Institutional Repository [SANDBOX]
Technical University of Crete
EN  |  EL

Search

Browse

My Space

Autonomous drone navigation for landmark position estimation using reinforcement learning

Galanis Michalis

Full record


URI: http://purl.tuc.gr/dl/dias/8E9F870F-EF8B-42A0-AFC6-488BF38B90DF
Year 2021
Type of Item Diploma Work
License
Details
Bibliographic Citation Michalis Galanis, "Autonomous drone navigation for landmark position estimation using reinforcement learning", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2021 https://doi.org/10.26233/heallink.tuc.90434
Appears in Collections

Summary

Unmanned aerial vehicles (UAVs) have been increasingly used for critical and challenging applications, which often require a substantial level of autonomy. Several approaches have been investigated to create autonomous navigation systems such as Simultaneous Localization and Mapping (SLAM) using real-time mapping and position estimation. Reinforcement leaning (RL) is a promising alternative that focuses on learning to perform a task through a trial-and-error procedure, in which an agent interacts with its environment and receives continuous feedback based on the actions taken, with no access to any information about the environment itself. Eventually, the agent’s objective is to find the best possible sequence of actions that lead to the maximum total reward in the long term. This thesis explores a mapless approach to UAV autonomous navigation in completely unknown 3D environments using deep reinforcement learning (DRL), a reinforcement learning approach that incorporates deep learning techniques (deep neural networks) to overcome dimensionality limitations. The goal of the agent is to safely navigate through this unknown environment, so as to detect and approach a predefined set of ArUco markers (landmarks) placed within the environment. The unknown environments are dynamically created and contain a number of procedurally generated obstacles. We evaluate our agent in five different environment profiles with increasing difficulty level and observe how environment complexity affects training performance. Results show that deep reinforcement learning can be effective and can be successfully used for autonomous navigation missions. The entire project was implemented using the Robot Operating System (ROS) platform within the Gazebo robot simulator environment.

Available Files

Services

Statistics