Decision making for autonomous vehicles: Combining safety and optimality

In this paper, we propose a novel decision maker design for an autonomous vehicle driving on a highway, considering safety and optimality, and which is scalable, i.e., remains computationally tractable for more complex situations. This is realized in two stages. First, all safe actions are found, and second, from these actions the optimal action is selected, according to (weighted) criteria that capture safety, comfort and efficiency. The design combines rule-based safety checks with the solution of a Markov decision process, found through a tree search algorithm, to fulfill the safe, smart and scalable requirements of the decision maker. The design is validated in simulation using eight different scenarios. The performance of the new design is compared to the performance of a rule-based controller. This comparison is done using three performance criteria that aim to capture safety, efficiency and comfort.