From Simulation to Reality: Reinforcement Learning for Real-Time Extreme Vehicle Control

Collision avoidance is a critical component of automotive safety systems and a key feature of autonomous vehicles. This paper focuses on the lateral control aspect of collision avoidance. The complexity of vehicle dynamics, environmental uncertainties, and the need for real-time computation present significant challenges to effective collision avoidance control. State-of-the-art approaches, such as Model Predictive Control (MPC), provide optimal solutions but are difficult to calibrate, suffer from high computational demands, and can be unstable when simplified models are used.This study investigates an alternative approach using reinforcement learning (RL). We demonstrate how an RL agent, trained in a simulated environment, can be successfully deployed and perform aggressive maneuvers in a real vehicle without additional training. The RL agent's performance is compared to state-of-the-art controllers, showing competitive results with lower computational requirements.