Learning to grasp objects based on ensemble learning combining simulation data and real data

In this study, deep learning based grasping using a robot has been discussed. A large amount of training data is required for good performance in deep learning. The training data is usually collected with a real robot. However, it is difficult to collect the data sufficient for training the network in terms of time and cost. Therefore, this study presents a method for collecting the training data based on a robot simulator as well as a real robot. The simulation system is composed of a robot, the work environment, and a 2-finger gripper. The convolutional neural network (CNN) was used for training where its input is the RGB image of the object and its output is the pose of the gripper. Furthermore, the ensemble learning method was used to combine real data and simulation data. It is shown that the ensemble learning method that combines multiple classifiers can lead to a higher grasping success rate than a single classifier.