Spatial Modeling and Robot Navigation

This habilitation thesis deals with the broad topic of robot navigation and with its foundations in spatial modeling, scene understanding, and action generation. Robot navigation is the process of autonomously making a sequence of decisions that allows a mobile robot to travel robustly to selected locations in the environment. This ability involves a large set of problems that need to be solved including sensor data interpretation, state estimation, environment modeling, scene understanding, learning, coordination, and motion planning. The ability to navigate autonomously has special practical relevance, since navigations tasks are embedded in most robotic applications. In this work, we present innovative solutions for acquiring models of the environment with mobile robots and for robustly navigating using these models. A representation of the environment is needed for a wide range of robotic applications. Here, a model of the environment representing its geometry is only a starting point. To allow for intelligent decision making or higher level functionalities, a robot needs information about the topology of the space as well as knowledge about objects in the scene in order to act robustly and interact in real world settings. These problems are addressed in this work. We furthermore present novel approaches for decision making, trajectory planning, localization, exploration, and related topics. We propose probabilistic solutions to a variety of these problems and consider different types of robots with different sensing modalities. This includes classical wheeled robots as well as computer-controlled cars, flying vehicles, and humanoid robots. Our solutions are important building blocks that enable robots to operate autonomously and effectively in the real world. The contributions of this habilitation thesis are innovative solutions to a variety of open problems in the context of model learning and efficient robot navigation which captured the attention of research community to-date. Most of the presented approaches are of probabilistic nature, they explicitly consider uncertainty, and include learning techniques. All developed approaches have been applied and evaluated on real mobile robots in realistic settings, have proven robustness, and have shown significant improvements over state-of-the-art methods in robotics.