How autonomous vehicles learn from us

Autonomously driving cars have long since ceased to be science fiction and are already being tested in road traffic at various degrees of autonomy. It is clear that the vehicles must pay attention to the course of the road, road markings and traffic signs and also take into account the current situation, for example construction sites, closures or weather influences. In order to participate safely in traffic, other road users play a central role, be they cars, cyclists or pedestrians. But how does a vehicle know how to behave in any situation? Is this completely describable by abstract rules? Do different reaction possibilities have to be weighed against each other? And how do you deal with situations for which suitable rules are lacking?

In urban traffic in particular, a large number of rules apply, some of which can also conflict with one another. This quickly leads to unnatural and hesitant, often even accident-prone driving behaviour. Machine learning methods could therefore complement the current rule-based methods. The idea is no longer to foresee all possible situations and to intercept them by appropriate rules, but to learn statistically directly from the behaviour of human drivers in numerous situations.

In the “Urban Mobility Assist” project, José María González Pinto and Wolf-Tilo Balke address the question of whether typical tactics for the same strategy can be determined with sufficient accuracy from a larger set of records of individual driving behavior (big data). If there were such dominant tactics, which could also be assigned to different local or currently identifiable situations, this would be a clear benefit for the improvement of autonomous mobility. The scientists focused on a mixture of current driving behavior models and focused on lane changes, intersections and roundabouts in urban areas. First results show that learning typical driving behaviour from vehicle fleet data is indeed possible.