A New Robotic Grasp Detection Method based on RGB-D Deep Fusion

Abstract

Grasping is one of the most widely used tasks of robots. The application of computer vision can improve robot intelligence. Previous methods simply treated the problem of robotic grasping detection similar to object detection, which ignores the characteristics of the grasping problem, leading to a loss of accuracy. Additionally, treating depth images equally with RGBs is unreasonable. This study proposes a new grasp detection model using an RGB-D deep fusion module that combines multi-scale RGB and depth features. An adaptive anchor box-setting method based on a two-step approximation was designed. With the network-sharing structures of target and grasp detection, the target category and appropriate grasp posture can be obtained end-to-end. Experiments show that compared with other models, ours achieves significant improvement in accuracy while maintaining real-time computing performance.