Poster: Unraveling Reward Functions for Head-to-Head Autonomous Racing in AWS DeepRacer

AWS DeepRacer is a fully autonomous 1/18th scale race car designed to help developers learn and practice reinforcement learning through cloud-based simulations and real-world racing. What drives the reinforcement learning model is the reward function, a way to provide positive or negative feedback to an agent, guiding its learning process in reinforcement learning by assigning numerical values. In the AWS training environment, there are multiple modes in which you can run training simulations and evaluations in. These modes include Time Trial, Object Avoidance, and a relatively new mode: Head to Bot. The research done in this project was primarily focused on testing reward functions in the Head to Bot mode as well as developing a research function that would be suited for training in this new Head to Bot mode. The developed algorithm outperformed the default Centerline reward function, as well as the Object Avoidance reward function.