State-of-the-art robot mapping approaches are capable of acquiring impressively accurate 2D and 3D models of their environments. To the best of our knowledge, few of them represent structure or acquire models of task-relevant objects. In this work, a new approach to mapping of indoor environments is presented, in which the environment structure in terms of regions and gateways is automatically extracted, while the robot explores. Objects, both in 2D and 3D, are modeled explicitly in those maps and allow for robust localization. We refer to those maps as structured object-oriented environment representations or Region & Gateway Maps (RG Maps). The process of building such maps is called Region & Gateway Mapping (RG Mapping). RG Maps and RG Mapping make several contributions to the field of map building of indoor environments. First, RG Mapping automatically recovers the structure of large classes of indoor environments and represents them explicitly. Therefore, novel algorithms have been developed for the detection and recognition of gateways, i.e. transitions between regions. Second, it detects rectangular 2D/3D objects using laser range as well as image data that are used for gateway detection and localization. Third, the semantic description, i.e. annotation of regions and objects, is obtained from human-machine interaction in the context of task assignment. Fourth, a compact Region & Gateway Graph is easily extracted from RG Maps, and it is used for efficient path planning on the global scale. It allows reasoning about the feasibility of a given path and learning of the properties of path segments while the robot moves through the environment. Fifth, due to the clustering of metric data into regions, region-based localization and path planning only need to consider the respective region data. Thus, the complexity of the data association problem for localization and the search space for metric path planning is substantially reduced. The RG Mapping and Navigation System has been fully implemented as a distributed (module-based) system, and runs in real-time on a real robot. Due to its architecture, the system can be easily ported to different platforms, provided that similar sensor data is available. In order to support reliable exploration of and navigation within RG Maps, the presented system features a novel approach to collision avoidance and low-level control. The low-level control allows for very precise and fast pursuing of short trajectory segments, which can be changed at any time. The collision avoidance generates trajectory segments based on the interpretation of the current sensor data and short-distance targets from the path planning process. As a result, the navigation behaviour of the robot is reproducible and predictable, which is a very desirable feature for high-level planning.
Übersetzte Kurzfassung:
Um komplexe Aufgaben autonom ausführen zu können, benötigen mobile Roboter sinnvolle Repräsentationen der Umgebung. Der Stand der Forschung ermöglicht die Generierung sehr genauer globaler metrischer Umgebungskarten und deren Nutzung für einfache Navigationsaufgaben. In dieser Arbeit wird ein neuer Ansatz zur Kartierung von Innenraumumgebungen vorgestellt, bei dem die Struktur einer Umgebung in Form von Regionen und Gateways beschrieben wird. Innerhalb der Regionen, wie z.B. Büros, werden 2D/3D Objekte explizit repräsentiert. Region & Gateway Mapping ist das Verfahren, das RG Karten automatisch aus den Sensordaten generiert. Die Arbeit beschreibt und evaluiert neue Methoden zur Erkennung und Klassifizierung von Gateways, wie z.B. Kreuzungen, zur Generierung von Objekthypothesen in 2D und 3D sowie die Anwendung der Karten für Lokalisation und Pfadplanung im Rahmen eines Robot Kontroll Systems. Dieses System wurde auf einem B21r Roboter in zahlreichen Experimenten erfolgreich getestet.
Veröffentlichung:
Universitätsbibliothek der Technischen Universität München
Login
BibTeX