A platform-agnostic deep reinforcement learning framework for effective Sim2Real transfer towards autonomous driving.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Autonomous driving systems face challenges in transferring agents from simulation to real-world environments due to simulation-reality discrepancies.
• Bridging the Simulation to Reality (Sim2Real) gap is crucial for reliable autonomous vehicle deployment.

Purpose of the Study:

• To propose a robust Deep Reinforcement Learning (DRL) framework for autonomous driving that facilitates Sim2Real transfer.
• To enable the training of a lane-following and overtaking agent in simulation with minimal adjustments for real-world application.

Main Methods:

• Developed a DRL framework incorporating platform-dependent perception modules for task-relevant information extraction.
• Trained a lane-following and overtaking agent within a simulated environment.
• Assessed agent performance in diverse simulated and real-world driving scenarios.

Main Results:

• The proposed DRL framework demonstrated efficient transfer of the trained agent to new simulated environments and the real world.
• The agent performed consistently across various driving scenarios, bridging the Sim2Real gap.
• Comparative analysis showed the framework's effectiveness against human drivers and baseline methods.

Conclusions:

• The DRL framework successfully addresses Sim2Real challenges in autonomous driving.
• The approach ensures consistent agent performance in both simulated and real-world settings.
• This methodology enhances the reliability and applicability of DRL agents for autonomous vehicles.