Mobile robot navigation in dynamic environments

Service robots which act in environments populated by humans have <br>become very popular in the last few years. A variety of systems exists <br>which act for example in hospitals, office buildings, department <br>stores, and museums. Furthermore, several multi-robot systems have <br>been developed for tasks which can be accomplished more efficiently by <br>a whole team of robots than just by a single robot. These tasks <br>include surface cleaning, deliveries, and the exploration of unknown <br>terrain. Whenever teams of mobile robots are operating in the same <br>environment their motions have to be coordinated in order to avoid <br>congestions or collisions. At the same time the robots should perform <br>their navigation tasks in a minimum amount of time. Thus, <br>sophisticated path planning techniques are needed that fulfill these <br>requirements. Since the joint configuration space of the robots is <br>typically huge and grows exponentially with the number of robots, <br>existing path planning methods for single robot systems cannot <br>directly be transferred to multi-robot systems. <br> <br>Many existing path planning methods for multi-robot systems are <br>decoupled, which means that they first plan paths for the individual <br>robots independently. Afterward, they check if the robots would get <br>too close to each other if the paths were executed. In such a case the <br>paths are recomputed to avoid these conflicts. Many decoupled methods <br>assign priorities to the individual robots. These priorities define an <br>order in which the paths of the robots have to be recomputed. By <br>computing the path of a robot, the paths of the robots with higher <br>priority are considered as fixed. This way, the size of the search <br>space is extremely reduced. Most of the existing prioritized decoupled <br>methods use a fixed priority scheme (order of the robots). However, <br>the order in which the paths of the robots are recomputed has a <br>serious influence on whether a solution can be found at all and on how <br>efficient the solution is for the overall multi-robot system. <br> <br>In the first part of this thesis we present an approach which searches <br>in the space of all priority schemes to find an order of the robots <br>for which a solution to the path planning problem can be computed. <br>During the search, we utilize constraints between the priorities of <br>the robots which are automatically derived from the task <br>specification. After an appropriate priority scheme has been found, <br>our technique tries to improve it by using a hill-climbing strategy. <br>Our search method can be used to find and optimize paths generated by <br>any prioritized path-planning technique. In several experiments with a <br>real-robot system as well as in simulation we show that our approach <br>produces efficient solutions even for difficult path planning <br>problems. <br> <br>The second part of this thesis is focused on robots acting in <br>environments populated by humans. These systems can improve their <br>behavior if they react appropriately to the activities of the <br>surrounding people and do not interfere with them. In contrast to a <br>multi-robot path planning system, the future movements of people are <br>not known. Therefore, the robots have to be able to detect people <br>with their sensors, to identify them, and to learn their intentions in <br>order to be able to make better predictions of their future behavior. <br>In this thesis we present an approach to learn typical motion patterns <br>of people from sensor data using the EM algorithm. Furthermore, we <br>describe how the learned patterns can be used to predict future <br>movements of the people. Afterward, we explain how this knowledge can <br>be integrated into the path planning process of a mobile robot. <br>Finally, we introduce a method which automatically derives Hidden <br>Markov Models (HMMs) from the learned motion models. These HMMs can be <br>used by a mobile robot to predict the positions of multiple persons <br>even when they are outside its field of view. To update the HMMs based <br>on laser-range data and vision information we apply Joint <br>Probabilistic Data Association Filters. In practice, the robot becomes <br>uncertain about the positions of people if it does not observe them <br>for a long period of time. We therefore propose a decision-theoretic <br>approach to determine observation actions that are carried out while <br>the robot is executing its tasks. Practical experiments carried out <br>with our mobile robot demonstrate that our method is able to learn <br>typical motion patterns of people, that the navigation behavior of the <br>robot can be improved by predicting the motions of people based on the <br>learned motion patterns, that the derived HMMs can be used to reliably <br>maintain a probabilistic belief about the current positions of <br>multiple persons even if they are currently not in its field of view, <br>and that our technique generates effective actions that seriously <br>reduce the uncertainty in the belief about the positions of people.

Author dc:creator

• Bennewitz, Maren

Contributors dc:contributor

• Burgard, Wolfram

• Mehrroboternavigation
• Lernen von Bewegungsmustern von Personen
• Roboternavigation
• Zustandsschaetzung
• Vorhersagen von Personenbewegungen
• Multi-robot navigation
• Learning motion patterns of people
• robot navigation
• state estimation
• predicting the motions of people

Repository record source_url

https://freidok.uni-freiburg.de/data/1362

OAI identifier oai:identifier

oai:freidok.uni-freiburg.de:1362