Application of Deep Learning Algorithm Based on Multimodal Regularization in Detection Robot Grasping

This paper uses the depth learning method to settle the question of robot detection and capture in the RGB-D view of the sight comprising the object. A large number of candidates can be evaluated by using a two-step cascade system, which is faster and more effective than manual design. There are two deep-step cascade system network, a top value of the first detection system is reappraised by the second system. The first meshwork runs faster because of its fewer functions and can prune a large number of candidate fetches effectively. The second network is more powerful, so the speed is relatively slow, and the top-level detection value of the first system can be re-evaluated. To deal with multimodal input effectively, e propose a weighted structure regularization method in view of multichannel group regularization. The experimental results indicate the depth learning arithmetic can effectively improve the performance of RGBD robot grasping data sets, And the new objects can also be captured successfully.