Navigating a road network using Reinforcement Learning

Agent based models are increasingly utilised in game development and film making because they allow the creation of large scale content with minimal effort. We use Reinforcement Learning and a curriculum learning approach to train a single model for agents navigating a road network encompassing different types of intersections, to respect traffic regulations. We train agents to navigate two game scenes: a two-way crossroad intersection governed by traffic lights and a four-way roundabout intersection, and we discuss the performance of the trained model with respect to robustness and success rates.