End-to-End Learning of Object Grasp Poses in the Amazon Robotics Challenge

Abstract

The Amazon Robotics Challenge (ARC) is a robotics competition aimed to advance warehouse automation. One of the engineering challenges is making the system robust to and being able to handle a wide variety of objects, as would be the case in a real warehouse. In this paper, we shortly describe our system used in ARC featuring a method to obtain object grasp poses containing the location of the object as well as orientation for the grasp by using a convolutional neural network with an RGB-D image as input. Through our entry in ARC 2016, we show the effectiveness of our method and the robustness of our network model to a large variety of object types in dense and unstructured environments wherein occlusions are possible.