Addressing uncertainty challenges for autonomous driving in real-world environments

Despite ongoing beta tests and research, significant challenges still need to be addressed before fully autonomous driving can be commercialized. At levels 3 or higher, the responsibility for the consequences of autonomous driving shifts from the human driver to the system itself and its OEM (Original Equipment Manufacturer). However, current autonomous vehicles undergoing real-world road tests have been reported to experience numerous failures and accidents, indicating that there is still a long way to go before achieving reliable and safe autonomous driving technology. The development of fully autonomous driving technology has been challenging due to the inherent uncertainty associated with deep learning and the unique domain of autonomous driving. As vehicles are complex systems directly impacting human life, it is crucial to minimize uncertainty in autonomous driving systems to ensure safety and integrity. This uncertainty must be addressed for successful commercialization by designing autonomous vehicles that comply with rigorous safety standards. Our study identified two types of uncertainty in autonomous driving: those related to deep learning and those related to the specific domain of autonomous driving. Uncertainties related to deep learning can be further classified into aleatoric uncertainty caused by noise and randomness in the data and epistemic uncertainty resulting from incomplete knowledge of the models. Uncertainties related to the domain of autonomous driving can be divided into three categories: driver propensity and intention, visibility, and occlusion, and traffic regulation and right of way. In this study, we proposed solutions to address each type of uncertainty and presented an example implementation of an autonomous driving system that incorporates these solutions through a SFF (Safety Force Field) based driving policy, which improves the overall quality of driving automation.