東華大學圖書館 |

Towards Deep-Reinforcement-Learning-Based Navigation Systems for Open-Path Navigation in Collaborative and Highly Dynamic Environments /

Record Type:	Electronic resources : Monograph/item
Title/Author:	Towards Deep-Reinforcement-Learning-Based Navigation Systems for Open-Path Navigation in Collaborative and Highly Dynamic Environments // Doan Hoang Linh Kastner.
Author:	Kastner, Doan Hoang Linh,
Description:	1 electronic resource (184 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 85-11, Section: A.
Contained By:	Dissertations Abstracts International85-11A.
Subject:	Deep learning. -
Online resource:	https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31078083
ISBN:	9798382654836

Towards Deep-Reinforcement-Learning-Based Navigation Systems for Open-Path Navigation in Collaborative and Highly Dynamic Environments /
Kastner, Doan Hoang Linh,

Towards Deep-Reinforcement-Learning-Based Navigation Systems for Open-Path Navigation in Collaborative and Highly Dynamic Environments /Doan Hoang Linh Kastner. - 1 electronic resource (184 pages)

Source: Dissertations Abstracts International, Volume: 85-11, Section: A.

Navigation of mobile robots in collaborative environments, employing various different actors like humans, robots, or other machinery, opens doors for a variety of complex task domains such as last-mile delivery, logistics, or healthcare services. However, technical limitations still prevent widespread operation of mobile robots in these highly dynamic environments. Moreover, currently employed approaches are cost-intensive due to their overly conservative navigation behavior and the need for extensive engineering efforts. At the same time, Deep Reinforcement Learning has been increasingly successful to generalize new problem instances and swiftly adapt to unexpected changes, thus revealing the potential to improve dynamic obstacle avoidance substantially. This work addresses three main points that currently hampers the widespread adoption of Deep Reinforcement Learning into mobile robot navigation to improve dynamic obstacle avoidance: the complex training of Deep Reinforcement Learning agents, the lack of integration efforts into conventional systems, and as a result, the difficult comparability with classic approaches. To address these aspects: first, an efficient training environment is provided to develop Deep Reinforcement Learning agents in a resource-efficient and intuitive manner. Second, methods to integrate the resulting Deep Reinforcement Learning agent into existing navigation stacks are presented by proposing an intermediate planner that interconnects and leverages the advantages of both learning-based and classic navigation. Finally, methods to extensively evaluate and compare the resulting navigation systems against state-of-the-art approaches are developed to address the lack of an unified assessment procedure for Deep Reinforcement Learning approaches. Results demonstrated the superiority of the proposed navigation system in terms of robustness, efficiency, safety, and smoothness. Within evaluations on real robots, the proposed system outperforms multiple baselines by up to 48 % in terms of navigational safety and by up to 14 % regarding the overall performance - producing the least collisions, highest success rates, and shortest paths to the goal, particularly in highly dynamic environments. This constitutes an essential step to facilitate natural navigation in collaborative environments, thus opening doors for more sophisticated and complex task domains in future.

English

ISBN: 9798382654836Subjects--Topical Terms:

