High-level environment representations for mobile robots

In most robotic applications we are faced with the problem of building a digital representation of the environment that allows the robot to autonomously complete its tasks. This internal representation can be used by the robot to plan a motion trajectory for its mobile base and/or end-effector. For most man-made environments we do not have a digital representation or it is inaccurate. Thus, the robot must have the capability of building it autonomously. This is done by integrating into an internal data structure incoming sensor measurements. For this purpose, a common solution consists in solving the Simultaneous Localization and Mapping (SLAM) problem. The map obtained by solving a SLAM problem is called ``metric'' and it describes the geometric structure of the environment. A metric map is typically made up of low-level primitives (like points or voxels). This means that even though it represents the shape of the objects in the robot workspace it lacks the information of which object a surface belongs to. Having an object-level representation of the environment has the advantage of augmenting the set of possible tasks that a robot may accomplish. To this end, in this thesis we focus on two aspects. We propose a formalism to represent in a uniform manner 3D scenes consisting of different geometric primitives, including points, lines and planes. Consequently, we derive a local registration and a global optimization algorithm that can exploit this representation for robust estimation. Furthermore, we present a Semantic Mapping system capable of building an \textit{object-based} map that can be used for complex task planning and execution. Our system exploits effective reconstruction and recognition techniques that require no a-priori information about the environment and can be used under general conditions.