Trajectory based lateral control: A Reinforcement Learning case study

Reinforcement Learning (RL) has been employed in many applications of robotics and has steadily been gaining traction in the field of Autonomous Driving (AD). This paper proposes a Deep Reinforcement Learning based approach for lateral Vehicle Motion Control (VMC), and explores the generalization capabilities of the approach. The proposed methodology uses a sequence of waypoints generated from a planning module of an AD stack as the input. The network has been trained to predict accurate steering commands to follow the given trajectory. In this paper we detail our implementation and share our learning experience on real-vehicle deployment of the RL based controller. Our experiments yield promising results with an agent trained on less than 4 h of simulated driving experience without any real-world data. The trained agent is able to successfully complete unseen and more complex tracks using different unseen vehicle models. The agent safely reached up to 150km/h in simulation and up to 60km/h in a real-life Sport Utility Vehicle (SUV) weighing more than 2000kg.