A LiDAR-based approach to autonomous racing with model-free reinforcement learning

This paper explores the use of reinforcement learning (RL) in the context of autonomous vehicle racing, specifically focusing on the F1TENTH simulation platform. While commercial autonomous driving often employs classic control algorithms, the state-of-the-art solutions, including those in the F1TENTH domain, increasingly rely on RL. Notably, RL-based approaches have shown superhuman performance in simulated environments, as seen in drone racing and the recent achievement by Sony in autonomous racing. In this paper we propose a novel LiDAR-only observation for learning vehicle dynamics, and test it with a widely accessible model-free RL method. The trained agent demonstrates the capability to transfer its driving skills to previously unseen tracks. Additionally, the paper provides recommendations for selecting hyperparameters, contributing valuable insights for newcomers to the field of autonomous racing.