); a Geography-based Route Selection method, finding geographically rare corner cases; and Route-Aware NPC Pedestrian Generation, more effective pedestrian generation using ego-vehicle's route. Experimental results demonstrate the methodology's superiority in lane coverage and road characteristics coverage, generating geographically rare corner cases, and achieving a higher interaction rate in NPC generation. The approach contributes to scenario diversity, broader area testing automation, and improved safety violation detection, marking advancements in autonomous vehicle testing methodologies. In summary, the paper's contributions extend beyond conventional methods, offering effective solutions for generating diverse corner case scenarios, automating testing over larger areas, and enhancing safety violation detection. The methodology proves valuable for advancing autonomous vehicle testing, providing realistic and reliable scenarios for thorough evaluation.