DefinePK

Abstract

The techniques of gripping elements by means of manipulative robots have had a wide advance during the last years, allowing them to perform complex tasks such as the follow-up of paths with evasion of both static and dynamic obstacles, in order to reach an objective, hold it and deliver it to the user. However, these algorithms do not allow to directly grasp the elements of interest when they present occlusions, which leads to the need to develop a new sequence of displacement that allows recognizing and eliminating possible occlusions on the desired object, before performing the grip. and delivery thereof. The development of the research work was carried out through the use of Convolutional Neural Networks (CNN) type DAG-CNN (Directed Acyclic Graph CNN), an anthropomorphic robot and VRML simulation, where an occlusion elimination sequence was programmed that allows remove unwanted elements located on the object of interest, before making its grip and delivery to the user, both in a physical and virtual environment. The program achieved 100% success in holding and delivering desired objects with less than 5 occlusion elements, with 99% accuracy in the DAG-CNN for the classification of the desired element with and without occlusions.