École Nationale Supérieure de
Techniques Avancées

Abstract

The grasping of objects by autonomous robots is still an important issue in robotics. Many methods have been tried, but today, learning, as in all fields, is the most popular and generates the most research. In order to train all these learning algorithms, it is necessary to have datasets with several thousand grasps, presenting diversity and labelled in a way consistent with the given problem. In the context of this study, the objective is to create a dataset allowing to train robust grasp prediction algorithms. To do this, I worked on the random creation of scenes including randomly placed objects as well as on the creation of random or non-random grasp positions. I also set up metrics for the analysis of grasps in order to measure their quality according to the robustness criterion. Finally, I propose a tool for setting up and simulating scenes and object grasp as well as grasp analysis tools. Unfortunately, this last point is not perfectly developed and still requires research so that it is perfectly operational.