Το work with title Autonomous drone navigation using visual gate detection and reinforcement learning by Karamailis Panteleimon is licensed under Creative Commons Attribution 4.0 International
Bibliographic Citation
Panteleimon Karamailis, "Autonomous drone navigation using visual gate detection and reinforcement learning", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2024
https://doi.org/10.26233/heallink.tuc.100586
In recent decades, drones have gained much recognition for their ability to carry out missions involving areas inaccessible to humans, requiring high costs and long times to approach. Given the ever-increasing complexity of UAV applications, the aim is to minimize their interaction with humans to reduce human error, which consequently may cause material damage and injury. Thus, a particularly desirable direction is the development of fully autonomous navigation systems. Such systems rely on the vehicle's sensors to guide the vehicle in order to reach a target position in unknown and dynamic environments, while avoiding possible collisions. Therefore, novel and innovative approaches are used to develop applications of such complex behavior beyond conventional methodologies. Among these relatively new techniques is Reinforcement Learning (RL), a branch of Machine Learning that has achieved excellent results in many different problems in recent years. This way of learning simulates the learning of living beings, as the agent interacts with its environment and, through constant trial and error, improves its behavior to achieve its goals. Thus, the topic of this diploma thesis concerns the development of a fully operational navigation system for unmanned aircrafts and sets as its ultimate goal the safe navigation through certain gates. The inspiration for this work is the annual AlphaPilot competition of heroX, which invites participants to develop a fully autonomous drone that can beat the best pilot team in a speed race. However, this thesis focuses on sailing the aircraft as safely as possible and not on the speed of flight. Thus, the first and most significant part of this work concerns the autonomous flight of the unmanned aircraft in unknown dynamic environments and the avoidance of possible obstacles using Deep Reinforcement Learning. The proposed approach uses Deep Neural Networks to approximate value functions within RL, addressing the high-dimensional problem that traditionally challenges RL. The second part of this thesis focuses on enhancing an existing optical gate recognition system, operating in real-time through the aircraft's fixed camera. The proposed integrated system, developed and tested in the Gazebo robot simulator using the Robot Operating System (ROS) framework, successfully avoids most obstacles and navigates through gates in numerous randomly-generated instances of variable difficulty, demonstrating significant potential for robust performance.