Navigation of an atonomous mobile tobot supported by supplementary information

The presented project deals with the localization and world-modeling for autonomous mobile robots that navigate in typical indoor environments. The main goal of this project was to develop reliable techniques for pose estimation which would satisfy the following requirements: The robot should be able to estimate its position in environments which are not modified for the navigation purposes in any way. The robot should be able to interpret sensory data and generate appropriate world models autonomously. The robot should determine its position through a comparison of the currently obtained sensory information and the known world models. The localization and world-modeling should work also with simple and cheap sensing systems (e.g. sonars and wheel encoders). Supplementary information provided by humans should easily be incorporated into the world-models and navigation processes. In general, reliable localization and world modeling in typical indoor environments turn out to be very challenging research topics. Namely, in typical indoor environments usually no global positioning systems are available. In addition, such environments are partially unpredictable and not completely observable; e.g. typical indoor environments are populated by humans and there are objects that can easily be moved. There are also other unpredictable aspects of the interaction between the robot and its environment such as severe sensing and odometric errors. However, in this work we show that reliable navigation in indoor environments can be achieved if we use very simple models of different processes and spatial classes in conjunction with appropriate heterogeneous information. The main contributions of the presented work are: Investigation of the information types and representations which are useful for the localization and world modeling in typical indoor environments. We show that very heterogeneous world models can be generated autonomously by using simple sensors, in particular wide-angle sonars and wheel encoders. We introduce a novel context-based approach to sonar data interpretation and present a new pose tracker, which can cope with severe odometric and sensing errors as well as inconsistent world models. We introduce the Geometric Consistency Filter which is used for the estimation of the geometric consistency of the used world models. We also show that the relative geometric inconsistencies, resulting from simultaneous localization and map-updating, remain bounded under certain conditions, which can easily be satisfied in typical indoor environments. We also show that robust Markov-localization within a network of rooms can be achieved by using heterogeneous information types. The presented approaches to localization and world-modeling have been verified through extensive experiments with a physical mobile robot and proved to be very robust in typical, partially unpredictable indoor environments, which have not been modified for the navigation purposes in any way. From the experimental results it is also evident that our solutions to certain crucial localization and world-modeling problems are superior or supplementary to other known approaches.