Exploiting the Unexpected: Negative Evidence Modeling and Proprioceptive Motion Modeling for Improved Markov Localization Jan Hoffmann, Michael Spranger, Daniel G¨ ohring, and Matthias J¨ ungel Institut f¨ ur Informatik LFG K¨ unstliche Intelligenz Humboldt-Universit¨at zu Berlin Unter den Linden 6 10099 Berlin, Germany http://www.aiboteamhumboldt.com Abstract. This paper explores how sensor and motion modeling can be improved to better Markov localization by exploiting deviations from expected sensor readings. Proprioception is achieved by monitoring tar- get and actual motions of robot joints. This provides information about whether or not an action was executed as desired, yielding a quality measure of the current odometry. Odometry is usually extremely prone to errors for legged robots, especially in dynamic environments where collisions are often unavoidable, due to the many degrees of freedom of the robot and the numerous possibilities of motion hindrance. A quality measure helps differentiate the periods of unhindered motion from pe- riods where robot motion was impaired for whatever reason. Negative evidence is collected when a robot fails to detect a landmark that it ex- pects to see. Therefore the gaze direction of the camera has to be modeled accordingly. This enables the robot to localize where it could not when only using landmarks. In the general localization task, the probability distribution converges more quickly when negative information is taken into account. 1 Introduction Selflocalization, the estimation of position and orientation of a mobile robot, remains an important and valuable task for mobile robotics. One of the most successfully applied approaches is called Monte-Carlo-Localization. This method is used in numerous robot navigation problem domains, such as office navigation [1], museum tour guides [13], RoboCup [7], as well as outdoor or less structured environments [9]. We propose 2 extensions affecting the sensor model as well as the motion model. 1. We show how negative information can be incorporated into Monte Carlo localization. The sensor model is extended by modeling the probability of non-detection events.