Visual Completion Of 3D Object Shapes From A Single View For Robotic Tasks

The goal of this paper is to predict 3D object shape to improve the visual perception of robots in grasping and manipulation tasks. The planning of image-based robotic manipulation tasks depends on the recognition of the object's shape. Mostly, the manipulator robots usually use a camera with configuration eye-in-hand. This fact limits the calculation of the grip on the visible part of the object. In this paper, we present a 3D Deep Convolutional Neural Network to predict the hidden parts of objects from a single-view and to accomplish recovering the complete shape of them. We have tested our proposal with both previously seen objects and novel objects from a well-known dataset.

I. INTRODUCTION

Knowing the complete 3D geometry of an object is indispensable for the physical interaction between the robots and the outside world such as object recognition, grasping, and object manipulation. In this work, we aim to tackle the problem of occlusion in grasping and manipulation tasks through predicting the complete 3D shape from a single 2.5D depth view. If the shape of the object was known, robots could get some ideas of what actions to consider like path planning and generating stable grasps. For this objective, we designed and trained a 3D convolutional neural network to do the shape reconstruction. This is a very challenging task because different 3D models can be obtained from the same single view. Therefore, our solution should have the ability of generalization.