A Case-based Reward Function Design for Reinforcement Learning-based Pure Pursuit Hybrid Controller

This paper presents an innovative approach to enhancing the Pure Pursuit algorithm for path tracking in autonomous vehicles by integrating Reinforcement Learning and curvature information. Traditional Pure Pursuit algorithms, while effective in low-speed scenarios, often require extensive manual tuning of the look-ahead distance to maintain tracking accuracy at varying speeds and complex paths. To address these limitations, we designed an RL-based pure pursuit controller incorporating future curvature into the state space and reward function to enhance learning a proper tracking policy at higher speeds. The controller is trained and evaluated in the CARLA simulator, demonstrating improved performance in terms of path-tracking accuracy and stability across different speeds and path complexities. By comparing the controller which considered curvature improvement with the original one, our results show that the improved method can achieve lower lateral deviation and lateral acceleration while maintaining almost the same average speed.

This paper presents an innovative approach to enhancing the Pure Pursuit algorithm for path tracking in autonomous vehicles by integrating Reinforcement Learning and curvature information. Traditional Pure Pursuit algorithms, while effective in low-speed scenarios, often require extensive manual tuning of the look-ahead distance to maintain tracking accuracy at varying speeds and complex paths. To address these limitations, we designed an RL-based pure pursuit controller incorporating future curvature into the state space and reward function to enhance learning a proper tracking policy at higher speeds. The controller is trained and evaluated in the CARLA simulator, demonstrating improved performance in terms of path-tracking accuracy and stability across different speeds and path complexities. By comparing the controller which considered curvature improvement with the original one, our results show that the improved method can achieve lower lateral deviation and lateral acceleration while maintaining almost the same average speed.