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

Autonomous driving faces significant challenges in transferring agents trained in simulation to real-world environments due to Sim2Real discrepancies. This work proposes a robust Deep Reinforcement Learning (DRL) framework that uses platform-dependent perception modules to extract task-relevant affordances and a universal DRL control module. The agent is trained in simulation for lane following and overtaking, then transferred across different simulators and to real-world robotic vehicles with minimal adjustments. We evaluate performance across diverse simulated and real-world scenarios, comparing against PID control, human drivers, and other DRL baselines. Results show consistent, effective driving and superior performance to baselines, demonstrating the framework’s ability to bridge platform gaps and the Sim2Real gap.