Autonomous exploration and object reconstruction with an MAV
Author(s)
Papatheodorou, Sotirios
Type
Thesis
Abstract
Deploying autonomous robots outside a controlled environment is a challenging task. Robots must be able to gather information about and model their environment, determine their next course of action, and safely navigate to their goal. As such, autonomous exploration is an integral component of many higher-level tasks tackled by autonomous robots. Micro Aerial Vehicles (MAVs) are a very popular robotic platform for autonomous exploration due to their capability to move freely in three dimensions, allowing them to navigate a variety of environments.
While there is prior work in autonomous MAV exploration, it is by no means a solved problem. Improvements can be made in terms of exploration speed, scalability to large environments, and robustness to odometry drift. Maybe most importantly, the integration of autonomous exploration with some higher-level robotic task has not been extensively researched.
In this thesis we propose autonomous exploration methods for MAVs that address these challenges. A fast and efficient exploration algorithm, a hybrid between sampling-based and frontier-based exploration is presented. Additionally, the traditional geometric formulation of autonomous exploration is extended to incorporate a downstream task, in this case the high-quality reconstruction of objects of interest. Finally, a submap-based large-scale exploration method suitable for both depth camera and LiDAR sensors is introduced. These contributions have the potential to enable the use of autonomous MAV exploration in a wider variety of environments and tasks.
While there is prior work in autonomous MAV exploration, it is by no means a solved problem. Improvements can be made in terms of exploration speed, scalability to large environments, and robustness to odometry drift. Maybe most importantly, the integration of autonomous exploration with some higher-level robotic task has not been extensively researched.
In this thesis we propose autonomous exploration methods for MAVs that address these challenges. A fast and efficient exploration algorithm, a hybrid between sampling-based and frontier-based exploration is presented. Additionally, the traditional geometric formulation of autonomous exploration is extended to incorporate a downstream task, in this case the high-quality reconstruction of objects of interest. Finally, a submap-based large-scale exploration method suitable for both depth camera and LiDAR sensors is introduced. These contributions have the potential to enable the use of autonomous MAV exploration in a wider variety of environments and tasks.
Version
Open Access
Date Issued
2024-11-14
Date Awarded
01/06/2025
License URL
Advisor
Leutenegger, Stefan
Davison, Andrew
Sponsor
Imperial College London
Munich Institute of Robotics and Machine Intelligence (MIRMI)
Munich Center for Machine Learning (MCML)
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)