MontiSim: Agent-Based Simulation for Reinforcement Learning of Autonomous Driving

Reinforcement learning is a machine learning method particularly interesting for the autonomous driving domain, as it enables autopilot training without the need for large and expensive amounts of manually labeled training data. Instead, agents are trained by evaluating the effects of their actions and punishing or rewarding them accordingly. In autonomous and particularly cooperative driving a core problem is however that multiple vehicles need to be trained in parallel while having an impact on each other's behavior. In this paper, we present a simulation solution providing cooperative training capabilities out-of-the-box and compare the quality of the resulting autopilots in an intersection scenario.