Go-Kart Racing Simulator for Reinforcement Learning with Augmented Sim2Real Adaptation

Training self-driving cars in real-world scenarios is inefficient due to the possibility of crashes with obstacles and borders. This paper introduces a virtual environment to enhance reinforcement learning training in a virtual Go-Kart racing simulator. The primary objective is to leverage augmented reality to enhance observations inside the simulation, improve policy networks, and make the Value function precise and robust. We develop the wrapper for the CARLA simulator, enabling a cost-effective Sim2Real transition. It is demonstrated that the augmented Sim2Real adaptation successfully integrates simulated training outcomes into real-world scenarios where the real Go-Kart can accomplish six laps in a single-race mode, reaching the maximum speed of 11.5 m/s.

I. Introduction

The main components of an Autonomous Driving System [1], [2] include perception with multiple sensors, localization and mapping, planning, and control. The planning module utilizes the perception state to generate the motion-level commands based on route-level plans. Control for trajectory tracking defines speed, steering angle, and braking actions. The path planning and trajectory following tasks are typically solved by Reinforcement learning (RL) techniques [3]. In RL, an autonomous agent learns to improve its performance by interacting with its environment and receiving rewards based on its actions [4] - [6]. The agent aims to maximize its cumulative rewards over time by reaching the balance between exploration and exploitation [7].