Tesi etd-12292022-153456
Link copiato negli appunti
Tipo di tesi
Dottorato
Autore
HERRERA ALARCON, EDWIN PAUL
URN
etd-12292022-153456
Titolo
Autonomous vision-based drone navigation in GNSS-denied environments
Settore scientifico disciplinare
ING-INF/04
Corso di studi
Istituto di Tecnologie della Comunicazione, dell'Informazione e della Percezione - PHD IN EMERGING DIGITAL TECHNOLOGIES
Commissione
relatore Prof. AVIZZANO, CARLO ALBERTO
Membro Prof.ssa DONZELLA, VALENTINA
Membro Prof. LOIANNO, GIUSEPPE
Membro Prof.ssa DONZELLA, VALENTINA
Membro Prof. LOIANNO, GIUSEPPE
Parole chiave
- Aerial Robotics
- Autonomous Navigation
- Motion and Path Planning
- Vision-based Perception
Data inizio appello
31/07/2023;
Disponibilità
parziale
Riassunto analitico
Aerial robotics has gained significant attention in various industries, including academic, commercial, industrial, and humanitarian sectors. Flying robots are going to have an important role in the future by accomplishing tasks such as delivering
goods, surveilling cities to guarantee safety, revising inventory in warehouses, inspecting structures that are unsafe for or inaccessible by human operators, reacting to natural disasters promptly by performing search-and-rescue missions among the rubble, between others. Still, it is worth noting that until recently, for most of the applications of these vehicles, they have been primarily operated in two modes: manually under the guidance of a pilot or using global positioning systems.
Autonomous navigation is a field in robotics that researches how to determine
a vehicle’s location within an environment and to be able to figure out a path that will take it to a goal. The navigation stack of an autonomous system is composed of modules that guarantee its capacity for guidance and navigation in different scenarios; these modules can be described into four major components: localization, mapping, path-planning, and control. Most of the existing literature in this field assumes that the vehicle can access global positioning information or process the data retrieved by the onboard sensors in offboard computational units with high processing performance. This assumption does not apply to complex scenarios where communication with external parties is not granted, and the robot’s behavior needs to rely only on its onboard resources. However, easier to be said than done, the deployment of autonomous agents in real-world scenarios represents a series of significant challenges that can be summarized through their SWaP (Size, Weight, and Power) constraints.
Multirotors are aerial robot platforms that stand out for their agility, maneuverability, simple electromechanical design, payload capacity, and vertical take-off capabilities. Moreover, a plausible option to guarantee autonomy with a limited payload platform is to rely on lightweight sensors, such as cameras and inertial measurement units (IMU). Cameras are promising sensors for limited-size drones because, for comparable mass, they can gather rich information and span wide fields of view, allowing them to understand the environment better, which can guarantee a certain performance during an autonomous mission.
Motivated by the difficulties associated with autonomous navigation for aerial platforms relying only on onboard resources in degraded scenarios, this thesis addresses the path-planning problem of autonomous vision-based multirotors. Considering aerial platforms often need to trade off their payload capacity and performance with their flight endurance, the research question addressed in this work is to decide where to go next efficiently in vehicles with critical SWaP constraints.
The path planning problem is tackled in two main categories: unknown and known GPS-denied environments. The solutions explored present a hybrid approach exploiting traditional planning algorithms with learning-based methods. Specifically, the proposed contributions use different machine learning techniques in collaboration with classic planning algorithms to enhance the path planning module performance for solving a specific problem.
Unknown environments are confronted through the exploration problem, which focuses on maximizing the volumetric representation of the surroundings in a completely unfamiliar scenario. The first contribution depicts an object-oriented exploration algorithm whose object detection module relies on Convolution Neural Networks. The exploration behavior is settled through a cost-utility formulation function that comprises both an exploration gain and an object gain. The key improvement of the method is the trade-off between both gains by exploiting the different working ranges of the sensors used to favor the exploration strategy. On the one side, the point cloud is provided by the depth sensor camera, and on the other side, the field of view is recovered from the RGB camera.
Exploration is addressed on a second occasion through the implementation of graph neural networks as a policy that defines the behavior of the robot. Classic approaches use the same exploration criterion during a mission to decide where to go next; conversely, learning-based approaches can learn the exploration behavior through heuristics. The studied solution is trained without a supervised methodology, bypassing the advantages and disadvantages of the teaching algorithm. This approach exploits the graph-alike structure of the exploration tree as a natural input embedding for the graph neural network, which is often already embedded with exploration information.
