Automatic 3D Object Recognition and Localization for Robotic Grasping.

With the emergence of Industry 4.0 and its highly re-configurable manufacturing context, the typical fixedposition grasping systems are no longer usable. This reality underlined the necessity for fully automatic and adaptable robotic grasping systems. With that in mind, the primary purpose of this paper is to join Machine Learning models for detection and pose estimation into an automatic system to be used in a grasping environment. The developed system uses Mask-RCNN and Densefusion models for the recognition and pose estimation of objects, respectively. The grasping is executed, taking into consideration both the pose and the object's ID, as well as allowing for user and application adaptability through an initial configuration. The system was tested both on a validation dataset and in a real-world environment. The main results show that the system has more difficulty with complex objects; however, it shows promising results for simpler objects, even with training on a reduced dataset. It is also able to generalize to objects slightly different than the ones seen in training. There is an 80% success rate in the best cases for simple grasping attempts.