Universal approximation theorems in deep learning and convolutional neural networks for industrial gripping

In this thesis, after having introduced the essential aspects of neural networks, we deal with the relevant theory of Universal Approximation Theorems and we propose a practical solution example of the development of a neural model for evaluating and searching the best gripping positions of sheet metal parts with a suction cup gripper. The work aims to be the summary of our development path in this regard. It presents the fundamental theory for understanding the functioning of a neural network (the structure of the models, the strategies and fundamentals of the training, and the testing methods), the more mathematical one for understanding the theory of Universal Approximation Theorems (passing through the Hahn–Banach Theorems, the theory of signed measures and their Fourier transforms, and finally the Riesz Representation Theorems), and the strategies and implementation of a concrete algorithm, that uses neural networks and other modern tasks, for searching the best gripping position for a sheet metal part with a suction cup gripper. In particular the overall approach is based on building a position evaluation algorithm, reproducing it with a neural network and using this network to search for the optimal solution with a genetic approach.

This seminar concerns the results of Nicola’s MSc thesis (advisor Dimitri Breda, co-advisor Andrea Pavan from Starmatik, Spresiano).