Testing Autonomous Driving Systems Through Blind-Spot Guided Fuzzing

Autonomous Driving Systems (ADS) must reliably perceive and react to complex environments, even when sensor blind spots obscure critical objects. While existing testing methods often focus on dynamic interactions, they significantly underestimate safety risks arising from both dynamic occlusions (e.g., vehicles) and persistent static occlusions (e.g., buildings). This research proposal introduces a novel, unified framework for occlusion-aware ADS testing. We present Blind-Spot Guided Fuzzing, a technique that systematically generates critical test scenarios by leveraging Large Language Models (LLMs) to synthesize realistic seeds from accident reports and employs multi-objective optimization to evolve them. This approach is implemented in our Occlusion-Sensitive Fuzzing (OS-Fuzz) framework, which encompasses two specialized modules: BlindSpotFuzz (BSF) for dynamic occlusions and StaticOcclu-Fuzz (SOF) for static environmental occlusions that persistently hide Vulnerable Road Users. At its core, OS-Fuzz integrates a generalized occlusion model and innovative metrics to guide test generation and quantify the exploration of obscured inputs. Preliminary results against Apollo show BSF identifies over 50% more blind-spot-related collisions. Our comprehensive evaluation plan will benchmark OS-Fuzz against state-of-the-art techniques, rigorously analyzing the relative impact of static versus dynamic occlusions. This research aims to significantly enhance ADS safety by incorporating the critical dimension of sensor and environmental limitations into automated testing.