Seite - 40 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Fig. 2. System overview of the transport robot [5]. safely navigate between buildings. The robot is depicted in Figure 1 and is based on a pioneer 3-AT platform which allows the robot to navigate indoors as well as outdoor. Additionally, the robot has three laser scanners to detect obstacles. Two laser scanner are mounted horizontally to detect obstacles, like cars or people. Furthermore, these two lasers are used to localize the robot. The third laser is mounted tilted down to scan the ground in front of the robot. This laser is used to build a map of the local terrain. Besides the laser scanners, the robot has a GPS sensor for the localization and mapping. To improve the accuracy of the odometer of the robot an inertial measurement unit (IMU) is used which is mounted on the robot. To perform the transportation task, the robot uses the system architecture depicted in Figure 2. The robot uses its sensors to estimate the current location. This is done using the robots odometer, the IMU, GPS and the horizontal laser scanners. Due to this redundancy, the estimation of the current location is stable in areas where one sensor may yield wrong results, e.g. the GPS sensor near tall buildings. To perform the estimation, the robot matches the sensor readings with the information of a topological map. The topological map consists of several small maps which are linked to each other to allow the robot to only keep small maps to be localized. Using the estimation of the current location together with a road map, the robot plans a high-level path for naviga- tion. The roadmap describes possible traversal routes in the environment on a higher level. Due to this abstraction, the planning can be done very efficiently even in the case of large environments. After generating a high-level plan, the plan is passed to the lower-level planner which tries to find a valid path in the environment for each path segment in the high-level plan. This is done by considering the current small local map of the environment as well as the sensor data which are used to build a cost map. If a valid path is found the robot tries to follow this path as accurate as possible. To incorporate the information of the terrain the robot uses the tilted laser to perform a terrain analysis. Afterward, the resultsof this terrainanalysisareused toupdate thecostmap. Thus, holes, as well as small objects which are below the horizontal laser scans but bigger than the robots clearance, are added to the cost map as obstacles. This allows the robot to consider the terrain in the low-level planning. In the following two sections, we will discuss the local- ization as well as the navigation in more detail. III. LOCALIZATION Starting from an initial known position the robot needs to know its location during the entire delivery. This is done throughonepartof the robot systemwhich isused to localize the robot. This localization should ensure that the robot has an estimation of its global position. First, the robot corrects its odometer to get a good estimation of its 2D position using dead reckoning. Afterward, it uses the created topological map to localize itself. To correct the odometer of the robot we use an unscented Kalman filter (UKF) [6]. The Kalman filter uses the raw odometer of the robot to perform a prediction of the robot pose.Thisprediction is formedinaprobabilisticmannerwith a position and a covariance matrix specifying the uncertainty. The covariance matrix is defined in such a way that the linear speed has a higher accuracy as the rotational speed, as the rotation is badly estimated through the raw odometry due to the slippage of the wheels during rotation. To correct the prediction the IMU data are used. The IMU data is used to provide an additional estimation of the robots velocity in all three axes as well as the global orientation the robot has in space. As in the case of the raw odometry, the IMU data update the estimate in a probabilistic manner with the help of a covariance matrix. The covariance matrix for the IMU data is formed in such a manner that the rotational speed is estimated more accurately than with the raw odometry. Due to the use of the Kalman filter, we have a better estimation of the robot pose instead of the very noisy raw odometer of the robot. After the odometer is corrected the robot can perform its estimation on the topological map. The topological map is a graph with vertices which represent positions in the world and edges which represent connections between those posi- 40