A System for Three-Dimensional RoboticMapping of Underground MinesMichael Montemerlo, Dirk H¨ahnel, David Ferguson,Rudolph Triebel, Wolfram Burgard, Scott ThayerWilliam Whittaker, and Sebastian ThrunOct 27, 2002CMU-CS-02-185School of Computer ScienceCarnegie Mellon UniversityPittsburgh, PA 15213This research is sponsored by by DARPA’s MARS Program (Contract number N66001-01-C-6018) and the National Science Foundation (CAREER grant number IIS-9876136 and regulargrant number IIS-9877033), all of which is gratefully acknowledged. The views and conclusionscontained in this document are those of the author and should not be interpreted as necessarilyrepresenting official policies or endorsements, either expressed or implied, of the United StatesGovernment or any of the sponsoring institutions.AbstractWe describe two robotic systems [6] for acquiring high-resolution volumetricmaps of underground mines. Our systems have been deployed in an operationalcoal mine in Bruceton, Pennsylvania, where they have been used to generateinteractive 3-D maps. Our approach includes a novel sensor head, assembledfrom multiple SICK laser range finders, and a real-time algorithm for scan match-ing that generates accurate volumetric maps. The scan matching algorithm per-forms horizontal and vertical simultaneous localization and mapping (SLAM).Data from the horizontal scans is used to remove artifacts in the vertical scans,and vice versa. The system can construct full 3-D volumetric maps hundreds ofmeters in diameter, even when no odometry information is available.Keywords: Robot mapping, mine mapping, mobile robotics, probabilisticrobotics1 IntroductionThroughout the industrialized world, the lack of accurate maps of inactive, un-derground mines poses a serious threat to public safety. According to a recentarticle [1], “Tens of thousands, perhaps even hundreds of thousands, of aban-doned mines exist today in the United States. Not even the U.S. Bureau of Minesknows the exact number, because federal recording of mining claims was not re-quired until 1976.”1In July of 2002, nine miners were nearly killed in the Que-Creek Mine in Somerset, Pennsylvania when they accidentally drilled into theabondoned Saxmon Mine, releasing millions of gallons of water in the QueCreekmine. This accident highlights the pressing need for accurate maps of abandonedmines.Hazardous operating conditions and difficult access routes suggest that roboticexploration and mapping of abondoned mines may be necessary. Robotic minemapping has been pursued by various research groups around the world. Corkeand colleagues [3] have built vehicles that can acquire and utilize accurate 2-Dmaps of flat mines. Similarly, Baily [14] reports 2-D mapping results of an un-derground area using advanced mapping techniques. The mine mapping problemis made challenging by the lack of global position information underground. Asa result, mine mapping must be approached as a simultaneous localization andmapping, or SLAM, problem [4, 8, 13, 16]. The robot must construct a map of themine, while estimating its own position at the same time. The SLAM problem isknown to be particularly difficult when the environment possesses loops [5, 15].Unfortunately, mines typically contain a large number of cycles, and we know ofno robotic system that could handle such maps. Moreover, none of the existingrobotic mine mapping systems produce accurate volumetric 3-D maps.The systems described in this paper are capable of generating volumetric 3-Dmodels of mines. Our first system makes the common (but unrealistic) assump-tion of a flat floor inside the mine. This system has been used to generate accuratevolumetric maps of relatively flat mines. The second, more elaborate system doesnot rely on a flat world assumption. It uses multiple range finders to generate1See the course page http://www.cs.cmu.edu/afs/cs.cmu.edu/academic/ class/16861-f02/www/for more information.1accurate volumetric maps for mines that change elevation. At the core of bothsystems are 2-D laser range finders, which are used for position referencing andfor the recovery of the volumetric structure of a mine. Our initial system usedtwo such sensors, one for each of the two functions described above. To accom-modate uneven terrain, our second system uses four sensors, effectively extendingthe mine mapping capabilities of our first system into a vertical dimension. Wepresent in this paper a new scan matching algorithm that exploits the overlappinglaser range scans, to correct for noise and alignment errors in the data. The result-ing mine maps are highly accurate 3-D models that can be visualized interactivelyby mining staff.2 The SystemsFigures 1 and 2 show our two volumetric mine mapping systems. Our first proto-type, shown in Figure 1, consists of a modified Pioneer AT robot. It is equippedwith two SICK laser range finders, one pointing forward parallel to the floor, andone pointing upward perpendicular to the robot’s heading direction. In addition,the robot is equipped with two wheel encoders to measure approximate robot mo-tion. The forward-pointing laser scanner is used for simultaneous localizationand mapping (SLAM) in 2-D. Using this data, the robot acquires an accurate 2-Dmap of the environment. The upward-pointing laser is used to reconstruct the 3-Dshape of the walls and the ceiling of the mine, registered in space according toposition estimates gathered from the 2-D map.The limitations of the robotic system are immediately apparent. First andforemost, the system is confined to flat surfaces, due to its inability to sense orincorporate variations in elevation while performing SLAM. In this way, the sys-tem bears close resemblance to existing work on volumetric mapping of indoorenvironments [11, 7, 9], which principally lacks an extension into the third, verti-cal dimension when performing SLAM. Additionally, the robot platform was notrugged enough to handle the uneven, frequently wet terrain common in mines.Most notably, the robot was not able to cross rail-road tracks used to transport oreinside the mine.To overcome these limitations, we developed the sensor cart assembly shown2Figure 1: Mine mapping robot with two laser range finders.in Figure 2. This system is equipped with four SICK laser range finders. Two ofthese sensors point forward, but with a ninety degree offset in orientation. Withthis configuration, SLAM can be performed horizontally and vertically, capturingthe missing dimension