Obstacles Avoidance of Self-driving Vehicle using Deep Reinforcement Learning

Abstract

Nowadays, there exist different self-driving vehicle functions that allow the vehicle to perform certain actions by itself while the driver is only monitoring it. However, it is difficult in real world to acquire training data for self-driving artificial intelligence algorithms because there are a lot of risks and the need of labeled data. This paper proposes a method to collect training data from Unity game engine’s Machine Learning Toolkit (ML-Agents Toolkit). With this toolkit, Unity allows its users to incorporate Reinforcement Learning (RL) algorithms to train a learning agent. The aim of this paper is to search for the best RL algorithm in order to train the self-driving vehicle to avoid obstacles in a 3D environment. For all study cases, the learning was done by using the two RL learning algorithms Proximal Policy Optimization algorithm (PPO) and Soft Actor-Critic (SAC) algorithm, both using single-instance and multi-instance training. In the data collection from virtual environment to learn, two types of sensors in comparison had been experimented using camera sensors and Light Detection and Ranging (LiDaR) sensors. The results of the research show the advantages and limitations of the used learning algorithms for learning behaviors, the importance of the demonstration provided for the learning algorithms. Experimental results for applying the virtual driving data to drive a vehicle shows the effectiveness of the proposed methodology.