On Correcting Sewer Robots' Odometry Errors by Reasoning

When inspecting sewers, it is required that discovered objects like damages or previously unknown laterals are entered into the existing sewer map at the correct metric position. On the other hand, odometry conditions in sewers can be very poor, and they may vary greatly with slip, slope, and sewage current. The paper describes a simple reasoning method for correcting the odometry error according to local conditions. The method is based on correcting past measurements, and hence the corresponding new map entries, according to safe localization points at individual landmarks and to a model of the local odometry error or errors in between the localization points. A sketch of the intended application of the method in an autonomous, multi-segment, articulated sewer robot is given. 1 THE APPLICATION PROBLEM Autonomous mobile robots live in space, must cope with space, and hence notoriously reason about space. Examples of such reasoning are path planning, trajectory planning, and navigation under robots' various perception impairments. In this paper, we are dealing with localization, i.e., the question: Given a map of the environment, where exactly am I? Some applications allow this problem to be solved satisfactorily by plastering the environment with beacons or by using external references like the Global Positioning System (GPS). Some applications don't: GPS signals, for example, are shielded in most buildings, and if the purpose of some service robot is to work in inaccessible areas, then installing beacons may be impossible or 0 cause high cost. Much research is currently done in localization methods for mobile robots; we will address some related work en passant. Our localization problem is somewhat particular, according to the particular application area: Sewers. The long-term background is the intention to build autonomous mobile robots for continuously inspecting inaccessible (i.e.,< 1m diameter) sewers in order to record problems like damages or hazardous substances. We do not further expose this area and the rationale for using autonomous robots; see [Hertzberg et al., 1998a] for an overview. The domain characteristics of interest for this paper can be explained easily. A metric map exists for every sewer, specifying, among other things, manholes, main pipes and laterals. The map is correct with respect to manholes, mains, and most laterals. During inspection, a number of interesting features like damages (cracks, occlusions, grown-in tree roots) or other items (unrecorded laterals) have to be entered into the map. Our users require measured positions to be exact within a range of 10 cm. Sensor conditions for autonomous sewer robots are poor. Sewers are cylindric, narrow, dark, dirty, slippery, and wet. Pipes vary greatly in diameter, material and wearout. Odometry is particularly di cult because of the much varying, sometimes very high, degrees of slip, slope, and sewage current. Conventional sewer inspection uses tethered camera platforms, where position is determined in the archaic (and somewhat error-prone) way of measuring the cable as it gets unreeled. Autonomous robots must do without, and so the intuitive problem is: How can an autonomous sewer robot know with sufcient metric precision where it is when spotting a reportable feature, to enter it correctly into the map? (1) Our own previous work for topologically correct positioning [Hertzberg & Kirchner, 1996] and metrically correct map entries [Schonherr, Hertzberg, & Burgard, 1998] has made some very weak assumptions about the correctness of turning maneuvers in the sewer and about the correctness of object detection and classi cation; in consequence, we had to use probabilistic (POMDP) methods for position estimation that resemble those used, e.g., in hallways, where translation imprecision, drift, and rotation imprecision would blur the position estimation (e.g., [Koenig & Simmons, 1998]). For a new sewer robot platform, which we are developing in a joint research project [Cordes et al., 1997; Hertzberg et al., 1998b], turning errors are no longer likely to occur. As rotation and lateral drift are physically impossible in pipes, a simple localization method may be used that compensates more e ectively for the translational localization error. The next section describes the basic idea of the method. Sec. 3 sketches re nements of the model of the locally varying odometry error. Sec. 4 concludes.