From 3D Point Cloud to Grasping by Using Deep Learning Techniques in Underwater Domain

Abstract

This work is based on the detection of a couple of points for optimal and robust grasping in an underwater domain. The objective is to provide a base for the development of a routine able to autonomously gather information from the surrounding environment in order to provide a robust grasp for underwater intervention robot such as the G500. For this purpose a neural network and a point cloud processing are considered: the former is meant to be able to classify shapes out of a segmented point cloud, the latter have the purpose to reconstruct those shapes out of the classified pat of the point cloud scene. The neural network has been described in its details and the routine explained step by step. Eventually results obtained on real scenes are provided.