MMH-GGCNN: Multi-Modal Hierarchical Generative Grasping Convolutional Neural Network

Abstract

The advent of deep learning has changed the trends of various research areas including robotics. Especially, with the fast developing computer vision technology, deep learning based grasp pose detection algorithms have been presented. Unlike traditional algorithms, deep learning based ones can generalize well to unknown environment. However, there still exists the problem of computation time due to common pipeline of sampling and ranking grasp candidates. To deal with this problem, recently, lightweight network with moderate performances called GG-CNN has been developed. To further boost the performance by exploiting multi-modality and hierarchy of grasp components, we propose multi-modal hierarchical generative grasping CNN (MMH-GGCNN) with a small number of parameters. In the experiments, MMH-GGCNN achieves the improved accuracy of 91.9679% accuracy on the Cornell Grasping Dataset.