Performance Evaluation and Actual Track Verification of Reinforcement Learning Algorithms Using the AWS DeepRacer

This study evaluated and analyzed the performance of the PPO and SAC reinforcement learning algorithms in various driving environments using the AWS DeepRacer. Factors affecting the performance of the reinforcement learning algorithm include the complexity of the driving track, maximum speed, and reward function. The PPO algorithm can be learned with less training time than the SAC algorithm at a maximum speed of 1.5 m/s on asimple track. , the SAC algorithm drives along the centerline more accurately than the PPO algorithm. When the maximum speed is 2.5 m/s or the driving track is complex, the PPO algorithm can complete the track even if the training time increases; in contrast, the SAC algorithm cannot complete the track owing to insufficient learning time. To verify the simulation results, we implemented an actual trackand uploaded the simulation model to a DeepRacer vehicle. The experimental confirmed that stable autonomous driving was possible on an actual track, similar to the simulation results.