For our mobile robots Rhino and Minerva, which operated in the Deutsches Museum Bonn and the US-Smithsonian's National Museum of American History, the robustness and reliability of our Markov localization system was of utmost importance. Accurate position estimation was a crucial component, as many of the obstacles were ``invisible'' to the robots' sensors (such as glass cages, metal bars, staircases, and the alike). Given the estimate of the robot's position [Fox et al. 1998b] integrated map information into the collision avoidance system in order to prevent the robot from colliding with obstacles that could not be detected. Figure 12(a) shows a typical trajectory of the robot Rhino, recorded in the museum in Bonn, along with the map used for localization. The reader may notice that only the obstacles shown in black were actually used for localization; the others were either invisible or could not be detected reliably. Rhino used the entropy filter to identify sensor readings that were corrupted by the presence of people. Rhino's localization module was able to (1) globally localize the robot in the morning when the robot was switched on and (2) to reliably and accurately keep track of the robot's position. In the entire six-day deployment period, in which Rhino traveled over 18km, our approach led only to a single software-related collision, which involved an ``invisible'' obstacle and which was caused by a localization error that was slightly larger than a 30cm safety margin.
Figure 12(b) shows a 2km long trajectory of the robot Minerva in the National Museum of American History. Minerva used the distance filter to identify readings reflected by unmodeled objects. This filter was developed after Rhino's deployment in the museum in Bonn, based on an analysis of the localization failure reported above and in an attempt to prevent similar effects in future installations. Based on the distance filter, Minerva was able to operate reliably over a period of 13 days. During that time Minerva traveled a total of 44km with a maximum speed of 1.63m/sec.
Unfortunately, the evidence from the museum projects is anecdotal. Based on sensor data collected during Rhino's deployment in the museum in Bonn, we also investigated the effect of our filter techniques more systematically, and under even more extreme conditions. In particular, we were interested in the localization results