Decision making in partially observable domains: autonomous driving applications

The advancement in the technology of autonomous vehicles has made possible to manage most urban and extra-urban scenarios with improvements in terms of performance and vehicle safety. A challenge still open regards the management of the autonomous vehicle in partially observable environments, such as an intersection of limited visibility due to the presence of an obstacle. After a careful analysis of the various methods present in literature, in this thesis work are presented two different methods to handle an autonomous vehicle approaching an urban occluded intersection. The first method consists in evaluating the collision risk, taking into account the presence of a possible phantom occluded vehicle with worst-case-like behavior, and in developing an intuitive decision-making algorithm. The second method models the urban scenario as a continuous Partially Observable Markov Decision Process (POMDP) and employs a continuous online solver (POMCP) to find the best policy for the autonomous vehicle. Comparing the two methods, it turned out that the one based on the POMDP allows to obtain, from the point of view of speed profile and safety, better performances compared to the collision risk method.