Obstacle avoidance in mobile robot with multiple sensors using deep reinforcement learning

• Κύριος συγγραφέας: Καραγεώργος-Γούδης, Κωνσταντίνος
• Άλλοι συγγραφείς: Karageorgos-Goudis, Konstantinos
• Γλώσσα: English
• Έκδοση: 2021
• Θέματα: Reinforcement learning; Obstacle avoidance; Neural networks; Mobile robots; Ενισχυτική μάθηση; Αποφυγή εμποδίων; Νευρωνικά δίκτυα; Αυτοκινούμενα ρομπότ

In this project, we first studied the fundamental Reinforcement Learning algorithms and implemented them to play games like CartPole and VizDoom. We then proceeded to solve the problem of obstacle avoidance in an un- known environment using the previously developed algorithms. The field of reinforcement learning was inspired from biological systems and enables an agent to acquire knowledge through a trial and error process, interacting with an environment, based on the feedback returned from this environment. A multi- plicity of algorithms have been developed over the last years using this approach, with spectacular results in a plethora of fields, outperforming previous human attempts. In the current thesis we aim to implement the most fundamental of these RL methods, namely DQN, Double DQN, REINFORCE, Actor Critic and Proximal Policy Optimization. Then we create a custom Reinforcement Learning Environment consisted of a mobile robot with a camera, low range distance sensors and a collision detection sensor, moving in an unknown area with obstacles in random positions. We, finally, train an RL agent to explore the environment, avoiding any obstacles on the way.