Real-Time Robotic Grasp Detection with Multi-Scale Feature Fusion

Grasping is an essential problem that is a serious challenge in area of robotics. In recent years, computer vision based methods have been proved to be effective to solve this problem. With the introduction of deep learning, superior strategies that provide advantages over traditional approaches have emerged. Robotic grasp detection needs to consider both coarse-grained and fine-grained information, while previous works did not take full advantage of the latter, leading to loss of accuracy. In this paper, we propose a novel grasp detection model to predict a five-dimensional representation for grasps with RGB-D images. A fully convolutional network is employed with multi-scale feature fusion structure, which can combine feature information on different scales. Experiments show that our multi-scale model has made significant progress in accuracy compared to a single-scale, while maintaining the performance of real-time computation.