Finally, the path planning problem in known environments is visited by retrieving a low-dimensional roadmap of the environment, used for fast and safe planning. The extracted graph represents the medial axis of the collision-free space of the environment and is extracted using an unsupervised learning method for clusterization. Furthermore, the resultant graph is a close representation of the 3D Voronoi diagram of the scene, guaranteeing a certain distance from obstacles and, consequently, safety planning. A fundamental aspect of the mentioned algorithm is using a stop-learning criterion based on the Laplacian of the graph that guarantees its sparse connectivity. The connectivity attribute is exploited for the path planning problem because if the initial and final configurations can reach the graph, then the solution is assured. The performance of the previously mentioned contributions is compared to state-of-the-art solutions solving the same problem under identical conditions for the testbed.
goods, surveilling cities to guarantee safety, revising inventory in warehouses, inspecting structures that are unsafe for or inaccessible by human operators, reacting to natural disasters promptly by performing search-and-rescue missions among the rubble, between others. Still, it is worth noting that until recently, for most of the applications of these vehicles, they have been primarily operated in two modes: manually under the guidance of a pilot or using global positioning systems.
Autonomous navigation is a field in robotics that researches how to determine
a vehicle’s location within an environment and to be able to figure out a path that will take it to a goal. The navigation stack of an autonomous system is composed of modules that guarantee its capacity for guidance and navigation in different scenarios; these modules can be described into four major components: localization, mapping, path-planning, and control. Most of the existing literature in this field assumes that the vehicle can access global positioning information or process the data retrieved by the onboard sensors in offboard computational units with high processing performance. This assumption does not apply to complex scenarios where communication with external parties is not granted, and the robot’s behavior needs to rely only on its onboard resources. However, easier to be said than done, the deployment of autonomous agents in real-world scenarios represents a series of significant challenges that can be summarized through their SWaP (Size, Weight, and Power) constraints.
Multirotors are aerial robot platforms that stand out for their agility, maneuverability, simple electromechanical design, payload capacity, and vertical take-off capabilities. Moreover, a plausible option to guarantee autonomy with a limited payload platform is to rely on lightweight sensors, such as cameras and inertial measurement units (IMU). Cameras are promising sensors for limited-size drones because, for comparable mass, they can gather rich information and span wide fields of view, allowing them to understand the environment better, which can guarantee a certain performance during an autonomous mission.
Motivated by the difficulties associated with autonomous navigation for aerial platforms relying only on onboard resources in degraded scenarios, this thesis addresses the path-planning problem of autonomous vision-based multirotors. Considering aerial platforms often need to trade off their payload capacity and performance with their flight endurance, the research question addressed in this work is to decide where to go next efficiently in vehicles with critical SWaP constraints.
The path planning problem is tackled in two main categories: unknown and known GPS-denied environments. The solutions explored present a hybrid approach exploiting traditional planning algorithms with learning-based methods. Specifically, the proposed contributions use different machine learning techniques in collaboration with classic planning algorithms to enhance the path planning module performance for solving a specific problem.
Unknown environments are confronted through the exploration problem, which focuses on maximizing the volumetric representation of the surroundings in a completely unfamiliar scenario. The first contribution depicts an object-oriented exploration algorithm whose object detection module relies on Convolution Neural Networks. The exploration behavior is settled through a cost-utility formulation function that comprises both an exploration gain and an object gain. The key improvement of the method is the trade-off between both gains by exploiting the different working ranges of the sensors used to favor the exploration strategy. On the one side, the point cloud is provided by the depth sensor camera, and on the other side, the field of view is recovered from the RGB camera.
Exploration is addressed on a second occasion through the implementation of graph neural networks as a policy that defines the behavior of the robot. Classic approaches use the same exploration criterion during a mission to decide where to go next; conversely, learning-based approaches can learn the exploration behavior through heuristics. The studied solution is trained without a supervised methodology, bypassing the advantages and disadvantages of the teaching algorithm. This approach exploits the graph-alike structure of the exploration tree as a natural input embedding for the graph neural network, which is often already embedded with exploration information.
Finally, the path planning problem in known environments is visited by retrieving a low-dimensional roadmap of the environment, used for fast and safe planning. The extracted graph represents the medial axis of the collision-free space of the environment and is extracted using an unsupervised learning method for clusterization. Furthermore, the resultant graph is a close representation of the 3D Voronoi diagram of the scene, guaranteeing a certain distance from obstacles and, consequently, safety planning. A fundamental aspect of the mentioned algorithm is using a stop-learning criterion based on the Laplacian of the graph that guarantees its sparse connectivity. The connectivity attribute is exploited for the path planning problem because if the initial and final configurations can reach the graph, then the solution is assured. The performance of the previously mentioned contributions is compared to state-of-the-art solutions solving the same problem under identical conditions for the testbed.
File
| Nome file | Dimensione |
|---|---|
Ci sono 1 file riservati su richiesta dell'autore. |
|