The use of infrastructure-to-vehicle communication technologies can enable improved energy-efficient autonomous driving. Traditional ecological velocity planning methods have a high computational burden, particularly when plug-in hybrid electric vehicles are considered. Consequently, to retrieve an optimal velocity profile in real-time, it is necessary to rely on significant approximations. The aforementioned issue can be addressed by exploiting deep reinforcement learning to "understand" an eco-driving velocity planner within a model-free approach. However, the integration of machine learning algorithms in contexts where safety is crucial, such as autonomous driving, requires the adoption of specific preventive measures. Therefore, the coupling of deep reinforcement learning with state-of-the-art safety controller is proposed to extend the ecological velocity planning to an urban context, in which compliance with the traffic lights needs to be guaranteed. The discussion is supported by simulative results which statistically demonstrate that the learning-based algorithm outperforms two benchmark controllers, and it can generalize across a variety of intersection configurations.

Download the slides for this talk.Download ( PDF, 1678.0 MB)