Integrating cyber-physical systems with embedding technology for controlling autonomous vehicle driving.

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Engineering

Background:

• Autonomous vehicle cyber-physical systems (CPSs) face challenges in dynamic environments.
• Deep reinforcement learning (DRL) struggles with safety and optimal behavior in uncertain settings.
• DRL's difficulty in understanding dynamic reward systems limits its application in autonomous driving.

Purpose of the Study:

• To improve the adaptability and decision-making of autonomous driving CPSs under uncertain conditions.
• To address the limitations of current DRL models in ensuring safety and optimal behavior.
• To investigate the integration of algorithmic enhancements with hardware acceleration for real-time reinforcement learning.

Main Methods:

• Incorporation of double deep Q networks (DDQN) for enhanced agent adaptability.
• Development of a structured reward system for real-time fluctuation accommodation.
• Implementation of field programmable gate arrays (FPGAs) for real-time reinforcement learning execution and hardware acceleration.

Main Results:

• The combination of FPGA hardware acceleration and DDQN significantly improves computational efficiency.
• Decision-making reliability in uncertain autonomous driving scenarios is enhanced.
• Key performance metrics including collision rate, behavior similarity, travel distance, and speed control show positive outcomes.

Conclusions:

• FPGA-based hardware acceleration with DDQN effectively tackles uncertainty in autonomous driving CPSs.
• The study advances reinforcement learning applications in CPSs, enhancing safety and efficiency.
• Future research can explore real-world generalization, adaptive rewards, and scalable hardware for autonomous systems.