Narunas Vaskevicius

• Spatial perception capability is a fundamental requirement for autonomous intelligent robots to be effective in real world scenarios. Modern robots are deployed in unstructured and increasingly complex environments. In such environments, 3D perception is indispensable for autonomous robot activities such as mapping, localization, recognition, modeling, manipulation, etc. With the advancement made in 3D sensing technologies these tasks have become more manageable in terms of accuracy and computation time. Typical 3D range sensors produce a point-cloud. This is a set of 3D points computed from the ranges to the nearest obstacles along a 2D raster of beam directions. Point-clouds are the most commonly used representation in 3D perception algorithms, especially in mapping. This thesis investigates an alternative approach to 3D perception and mapping, which is based on surface patches extracted from noisy range measurements. This approach offers many advantages in terms of storage requirements, noise reduction, computational efficiency, and semantic interpretation of the environment. However, it also raises new challenges. Many of these challenges are addressed in this thesis within the context of 3D mapping and object recognition for manipulation tasks. In the first part of the thesis, contributions to surface segmentation and surface fitting with uncertainty estimation for 3D mapping are presented. In the second part of the thesis, a scenario focused on object recognition and localization in the context of autonomous container unloading is considered. The theoretical contributions presented in this thesis are substantiated by experiments on real-world data. Moreover, the majority of algorithms have been successfully integrated in autonomous robotic demonstrators, thus demonstrating their applicability to realistic scenarios.