Enhancing Autonomous Driving in Urban Scenarios: A Hybrid Approach with Reinforcement Learning and Classical Control.

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Science

Background:

• Autonomous vehicle decision-making systems often struggle with real-world complexities.
• Deep Reinforcement Learning (DRL) shows promise but lacks reliability for critical applications.
• Classical methodologies offer reliability but lack adaptability.

Purpose of the Study:

• To develop and validate a hybrid decision-making module for autonomous driving.
• To integrate DRL's learning capabilities with the robustness of traditional control methods.
• To generate reliable steering and velocity signals using HD maps and sensor data.

Main Methods:

• Formulated the decision-making problem as a Partially Observable Markov Decision Process (POMDP).
• Implemented a hybrid architecture combining DRL algorithms with a classical control module.
• Validated the system in the CARLA simulator using concatenated driving scenarios.

Main Results:

• The proposed hybrid system successfully generated steering and velocity commands.
• The architecture demonstrated enhanced safety and comfort compared to existing methods.
• Outperformed the CARLA Autopilot in scenario completion time.

Conclusions:

• A hybrid decision-making system effectively merges DRL and classical approaches for autonomous driving.
• The validated system offers a realistic and efficient solution for autonomous vehicle navigation.
• This approach improves upon existing autonomous driving stacks in simulated environments.