End-to-end autonomous driving based on reinforcement learning

Abstract

In this project, an RGB camera will be used as data input to explore an end-to-end method based on visual based reinforcement learning. The project will be carried out with the Unity game engine as the training environment, along with Unity’s ML-Agents package that provides out of the box deep Reinforcement Learning (RL) algorithms to interface with their environment. The results of training a simulated donkey car to drive in its own lane with an on-policy method, Proximal Policy Optimization (PPO), and an off-policy method, Soft Actor-Critic (SAC) will be compared. An ablation study, consisting of adding Generative Adversarial Imitation Learning (GAIL), semantic segmentation and stacked visual inputs, will be performed. Additionally, RL based obstacle avoidance will be explored. The results, based on stability of control and ability to stay in lane, indicate that the best performing method is PPO. Code is available at: https://github.com/MrOCW/Autonomous-Driving-RL-Unity