Search

In robotic planetary exploration missions, robots are deployed to autonomously explore and map the large and unstructured environments of planetary surfaces. While a robot should be able to execute a mission task mainly autonomously, for space exploration missions, it is important to have the opportunity to observe and adapt the robotic exploration task. Operators and scientists require to supervise the robot at the available communication time slots and understand the decisions made by the robot. For this we propose a generalized concept for robotic exploration based on Multi-Criteria Decision Making (MCDM) to model, implement and conduct exploration tasks. Our general formulation supports scientists by designing the autonomous exploration behavior of a robot to reach specific missions goals. In robotic exploration tasks, robots repeatedly decide where to move next. We define locations at the boundary to unknown areas - exploration goals - and locations in already visited areas - revisiting goals - to be the solution space of this decision problem. To model a certain exploration behavior, the goal locations are evaluated by a set of criteria and conditions. The criteria and condition values for each goal location are compared, applying a MCDM method to find the next goal location, which best matches the defined mission goal. Thereby, we introduce two novel multi-attribute utility functions and transfer the Preference Ranking Organization Method for Enrichment Evaluation (PROMETHEE II) to solve decision making in robotic exploration. To cope with the limited computational resources of space rovers, we extend the PROMETHEE II algorithm to decrease the required computational resources. Applying our generalized concept, we examine four exploration use cases, deduced from the Exploration Roadmap of the International Space Exploration Coordination Group (ISECG). In the first use case, the robot has to autonomously survey a region of interest. To tackle the trade-off between exploration efficiency and map quality, we implement an integrated exploration, which applies active loop closing to optimize an underlying SLAM graph. In our second use case, we implement a directed exploration to increase the scientific output while exploring a region of interest. It incorporates knowledge about the probability of detecting a feature of interest, i.e., a specific type of rock requested by the scientists. As our third use case, we implement an exploration behavior in the fashion of drive-by science, whereby the robot is directed to a predefined point of interest, while simultaneously gathering new information about the environment on its way. For our fourth use case, we apply the same concept to model a multi-robot exploration task, which coordinates a heterogeneous team of two robots. We demonstrate all four use cases on real or simulated space rover prototype hardware. In a total of more than sixty experiments, we evaluate our methods and analyze the implemented exploration behaviors. Bei der Exploration von planetaren Oberflächen erkunden Roboter autonom fremde Umgebungen und bauen eine Karte dieser auf. Auch wenn ein Roboter in der Lage sein sollte, eine planetare Explorationsmission autonom durchzuführen, ist es unerlässlich, eine Möglichkeit zur Überwachung und gegebenenfalls zur Anpassung der Exploration zu haben. Operatoren und Wissenschaftler müssen den Roboter während der kurzen Zeitabschnitte, in denen eine Kommunikation zum Roboter aufgebaut werden kann, überwachen können. Wir schlagen ein allgemeines Konzept für robotische Exploration basierend auf multikriteriellen Entscheidungsverfahren vor, um verschiedene robotische Explorationsaufgaben zu modellieren, zu implementieren und auszuführen. Unsere allgemeine Formulierung ermöglicht es Wissenschaftlern, das Explorationsverhalten eines Roboters so zu gestalten, das verschiedene zuvor definierte Missionsziele erreicht werden können. Bei der robotischen Exploration entscheidet ein Roboter immer wiederholend, welchen Zielpunkt er als Nächstes ansteuert. Der Lösungsraum dieses Entscheidungsproblems beinhaltet dabei Zielpunkte an der Grenze zu bis dahin unbekannten Gebieten und Zielpunkte in Gegenden, die der Roboter bereits besucht hat. Um das Explorationsverhalten des Roboters zu modellieren, werden mehrere Kriterien und Konditionen über die Ziele ausgewertet. Um das nächst beste Ziel zu bestimmen, werden die Kriterien und Konditionen für die einzelenen Ziele mithilfe eines multikriteriellen Entscheidungsverfahrens verglichen. Wir stellen zwei neue multikriterielle Nutzenfunktionen vor und übertragen die bekannte Methode Preference Ranking Organization Method for Enrichment Evaluation (PROMETHEE II) auf das Entscheidungsproblem der robotischen Exploration. Die Rechenleistung eines Roboters, der für den Weltraumeinsatz konzipiert wurde, ist begrenzt. Darum erweitern wir PROMETHEE II, um die nötige Rechenleistung der Methode zu verringern. Wir leiten vier verschiedene Anwendungsfälle von der Exploration Roadmap der International Space Exploration Coordination Group (ISECG) ab und wenden unser allgemeines Konzept an, um diese Anwendungsfälle zu modellieren und zu untersuchen. Im ersten Anwendungsfall exploriert und vermisst der Roboter autonom eine unbekannte Region. Wir implementieren eine Explorationsstrategie, bei welcher aktiv nach Schleifenschlüssen im darunter liegenden SLAM Graphen gesucht wird, um diesen zu optimieren. Damit lösen wir den Konflikt zwischen einer effizienten