Reinforcement Learning-Based Autonomous Driving at Intersections in CARLA Simulator.

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Science

Background:

• Autonomous driving systems face significant challenges in complex intersection scenarios.
• Uncertainty in surrounding vehicle behavior complicates navigation and decision-making.
• Existing frameworks require robust methods for safe and efficient intersection traversal.

Purpose of the Study:

• To develop a Deep Reinforcement Learning (DRL) approach for autonomous intersection handling.
• To integrate Curriculum Learning (CL) to enhance the DRL training process.
• To propose a hybrid architecture for a complete autonomous driving system, focusing on high-level decision-making.

Main Methods:

• A DRL approach utilizing an actor-critic model with Proximal Policy Optimization (PPO).
• Curriculum Learning integrated to progressively increase environmental complexity during training.
• A hybrid architecture comprising operative, strategy, and tactical levels for decision-making.
• State space defined by adversary and ego vehicle information, processed by a feature extractor module.

Main Results:

• The PPO algorithm successfully inferred ego vehicle desired behavior across various intersection types (traffic lights, signs, uncontrolled).
• The DRL approach demonstrated effective handling of complex intersection scenarios based on adversarial vehicle behavior.
• Curriculum Learning improved the training efficiency and robustness of the autonomous driving system.

Conclusions:

• Deep Reinforcement Learning combined with Curriculum Learning provides a viable solution for autonomous intersection navigation.
• The proposed hybrid architecture and DRL method can effectively manage complex and uncertain traffic environments.
• The system's ability to infer behavior solely from surrounding vehicles marks a significant advancement in autonomous driving decision-making.