Robot Positioning by Supervised and Unsupervised Odometry Correction

The aim of this thesis is to perform robot positioning, based on an odometry which is continuously corrected by different landmark detection systems demanding as less modifications as possible for the environment. Two independent correction systems (a supervised and an unsupervised) were implemented into two different experiences which represent the subject of this thesis. The supervised experiment uses grid lines painted on the floor which are detected by a single light sensor underneath the robot which cannot distinguish between horizontal and vertical lines. The robot knows the geometry of the grid lines and its estimated position which is calculated by odometry. A new position probability model calculates the assumed robot position and transforms this single sensor information into a reliable position and orientation indication of the robot. The intended trajectory is slightly modified in order to optimize the correction algorithm by guiding the robot more efficiently over close grid lines. The theory was implemented and tested on a real Khepera robot. Investigation and modeling of the odometry error is the main subject of this first experiment. The second experiment demonstrates continuous odometry correction by an unsupervised correction system. Different kind of unsupervised neural networks classify the robot’s rough sensor signals. A statistical algorithm extracts “meaningful” classes from the generated pool of classes and uses these as reference points for odometry correction. The robot position will be calibrated to these reference points if the robot detects the same “meaningful” sensor information which generated earlier the corresponding reference point. What is going to be used as a reference point is initially unknown, and no extended preprocessing of the sensor signals is done. The localization method allows a robot to correct its position calculated by odometry without any prior information about its environment and its sensor configuration. This localization approach takes full advantage of the sensor abilities, because no model unsuitable for the sensor configuration was imposed. In a higher level, sequences of reference points are grouped to places in order to establish a more reliable indication for a certain region in the environment. In opposite to reference points, places do not indicate an exact robot position, but more a fuzzy region which can be still detected even if the robot trajectory or the environment is slightly influenced by noise or other circumstances. Place recognition is done by Markov chain detecting reference point transitions. The result is a topological map of the environment representing these places and the common transitions between. Thesis: Robot Positioning by Supervised and Unsupervised Odometry Correction iii