End-to-end autonomous driving based on reinforcement learning

• Main Author: Beh, Chun Jye
• Other Authors: Lyu Chen
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Mechanical engineering::Mechatronics; Engineering::Mechanical engineering::Robots
• Online Access: https://hdl.handle.net/10356/141863
• Institution: Nanyang Technological University

Autonomous vehicles (AVs) have been developed for many years. Perception, decision making, path planning and tracking control are the four key components in AV. To highlight, lane keeping assistant is one of the most important scenarios in AV as it provides automatic control to the steering and braking to ensure the vehicle stays in the lanes. Lane keeping assistant can be achieved by different techniques such as PID controller and supervised learning method. In this paper, we focus on deep reinforcement learning-based (DRL) method for lane keeping assist system. Then, we move on to examine the simulatorreality gap and feasibility of DRL in real world. We train deep reinforcement learning models that images are taken by RGB camera in its first-person view to serve as the only input while the throttle and steering angle will be the output of the model. Reinforcement learning is an extension of deep learning which an autonomous agent must learn to perform with sequential decision-making task without the complete knowledge or control of the environment. It collects information and experience by interacting with the environment. In this paper, we present a comparative analysis between various learning algorithms such as DDQN, PPO, DDPG and SAC to perform lane keeping assist task. We set episode reward and learning efficiency as the criteria to determine the efficiency of algorithms. Carefully designing a good reward function can mean the difference between an effective and a misbehaving agent. Different reward function will be tested to establish the best design in our experiment. To evaluate the performance, different reinforcement learning algorithms will be trained using the same track in a Unity simulator. The performance will first be tested from one simulator environment to another one. The best model will then be applied on real-world track to investigate the consistency and simulator-reality gap