Fitting Primitive Shapes to Point Clouds for Robotic Grasping

Due to the growth of service robots in the last few years, made it possible that robots can carry out human tasks. Houseworks that for humans would seem as common and simple as grasping, moving and interacting with objects, are a challenge both mechanically and computationally in robotics. Therefore, considering the process of grasping objects by robotic machines, this thesis introduces an algorithm that makes the identification of complex objects easier by using a set of geometric primitive shapes already known to the robot. Thus, taking as input the point clouds which define the 3D objects once they have been scanned, an algorithm based on the iterative method RANSAC has been developed. The proposed algorithm is able to fit a set of primitive shapes to an incomplete and noisy point cloud. By splitting input point clouds into smaller sets, the suggested algorithm attempts to find the primitive shape that best fits each of these divisions in order to simplify the 3D objects. Consequently, a robot will be able to recognize unknown objects without the need of having previously stored objects in its database. The huge amount of possible grasps that can be applied to an object is therefore reduced to just a few based on the division into primitive shapes. Additionally, this thesis explains how each of the estimation algorithms have been designed, as well as evaluates the performance of these under different noise conditions.