In the future, autonomous vehicles are expected to safely move people and cargo around. However, as of now, automated entities do not necessarily outperform human drivers under all circumstances, particularly under certain road and environmental factors such as bright light, heavy rain, poor quality of road and traffic signs, etc. Therefore, in certain conditions it is safer for the driver to take over the control of the vehicle. However, switching control back and forth between the human driver and the automated driving entity may itself pose a short-term, elevated risk, particularly because of the out of the loop (OOTL) issue for humans. In this study, we develop a mathematical framework to determine the optimal driving-entity switching policy between the automated driving entity and the human driver. Specifically, we develop a Markov decision process (MDP) model to prescribe the entity in charge to minimize the expected safety cost of a trip, considering the dynamic changes of the road/environment during the trip. In addition, we develop a partially observable Markov decision process (POMDP) model to accommodate the fact that the risk posed by the immediate road/environment may only be partially observed. We conduct extensive numerical experiments and thorough sensitivity and robustness analyses, where we also compare the expected safety cost of trips under the optimal and single driving entity policies. In addition, we quantify the risks associated with the policies, as well as the impact of miss-estimating road/environment condition risk level by the driving entities, and provide insights. The proposed frameworks can be used as a policy tool to identify factors that can render a region suitable for level four autonomy.

未来，自动驾驶汽车有望安全运送人员和货物。但是，到目前为止，自动化实体并不一定在所有情况下都胜过人类驾驶员，特别是在某些道路和环境因素下，例如强光，大雨，道路和交通标志的质量较差等。因此，在某些情况下让驾驶员更安全地接管车辆的控制权。但是，在人类驾驶员和自动驾驶实体之间来回切换控制本身可能会带来短期的高风险，特别是因为人类的循环外（OOTL）问题。在这项研究中，我们开发了一个数学框架来确定自动驾驶实体和驾驶员之间的最佳驾驶实体切换策略。特别，考虑到旅途中道路/环境的动态变化，我们开发了一个马尔可夫决策过程（MDP）模型来规定负责实体以最小化旅途的预期安全成本。此外，我们开发了部分可观察的马尔可夫决策过程（POMDP）模型，以适应以下事实：直接观察道路/环境带来的风险可能仅能部分观察到。我们进行了广泛的数值实验，并进行了全面的敏感性和鲁棒性分析，在此基础上，我们还比较了最佳和单一驾驶实体政策下的出行预期安全成本。此外，我们将量化与政策相关的风险，以及对驾驶实体未正确估计道路/环境状况风险水平的影响，并提供见解。