Exploration und einer guten Kartenqualität. Im zweiten Anwendungsfall implementieren wir eine Explorationsstrategie, welche im ”Vorbeifahren” wissenschaftliche Entdeckungen machen soll. Der Roboter muss eine vorgegebene Reihe von globalen Zielen nacheinander ansteuern und während der Fahrt zu den einzelnen Zielpunkten möglichst viele neue wissenschaftliche Informationen über die Region, welche er traversiert sammeln. Im dritten Anwendungsfall stellen wir eine Explorationsstrategie vor, welche es zum Ziel hat, den wissenschaftlichen Ertrag zu erhöhen, während der Roboter eine Region exploriert. Dafür berücksichtigen wir die Wahrscheinlichkeit, ein Objekt von wissenschaftlichem Interesse an einem Ziel zu detektieren. Ein solches Objekt kann z. B. ein bestimmter Stein sein, der von Wissenschaftlern als wissenschaftlich relevant eingestuft wurde. In unserem fünften Anwendungsfall stellen wir eine Strategie vor, um ein heterogenes Team von Robotern zu koordinieren, welche gemeinsam eine Region explorieren. Dafür verwenden wir das gleiche allgemeine Konzept wie für die Anwendungsfälle, in denen nur ein Roboter alleine agiert. Wir demonstrieren alle vier Anwendungsfälle auf realen robotischen Systemen oder mithilfe einer Simulation. Insgesamt führen wir mehr als 60 Experimente durch, um alle Explorationsstrategien und unser allgemeines Konzept zu analysieren und zu validieren.

Robotic platforms represent a new frontier for data acquisition in a wide range of monitoring and exploration missions, including agriculture, surveillance, post-disaster assessment, and environmental sensing. A key challenge to realize their full potential is deciding how an autonomous agent should act to gather the most useful data about an uncertain environment within the set of its resource constraints. To address this, this thesis investigates the problem of active sensing: in such scenarios, how can the actions of the agent be planned in order to maximize the efficiency of the data collection process? It proposes new methods for environmental mapping and informative path planning as crucial elements towards achieving this goal. The strategies are developed in the context of two distinct applications: (a) terrain monitoring, and (b) active sensing under localization uncertainty, with a focus on Unmanned Aerial Vehicle (UAV) systems. The main contribution of this thesis is an informative planning framework that is applicable for general active sensing tasks. The overall approach integrates individual contributions on three different fronts, which aim to address the challenges associated with information gathering in uncertain 3-D environments with computationally limited systems. Firstly, a new method for environmental field mapping is presented for terrain monitoring scenarios. The strategy exploits a Gaussian Process (GP) model as a prior for recursive Bayesian data fusion with probabilistic, variable-resolution sensors. In doing so, it supports mapping using dense visual imagery without the computational burden of standard GP regression, making it suitable for online on-platform applications. Moreover, it accommodates noisy sensors with altitude-dependent performance, as relevant for UAV-based systems. Secondly, an online, adaptive informative path planning algorithm is introduced for generating continuous trajectories to collect data in resource-constrained missions. A key feature of the method is that it uses the output from a discrete grid search as an informed prior to initialize a trajectory optimization routine and thereby improve its convergence in a large and complex objective space. This strategy also enables trading off between computational efficiency and solution accuracy for deployment on systems with limited computing power. Finally, methods are proposed to account for the robot pose uncertainty in active sensing tasks. Unlike prior work, the approach propagates this uncertainty into both the mapping and planning modules towards improving the robustness and accuracy of information gathering. A new utility function is developed that allows the robot to auto matically trade off between exploiting the existing map to maintain good localization and exploring areas to acquire new data in a principled way, without relying on any manually-tuned parameters. The formulation is derived in the context of a GP-based monitoring scenario and is also applicable across different learning problems. The developed framework is modular, and can be tailored to a wide range of active sensing problems. Extensive simulation studies were conducted to evaluate the approach, examining how it performs against existing methods both as an integrated system as well as in terms of its key components. The main findings show that the proposed approach effectively: (a) produces maps with similar certainty and accuracy in significantly less time compared to current planning strategies; (b) can focus on adaptively mapping specific areas of interest; and (c) improves upon both field map accuracy and irobot localization by accounting for the pose uncertainty in informative planning. Results using an experimental dataset demonstrate system integration and validation in a photorealistic UAV-based terrain monitoring scenario. Finally, field tests are presented to demonstrate the algorithms implemented and running in real-time on robots for various data gathering tasks, including vegetation mapping on a farm. The framework is made publicly available as an open-source package.