Reinforcement Learning Approaches for Racing and Object Avoidance on AWS DeepRacer

Developing autonomous driving models through reinforcement learning is gaining widespread prominence. However, a pervasive problem is developing obstacle avoidance systems. Specifically, optimizing path completion times while avoiding objects is an underdeveloped area of research. AWS DeepRacer’s platform provides a powerful architecture for engineering and analyzing autonomous models. Using AWS DeepRacer, we integrate two pathfinding algorithms, A* and Line-of-Sight (LoS), into this paradigm of autonomous driving. LoS is a novel algorithm that incrementally updates the model’s heading angles to amply reach its destination. We trained three types of models: Centerline, A*, and LoS. The Centerline model utilizes logic from AWS and is practically the only model used by the AWS DeepRacer community that avoids objects. We developed models from A* and LoS that outperformed the default models in time per lap while maintaining commensurate stability.