Boosting Race Car Performance Through Reinforcement Learning from AI Feedback (RLAIF)

In this paper, we present a novel application of reinforcement learning with artificial intelligence frameworks (RLAIF) to enhance race car motion planning. Leveraging Proximal Policy Optimization (PPO) as the foundational algorithm, we develop two distinct artificial intelligence systems built on Convolutional Neural Networks (CNNs). These systems are designed with specific objectives: the first aims to maintain the race car's position in the center of the road, while the second encourages the adoption of shorter, more efficient paths. To integrate these objectives into a coherent learning process, we devise a reward modeling strategy that translates the outcomes of the CNN-based classifications into meaningful reward signals. Experimental results demonstrate the effectiveness of our approach, showcasing significant improvements in race car motion planning.