Project Summary
Artificial Intelligence (AI) has been bringing revolutionary changes to transportation systems, from single- vehicle autonomy such as robo-taxi to potential collaborative driving autonomy among connected vehicles and infrastructure. However, latest AI technologies, including the latest generative ones, are known to be fundamentally limited in their generalizability and adaptability to complex, dynamic, and long-tailed driving scenarios, making their safety risks and resilience in real-world deployment scenarios questionable. In this project, the research aims to systematically understand and address the AI safety and resilience challenges in the latest AI technology penetration in the intelligent transportation systems sector. Specifically, the research aims to address such challenges on both the single-vehicle autonomy and collaborative driving autonomy fronts. For the former, the research team plans to focus on the latest large Vision-Language Model (VLM) / Vision- Language-Action (VLA) model based autonomous vehicle technology, which the team found to face direct driving safety risks in real-world driving scenarios despite their widely-believed and demonstrated general driving performance potential (already at deployment stage for Waymo, Mercedes-Benz, Wayve). For the latter, the research team plans to study the latest collaborative perception systems with the goal of addressing the research gaps in both understanding and provably addressing their safety and resilience under shared perception data corruptions.