Service robots that operate in domestic environments and assist in the daily housework are still beyond reach when considering the current state of the art. Hence, they pose many interesting technical challenges that currently motivate a considerable amount of basic research in mobile robotics. One of these challenges is the question how service robots could attain the required level of autonomy and reliability to be truly useful. Ideally, if a service robot is deployed in a domestic environment it should require only a short setup time and as few user interactions as possible. When tidying up, it should know where the objects usually belong to. If the robot should set the table, it needs to know which objects are required, where to find them, and how they should be arranged on the table. It is apparent that most of these tasks involve knowledge about the spatial context of objects and how objects are usually arranged in such environments. It would be desirable to have a flexible approach in which the robot adapts to the specifics of its environment by observing it. It thereby could learn the usual object arrangements. Hence, as a basic requirement for achieving a high level of autonomy, service robots need to represent, learn, and utilize knowledge about the relevant spatial relations between objects in man-made environments. If robots are able to utilize representations that take into account the interdependencies between objects in such environments then this would enable them to more efficiently carry out their tasks or to address completely new tasks. For example, robots could more efficiently search for objects, or a robot could reason about missing or misplaced objects in a room or on a table.

In this thesis, we propose several techniques for learning and utilizing spatial object relations. First, we consider the problem of localizing objects. As future household objects and retail products might be equipped with an RFID tag, we first present a technique to localize RFID tags. Further, we show that same technique can also be applied to the complementary situation in which we want to localize a mobile robot based on a known map of RFID tag locations. Given that we can localize objects, we move on to address the question of how a robot can efficiently search for objects in an unknown environment. For this, we present two techniques that both aim at speeding up the search process by taking advantage of background knowledge about usual object arrangements acquired in previously seen, similarly structured environments. Specifically, we are interested in efficiently finding a product in an unknown supermarket. The main idea is to exploit knowledge about the co-occurrence of objects to focus on searching the promising regions first and to postpone the non-promising regions. While both approaches utilize learning techniques to leverage the information of previously seen environments, they both rely on predefined spatial relations, like an object being "in the same aisle" as another object. This motivates the final part of this thesis, in which we aim at learning stable spatial relations between objects. More specifically, we wish to learn spatially coherent object constellations in an unsupervised manner from complex multi object scenes. As an application scenario, we consider tabletop scenes in which the object constellations correspond to place covers. For this, we propose a novel hierarchical nonparametric Bayesian model that represents a prior distribution over scene structures in terms of object constellations. For posterior inference in this model, we present an efficient Markov chain Monte Carlo (MCMC) sampler. By basing our model on the Dirichlet process and the beta-Bernoulli process, the number of object constellations in our model is not fixed. This has practical benefits in the context of lifelong learning, as the robot is able to recognize and integrate previously unseen object constellations into its model in an open-ended fashion and within a single coherent probabilistic framework.