Toward the robustness of autonomous vehicles in the AI era

In modern transportation, autonomous driving (AD) stands at the forefront of the AI technological revolution, promising to transform the way we commute and interact with urban environments. This shift toward automation not only offers increased convenience and reduced human error in driving but also opens avenues for substantial economic and environmental benefits. However, accidents involving self-driving and driver assistance systems underscore the critical challenges in ensuring their robustness. According to the National Highway Traffic Safety Administration, Tesla's Autopilot has been associated with 17 fatalities and 736 crashes since 2019. Such statistics highlight the urgent need for robustness in AD systems, which operate reliably across diverse real-world scenarios, including varying weather conditions, complex traffic patterns, and unforeseen road events. The criticality of robustness in ensuring public trust and the practical viability on a global scale can never be underestimated. Developing AD systems that can handle these real-world challenges is not only a technological imperative but a societal necessity.

Figure 1 illustrates the key elements and challenges in developing robust AD systems. It highlights the diverse data sources needed, core system functions, and the importance of high-fidelity simulations. It also emphasizes integrating control theory and large language models to enhance system reliability and specificity. Empowerment technologies like deep generative models and reinforcement learning are crucial for balancing innovation and safety toward higher autonomy levels.