Industrial bin picking (extraction of one piece from a stack) is more and more used in many areas (automotive, food, waste treatment ...) and is a complex task to automate. A bin picking system should indeed be able to analyze the data of a visual sensor, to extract a potential grasp and then to transpose it into actions performed by the robot. The current approaches work but they still need a human parametrization for each processed object.

To deal with the increasing automation of these tasks, it is necessary to minimize human interventions when introducing new object in the system. Our main goal is to use deep learning techniques to allow the robot to use its previously acquired knowledge on new unseen objects. This will lead to more autonomous and adaptive robotic systems.