Deep Learning Based Intelligent Robot Grasping Strategy

Abstract

For human, grasping is based on learning. Kids generally have poor grasping strategy compared to a grown up individual who essentially learns better grasping strategy based on experience. This research presents a vision based humanoid learning strategy for robot grasping. Given the RGB-D information of a scene with an object to be grasped, the problem of estimating the optimal grasp of a rectangle has been addressed. To solve this problem a model based on deep-learning approach with convolutional neural network has been developed. For training the model Cornell Grasping Dataset [9] has been used. The implementation involves pre-processing, whitening, normalizing and flattening the data before the features being fed to a deep neural network. The research has been implemented on Intel i7-6700HQ with RAM: 16GB, GPU: Nvidia GTX960M 4GB. Our model performs extremely well on the dataset and produces an accuracy of 94.1%. The model is characterized by simplicity, robustness, generic nature having sufficient scalability.