Q-Learning for Autonomous Vehicle Navigation

In this work, we proposed and developed a reinforcement Q-learning method to do the lane-keeping and obstacle evasion driving maneuvers. We detail how to design a simple car simulator and how to use it to do the training. For each problem, we define different states, actions, and reward functions to obtain a Q-table. Next, we use it as a driving maneuver controller in a different simulation environment. With this method, our car successfully droves on a road different to where it was training. An important conclusion is the possibility to build, more complex controllers to do passing or behavior selectors.