Retrieving a driving model based on clustered intersection data

In order for autonomous vehicles to learn how to naturally navigate through an intersection, we present a method of learning from expert drivers using inverse reinforcement learning. We cluster different intersections according to how experts drive through them and train a model based on each class of intersection. With these models to choose from, as the agent encounters various situations, it can retrieve the corresponding model and execute the most fit policy. The models were trained from 308 trajectories recorded from 5 different drivers at 3 intersections. These clusters were then used to classify a further 5 more intersections. The approach is evaluated in a simulator where we provide a global path to be followed by the vehicle where at each intersection it drives according to the trained policies.