State Estimation, Planning, and Behavior Selection Under Uncertainty for Autonomous Robotic Exploration in Dynamic Environments

Lidoris, Georgios

kassel university press, ISBN: 978-3-86219-062-1, 2011, 167 Pages

URN: urn:nbn:de:0002-30633

Zugl.: Kassel, Univ., Diss. 2011

| Price and available forms -->

Content: One of the most challenging visions of contemporary robotics is to bring autonomous robots into the real world and enable them to operate efficiently and safely in natural, populated environments. Perhaps the greatest obstacle for the creation of reliable and truly versatile autonomous robotic assistants is the uncertainty that exists in the physical world. As a consequence, robots always have to rely on incomplete information and are unable to build exact representations of the world. Only estimations can be made and all decisions about the actions of the robot have to be made based on these estimations. At the same time actions have unpredictable results and lead to unpredicted states. Dealing with uncertainty is therefore a prerequisite in order to create systems that are capable of assisting humans
in their natural environment in intelligent and versatile ways.

The contributions of this thesis are tractable and scalable methods for state estimation, autonomous exploration, planning and action selection for autonomous mobile robots operating under uncertainty in the real world. Probabilistic approaches are used to represent uncertainty, which outperform alternative techniques in many real-world applications. Sensing, planning and motion execution are interconnected and are examined in a joint framework.

Novel algorithms for autonomous robotic exploration with active sensors are introduced. It is shown how the trajectory of mobile robots as well as the positioning of active on-board sensors can be planned over a finite time-horizon so that estimation errors are minimized, while at the same time the unknown environment is explored.

A Bayesian framework is introduced that takes into account state estimation uncertainty in the behavior selection process of autonomous robots. It integrates recursive estimation of dynamic environment models and efficient action selection based on these uncertain estimates. This is the first approach that uses a formal Bayesian framework to fully analyze the domain of autonomous robotic exploration. Starting from the joint distribution, which includes no assumptions and simplifications but is impossible to be calculated, it is shown how this can be decomposed to a hierarchy of distributions and models that are interconnected
and can be calculated efficiently with state-of-the-art approximation techniques. Also expert knowledge or the preferences of the system designer can be integrated in the proposed action selection framework. This knowledge can be acquired through learning from humans. An approach that allows the robot to learn probabilistic behavior selection models from humans is introduced. It is also shown how the learned models can be used in the proposed Bayesian framework for action selection.

Finally the navigation and behavior selection system of the Autonomous City Explorer (ACE) robot is analyzed as well as a novel method for system component interdependence analysis. Probabilistic models of the interdependencies between system components, such as perception, planning and execution, can be learned, enabling the determination of crucial system components with respect to robustness. The learned models facilitate design choices, without the need of fully examining the deterministic interrelations in the system. Furthermore the gained knowledge can be integrated into the on-line reasoning process
of the system itself to enhance its autonomy. Results from extensive and unique field experiments are presented in order to evaluate the proposed algorithms.

Publication is available in following forms:

Full text (pdf-file, not printable - 17.00 MB)
view PDF
Full text (pdf-file, printable, with costs - 17.00 MB) 20.00 Euro
(free of charge in kassel University network - you are in kassel University network if you are in the workplace, or you are using a pc in the ITS or in the library)
download PDF - Attention! with costs, because you are not in kassel University network!