Page - 43 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Fig. 6. Separator to detect grass with the help of the laser scanner [5]. Fig. 7. Detection of the grass though the laser scanner [5]. We start our discussion of related research with the robot presented in [11]. The robot could take a long tour through Munich without a prior created map or GPS information. Instead, the robot was using its sensors to react locally in a safe manner and asked humans for information about the direction. This was done by approaching humans and the recognition of basic commands to derive the direction of the desired destination. In contrast, our robot has a prior created map which allows it to move autonomously without asking for directions. This is also desirable in the case of a transport robot which should transport goods to a customer. In [12] the method to deal with large maps was described. The authors use a topological map to allow an efficient representation of large areas. The vertices in the topological map are spots of interests such as a square or a crossing. The edges represent paths between these places. For each edge, a traversal behavior is defined. Thus, one can use different behaviors to perform the traversal. With the help of this method, the robot could drive autonomously in a park. Our robot uses, in contrast, a topological map which contains enough information to allow the robot to be always localized not only in interesting places. Furthermore, the robot uses a denser road map allowing it to plan its route more accurate. A very close related work to ours was presented in [13]. The robot navigated more than 3km in the city Freiburg in an autonomous fashion. To localize itself, the robot used a topological map where each vertex in the graph contains a map of one part of the environment. In contrast, our approach additionally used the GPS signal for estimating the robot pose within the particle filter. To navigate the robot, the method presented in [13] created a high-level plan using the graph of the topological map. Each vertex is connected to those vertices in the graph which allow moving between these two locations. Thus, using this graph the robot can derive a simple high-level plan for the navigation. Whereas the robot uses a planner on grid map basis to navigate between different vertices of the topological map. This contrasts with our approach as we use a finer grained road map for the high-level planning which allows us to choose the path more precisely. VI. CONCLUSION AND FUTURE WORK The transportation of goods is an essential part of our today’s economy. The transportation often takes place in outdoor environments by delivering goods to costumers. To provide cost-efficient and flexible deliveries, robots are a promising solution. In this paper, we presented an autonomous transport robot which is capable of navigating in large scale outdoor envi- ronments. To perform this transportation, the robot addresses the problem of a large-scale environment, uneven ground, and grass which should only be traversed if necessary. To deal with the large scale of the environment the robot uses a topological map. This map stores areas of the environment which are loaded on demand. This allows that the robot only needs to keep a small part of the environment in its memory and perform the localization on it. We furthermore showed how the robot can exploit the topological map to switch between the different parts to allow the robot to be localized during the complete delivery. To deal with the uneven ground, the robot builds an elevation map for its local environment. Afterward, the robot determines within the elevation map dangerous terrain and avoids it. To deal with the grass we have shown a simple solution with a linear classification for laser scan measurements. This detection allows the robot to detect grass precisely enough to avoid the grass if possible. The robot presented in this paper mainly used several laser scanners to localize itself and it is left for future work to add more sensors to perform localization as well as navigation. Especially cameras would be of interest as they allow a detailed localization in many areas which don’t offer features for a laser scanner. The additional use of a camera would increase the quality of the terrain classification. REFERENCES [1] P. R. Wurman, R. D’Andrea, and M. Mountz, “Coordinating hundreds of cooperative, autonomous vehicles in warehouses,” in Proceedings of the 19th national conference on Innovative applications of artificial intelligence - Volume 2, ser. IAAI’07. AAAI Press, 2007, pp. 1752–1759. [Online]. Available: http://dl.acm.org/citation.cfm?id=1620113.1620125 [2] E. Guizzo, “Three Engineers, Hundreds of Robots, One Warehouse,” Spectrum, IEEE, vol. 45, no. 7, pp. 26–34, 2008. [3] C. Mu¨hlbacher, S. Gspandl, M. Reip, and G. Steinbauer, “Improving dependability of industrial transport robots using model-based tech- niques,” in Robotics and Automation (ICRA), 2016 IEEE International Conference on. IEEE, 2016, pp. 3133–3140. [4] S. Thrun, W. Burgard, and D. Fox, Probabilistic robotics. MIT press, 2005. 43