This paper proposes a robust Deep Reinforcement Learning (DRL) framework for autonomous driving that addresses the Sim2Real transfer challenge. The framework incorporates platform-dependent perception modules to extract task-relevant information, enabling training of a lane-following and overtaking agent in simulation. This agent can then be efficiently transferred to new simulated environments and the real world with minimal adjustments. The agent's performance is evaluated in various driving scenarios, comparing it to human drivers and PID baselines. The proposed approach bridges the Sim2Real gap, enabling consistent performance in both simulation and real-world scenarios.

Dianzhao Li, Ostap Okhrin