Convolutional neural networks for bin picking applications with a collaborative robot and a 2D camera

In bin picking applications, robots manipulate randomized objects placed in a bin. For that, the objects must be located before picking. The procedure of localization relies heavily on data from cameras. In the literature many approaches are available by means of the traditional computer vision approaches that work very well in an industrial scenario, while for more challenging situations artificial intelligence must be adopted. So, this thesis addresses this problem by presenting a possible solution for detecting objects and estimating their poses. Many convolutional neural networks exist to detect objects but still a reliable approach for estimating the pose of casual objects does not exist, in this thesis the problem will be solved for a certain category of objects, that is objects in which the orientation is clear from a purely geometrical point of view. In addition, the problem will be solved by using a collaborative robot, so that the task can be performed in a shared environment with people without hurting them.