Development of F1 Tenth-specific Autonomous Navigation Algorithm

Abstract

This research proposes an autonomous navigation algorithm for the ROS-based F1 Tenth racing platform, designed to excel in dynamic and challenging racing environments. It utilizes LiDAR data downsampling and a custom Gym environment to enable reinforcement learning-based control. The novelty in this research is the development of an efficient LiDAR data downsampling algorithm. Unlike traditional methods, this novel approach ensures computational efficiency, retaining essential information and preserving the graph’s geometric shape. Comparative computation analysis with normal sampling and polyline approximation demonstrates that the proposed approach is 217 times faster than polyline approximation and is efficient in preserving the valley and peak points despite a slightly higher RMSE. This algorithm is effective as downscaled data also needs to be of a fixed size for RL model training. The algorithm easily integrates with the F1/10 simulator through ROS2, enabling rapid algorithm development and testing in a simulated environment. This integration speeds up development and enables thorough testing before real-world deployment. After training for 700 thousand timesteps, the best trained model was able to successfully navigate 80% of the track length. The project contributes to advancing autonomous technology and its application in dynamic, fast-paced environments, where quick decision-making and precise navigation are crucial.