Learning to Grasp: an end-to-end sampling approach for robotic grasping in 6-DoF pose

The ability to grasp objects is essential in multiple contexts, ranging from the mimic of human activities to automation of industrial tasks. Recent works based on datasets obtained through simulation have proven to achieve promising performances on unseen objects also in real-case scenarios. Nevertheless, many of them still strongly rely on approaches built on top of ad-hoc geometric heuristics to generate grasp candidates, failing in generalizing to different settings and making it hard to reproduce the same approach in different environments. Moreover, some of the heuristics require different and sometimes independent modules to reduce the redundancy of generated grasps. In this thesis, we propose a lightweight end-to-end solution for the generation of 6-DoF parallel-jaw pose grasps starting from partial view of the object which relies on completely learned sampling techniques, making it easy to translate the same approach to different datasets and settings. We start by introducing a self-supervised pre-trained feature encoder which is able to extract both local and global informations of the input shape in a combined embedding and show how this method outperforms traditional encoders. Moreover, we designed a sampling technique to suit the grasping task. We leverage pre-existing literature on learning to sample to develop a module able to select grasp contact points without imposing geometric custom contraints. This approach makes it possible to better generalize to different object types and shapes.