End-to-End Intersection Handling using Multi-Agent Deep Reinforcement Learning

Navigating through intersections is one of the main challenging tasks for an\nautonomous vehicle. However, for the majority of intersections regulated by\ntraffic lights, the problem could be solved by a simple rule-based method in\nwhich the autonomous vehicle behavior is closely related to the traffic light\nstates. In this work, we focus on the implementation of a system able to\nnavigate through intersections where only traffic signs are provided. We\npropose a multi-agent system using a continuous, model-free Deep Reinforcement\nLearning algorithm used to train a neural network for predicting both the\nacceleration and the steering angle at each time step. We demonstrate that\nagents learn both the basic rules needed to handle intersections by\nunderstanding the priorities of other learners inside the environment, and to\ndrive safely along their paths. Moreover, a comparison between our system and a\nrule-based method proves that our model achieves better results especially with\ndense traffic conditions. Finally, we test our system on real world scenarios\nusing real recorded traffic data, proving that our module is able to generalize\nboth to unseen environments and to different traffic conditions.\n