Training an Autonomous Vehicle

Our goal is to train a model car to maneuver autonomously through a cluttered lab or hallway. Because of the probably complicated structure of any real-world environment, programming the vehicle using conventional means is implausible. Instead a better method to train the vehicle is to let the vehicle encounter the environment and update its decision making process based on past actions. In this situation, called reinforcement learning, the vehicle needs to interact within the environment multiple times in a variety of ways (e.g., go straight or turn left or slow down) in order to determine the best approach to achieve the final reward. In this research we consider two methods to train the vehicle in simulation in order to get good results. In the first method, we train the vehicle on complicated (that is, lots of turns) tracks and then test on a simple track. In the second method, we randomize the background (use different colors) that the vehicle “sees” with its camera during training with the assumption that the vehicle will learn to ignore the background and focus solely on the track. Our results show that training on complicated tracks gives improved performance. There is improved performance on the real track but not in simulation as the amount of randomization training increases