3554982
Deep learning.

Towards Deep-Reinforcement-Learning-Based Navigation Systems for Open-Path Navigation in Collaborative and Highly Dynamic Environments /
LDR:06549nmm a22004573i 4500 001 2437667
005 20251124054954.5
006 m o d
007 cr|nu||||||||
008 261011s2023 miu||||||m |||||||eng d
020 $a 9798382654836
035 $a (MiAaPQD)AAI31078083
035 $a (MiAaPQD)TechBerlin_20231130318926
035 $a AAI31078083
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Kastner, Doan Hoang Linh, $e author. $3 3828733
245 1 0 $a Towards Deep-Reinforcement-Learning-Based Navigation Systems for Open-Path Navigation in Collaborative and Highly Dynamic Environments / $c Doan Hoang Linh Kastner.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2023
300 $a 1 electronic resource (184 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 85-11, Section: A.
500 $a Advisors: Lambrecht, Jens; Hafner, Verena; Xiao, Xuesu.
502 $b Ph.D. $c Technische Universitaet Berlin (Germany) $d 2023.
520 $a Navigation of mobile robots in collaborative environments, employing various different actors like humans, robots, or other machinery, opens doors for a variety of complex task domains such as last-mile delivery, logistics, or healthcare services. However, technical limitations still prevent widespread operation of mobile robots in these highly dynamic environments. Moreover, currently employed approaches are cost-intensive due to their overly conservative navigation behavior and the need for extensive engineering efforts. At the same time, Deep Reinforcement Learning has been increasingly successful to generalize new problem instances and swiftly adapt to unexpected changes, thus revealing the potential to improve dynamic obstacle avoidance substantially. This work addresses three main points that currently hampers the widespread adoption of Deep Reinforcement Learning into mobile robot navigation to improve dynamic obstacle avoidance: the complex training of Deep Reinforcement Learning agents, the lack of integration efforts into conventional systems, and as a result, the difficult comparability with classic approaches. To address these aspects: first, an efficient training environment is provided to develop Deep Reinforcement Learning agents in a resource-efficient and intuitive manner. Second, methods to integrate the resulting Deep Reinforcement Learning agent into existing navigation stacks are presented by proposing an intermediate planner that interconnects and leverages the advantages of both learning-based and classic navigation. Finally, methods to extensively evaluate and compare the resulting navigation systems against state-of-the-art approaches are developed to address the lack of an unified assessment procedure for Deep Reinforcement Learning approaches. Results demonstrated the superiority of the proposed navigation system in terms of robustness, efficiency, safety, and smoothness. Within evaluations on real robots, the proposed system outperforms multiple baselines by up to 48 % in terms of navigational safety and by up to 14 % regarding the overall performance - producing the least collisions, highest success rates, and shortest paths to the goal, particularly in highly dynamic environments. This constitutes an essential step to facilitate natural navigation in collaborative environments, thus opening doors for more sophisticated and complex task domains in future.
520 $a Navigation in kollaborativen Umgebungen mit verschiedenen Akteuren wie Menschen, Robotern, und anderen Maschinen, offnet Turen fur eine Vielzahl komplexer Aufgabenbereiche, z.B. in Bereichen der Last-mile Lieferung, der Logistik oder der Pflege. Die Navigation in diesen hochdynamischen Umgebungen stellt jedoch weiterhin ein offenes Problem dar. Aktuell verwendete Ansatze haben immernoch technische Limitationen, agieren oft zu konservativ oder meiden diese Bereiche ganzlich und sind dadurch kostenintensiv. Gleichzeitig ist Deep Reinforcement Learning zunehmend erfolgreich in der Lage, neue Problemfalle zu verallgemeinern und sich schnell an unerwartete Veranderungen anzupassen und hat somit das Potential die dynamische Hindernisvermeidung erheblich zu verbessern. Diese Arbeit adressiert drei Hauptpunkte, welche derzeit die weit verbreitete Anwendung von Deep Reinforcement Learning in der mobilen Roboternavigation zur Verbesserung der dynamischen Hindernisvermeidung behindern: komplexes Training von Deep Reinforcement Learning Agenten, mangelnde Integrationsbemuhungen in konventionelle Systeme aufgrund grundlegend unterschiedlicher Paradigmen und als Folge davon, schwierige Vergleichbarkeit mit klassischen Ansatzen. Erstens wird eine effiziente Trainingsumgebung bereitgestellt, um Deep Reinforcement Learning Agenten ressourceneffizient und intuitiv zu entwickeln und zu trainieren. Zweitens werden Methoden zur Integration des resultierenden Deep Reinforcement Learning Agenten in bestehende Navigationssysteme vorgestellt, indem ein Zwischenplaner vorgeschlagen wird, der die Vorteile beider Paradigmen miteinander verbindet und nutzt. Schlieslich werden Methoden entwickelt, um die resultierenden Navigationssysteme ausgiebig zu evaluieren, mit dem Stand der Technik zu vergleichen, und dadurch eine einheitliche Bewertung von Deep Reinforcement Learning Ansatzen zu ermoglichen. Die Ergebnisse belegen die Uberlegenheit des vorgeschlagenen Navigationssystems in Bezug auf Robustheit, Effizienz und Sicherheit. Bei Evaluationen auf realen Robotersystemen ubertrifft das vorgeschlagene System mehrere Vergleichsansatze um bis zu 48 % in Bezug auf die Navigationssicherheit, und um 14 % in Bezug auf die Gesamtperformance mit den wenigsten Kollisionen, den hochsten Erfolgsraten und den kurzesten Wegen zum Ziel, insbesondere in hochdynamischen Umgebungen. Dies ist ein wesentlicher Schritt zur Ermoglichung der Open-Path Navigation in kollaborativen Umgebungen und offnet Turen fur noch anspruchsvollere und komplexere Aufgabenbereiche in Zukunft.
546 $a English
590 $a School code: 1119
650 4 $a Deep learning. $3 3554982
650 4 $a Curricula. $3 3422445
650 4 $a Systems design. $3 3433840
650 4 $a Neural networks. $3 677449
650 4 $a Robots. $3 529507
650 4 $a Navigation systems. $3 3686148
650 4 $a Packaging. $3 585030
650 4 $a Markov analysis. $3 3562906
650 4 $a Semantics. $3 520060
650 4 $a Robotics. $3 519753
650 4 $a Design. $3 518875
650 4 $a Logic. $3 529544
690 $a 0771
690 $a 0549
690 $a 0800
690 $a 0389
690 $a 0395
690 $a 0796
710 2 $a Technische Universitaet Berlin (Germany). $e degree granting institution. $3 3819406
720 1 $a Lambrecht, Jens $e degree supervisor.
720 1 $a Hafner, Verena $e degree supervisor.
720 1 $a Xiao, Xuesu $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 85-11A.
790 $a 1119
791 $a Ph.D.
792 $a 2023
856 4 0 $u https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31078083