Generic 2D/3D SLAM with NDT Maps for Lifelong Application Erik Einhorn and Horst-Michael Gross Ilmenau University of Technology, Germany Abstract—In this paper, we present a new, generic approach for Simultaneous Localization and Mapping (SLAM). First of all, we propose an abstraction of the underlying sensor data using Normal Distribution Transform (NDT) maps that are suitable for making our approach independent from the used sensor and the dimension of the generated maps. We present some modifications for the original NDT mapping to handle free-space measurements explicitly and to enable its usage in dynamic environments with moving obstacles and persons. In the second part of this paper we describe our graph-based SLAM approach that is designed for lifelong usage. Therefore, the memory and computational complexity is limited by pruning the pose graph in an appropriate way. I. I NTRODUCTION Simultaneous localization and mapping (SLAM) is one of the fundamental challenges in mobile robotics. It constitutes a difficult problem as consistent mapping depends on the knowledge of the current robot’s position, while robust self- localization on the other hand requires an accurate map of the environment. Therefore, the localization and the map- ping process are inherently coupled [22]. Consequently, the SLAM problem has been thoroughly analyzed for decades and researchers came up with many different solutions. However, when it comes to the practical application of SLAM, it often is used for map acquisition only during an offline map learning phase as part of the initial setup of the robot in its novel environment [7]. During the robot’s operation phase, this map then is used for robot localization, i.e. pose tracking, for instance by using particle filter based Monte Carlo localization [6]. In our previous real-world applications where we imple- mented tour guide robots and interactive shopping assistants [10], we also followed the philosophy of a map learning phase and a separate operation phase. However, today’s complex applications such as robot companions that assist elderly people in their home environments [11] require a paradigm shift. Typically, these environments are semi- static or dynamic, i.e. the location of obstacles like chairs or tables change over time. Therefore, a separated map learning phase is no longer acceptable. Instead, the mapping phase must continue during the whole operation time of the robot to adapt the map permanently to the changes in the environment. This results in the so called lifelong SLAM problem. A lifelong SLAM algorithm that is suitable for such scenarios must be able to constantly update the map of *This work has received funding from the Federal State of Thuringia and the European Social Fund (OP 2007-2013) under grant agreement N501/2009 to the project SERROGA (project number 2011FGR0107) the environment without increasing the complexity for map updates with new measurements. Moreover, it must operate in realtime to continuously provide estimates of the robot’s location to other navigation modules, like path planners. Modern assistance robots typically have a variety of sensors that provide different kinds of information about the robots environment. Laser Range Finders provide two dimensional range data that can be used to create 2D maps. Depth cameras on the other hand provide depth images that are suitable to create 3D maps. Also a single camera mounted in front of the robot can provide such 3D information when it is used for monocular scene reconstruction [4]. Therefore, we are interested in applying a generic SLAM approach that is able to process such 2D and 3D information equally well. Our contribution in this paper is twofold: We first intro- duce a novel, generic mapping technique that can operate with various range sensors of different dimensionality to generate compact 2D and 3D maps. Based on this mapping technique, we present a lifelong SLAM approach that sat- isfies all of the aforementioned requirements and hence is suited for real-world applications. In summary, our proposed approach: 1) is implemented in a generic way for 2D and 3D mapping 2) operates in realtime and allows online robot localization 3) allows lifelong mapping with constant complexity 4) operates in semi-static or dynamic environments and adapts the map to the changing environment This paper is organized as follows. The next section outlines the state of the art in SLAM and robot mapping. In section IV, we describe our approach in detail. In section V, we show several results that we have obtained using the presented approach for different kinds of sensors. Finally, we conclude with an outlook for future work. II. RELATED WORK As stated before, a large variety of different SLAM ap- proaches are available. Some techniques interpret the SLAM problem as a filtering problem and apply Extended Kalman filters [2] or Rao-Blackwellized Particle Filters [9], [22] to solve it. Others apply smoothing techniques [13], [12] to solve the full SLAM problem, i.e. beside the estimation of the most consistent map they keep the complete robot trajectory as part of the estimation problem. While these approaches provide direct solvers for the SLAM problem, others, e.g. g2o [16] exploit the sparsity of the SLAM problem by formulating it as a pose graph optimization problem. The problem of such optimizers is that they are not robust against outliers in data association. Hence, wrong in: Proc. 6th European Conference on Mobile Robots (ECMR 2013), Barcelona, Spain