Bin Picking Using Deep Learning and Other Methods

Abstract

This thesis focuses on various methods to solve the bin picking problem. An emphasis was put on methods utilizing deep learning. Bin picking systems consist of a robotic arm used to grasp randomly placed objects from a bin. The objects need to be detected, and the grasping needs to be planned without collisions with the other objects or the bin itself. The main objective of this thesis was to study and compare different studies made about the subject.

This paper also contains a general overview of the different components of neural networks, and how the networks are trained to reach an optimal output. One of the methods focused on in this thesis used deep learning to detect the objects, and one used it to plan the grasping of the objects. The method using deep learning to plan the grasping performs well with various objects with different physical properties. Different methods have their own advantages and disadvantages. Because of this, when choosing the method for a bin picking system, the specifc requirements of that system should be considered.