Reinforcement Learning for Real-Time Frenet Path Generation and Adaptation with Dynamic Trajectory Control in CARLA Simulation

Autonomous driving has obtained significant attention in recent years, fueled by advancements in computational power that enable new opportunities. This progress allows for the application of complex machine learning (ML) models to enhance autonomous driving stacks and address the limitations of traditional methods. Path planning, particularly for obstacle avoidance, is a key area where ML can offer more robust and adaptable solutions compared to the geometric approaches often used in current systems like Autoware Universe. These geometric methods can exhibit significant limitations in dynamic scenarios, creating a need for the adaptive solutions that ML can provide. A critical challenge in autonomous navigation remains the development of path-planning mechanisms capable of generating safe, efficient, and comfortable trajectories in real time. This thesis addresses these limitations by proposing a Reinforcement Learning (RL) agent capable of generating adaptive paths within the Frenet coordinate system. The core of the method is a Soft Actor-Critic (SAC) agent trained to dynamically modify both longitudinal speed and lateral displacement, informed by various sensor data inputs like processed LiDAR and camera feeds. The agent was trained in the high-fidelity CARLA simulator using a custom vehicle model and a two-stage curriculum that first mastered lane following before learning complex obstacle avoidance maneuvers. A comprehensive reward function guided the agent to balance safety, efficiency, and path quality. Results demonstrate that the agent can successfully generate dynamically feasible and contextually appropriate trajectories by intelligently coordinating speed and lateral control. The key novelty of this work lies in the agent’s ability to learn the coupled control of both longitudinal and lateral motion within the Frenet frame. This offers a more integrated and adaptable approach than traditional geometric methods and lays the groundwork for future integration into real-world autonomous driving stacks