| Robot localization is the problem of estimating the position of a robot given its sensory observations, control actions and a map. There exist efficient approaches to this problem, for example, Kalman and extended Kalman filters and Monte Carlo localization. These approaches address the problem by using probabilistic models for the motion and sensing of the robot. The effectiveness of these approaches is naturally dependent on the accuracy of these models.; Finding good models, or calibration, traditionally involves long and error-prone measurements and manual tuning. Therefore, an automatic calibration technique, which is the subject of this research, is highly desirable.; This thesis takes a Bayesian approach to calibration. Expert knowledge about the models is encoded as a prior distribution in parameter space. A new belief about parameters' distribution is inferred from data collected onboard the robot. Since analytical inference of this distribution is not feasible, a Markov Chain Monte Carlo algorithm is used to draw samples from this distribution. The effectiveness of our technique is demonstrated both in simulation and on a real robot. |
