Tuning path tracking controllers for autonomous cars using reinforcement learning.

Area of Science:

• Robotics and Control Systems
• Artificial Intelligence
• Autonomous Driving

Background:

• Autonomous vehicles require robust path tracking for safe navigation.
• Traditional control systems often struggle with dynamic trajectory changes and complex maneuvers like roundabouts.

Purpose of the Study:

• To develop an adaptable path tracking control system for autonomous cars.
• To minimize lateral and steering trajectory errors using a novel reinforcement learning approach.
• To ensure safe and accurate navigation through various trajectories, including lane changes and roundabouts.

Main Methods:

• Implementation of a four-parameter controller for vehicle behavior shaping.
• Utilizing an 'educated' Q-Learning algorithm for controller tuning.
• Employing the CARLA (CAR Learning to Act) simulator for system training and testing.
• Integrating a robot operating system (ROS) bridge for realistic simulation and hardware-in-the-loop compatibility.

Main Results:

• The autonomous vehicle demonstrated adaptability to diverse reference trajectories.
• The system achieved safe navigation with significantly reduced tracking errors.
• The ROS bridge facilitated a realistic simulation environment and simplified real-world vehicle integration.

Conclusions:

• The proposed reinforcement learning-based path tracking system offers an adaptable and effective solution for autonomous driving.
• The 'educated' Q-Learning approach successfully minimizes trajectory errors, enhancing driving safety.
• The system architecture, supported by dependability analysis, is suitable for both simulation and real-world deployment.