Deep Learning Based Monocular Robot Grasp Algorithm

Intelligent grabbing of industrial robots is an important part of smart factories and has a great impact on improving production efficiency.Current solutions can now be divided into three types.A normal practice is manually placing scattered parts on the conveyor belt,which is a semi-automated and requires additional labor costs.Another practice is using the depth camera or binocular stereo to obtain the depth map and training the gripping algorithm with artificial design feature or label data training,which increases the hardware cost.Some people collect a big dataset from a lot of robot grasping tries,which is time-consuming.The big datasetset is used to improve grasp success rate.This dissertation will use monocular image and industrial small datasets to conduct research on the scattered parts grasp.Our monocular robot parts grasp can be divided into three stages: grasp detection,monocular depth estimation and key point estimation.In the stage of grasp detection,this dissertation will optimize the performance of the state-of-the-art detection algorithm on the industrial scenes and obtain a better algorithm for grasp detection.In the monocular depth estimation stage,this dissertation will design a neural network and corresponding training method for industrial capture scenarios,which takes high-resolution input images as input and obtain high-quality depth estimation maps.In point estimation stage,this dissertation will obtain the RGB-D patches from the corresponding regions in RGB images and depth estimation map with the grasp detection results.The point estimation network takes RGB-D as input and the final predicted coordinates are the accumulation of multi-sacle predictions.In order to improve grasp success rate,this dissertation will propose the grasp probability calculation and serialization capture strategy to reuse the context information in the serialization grasp.This dissertation will establish a small data set of industrial scenes for the proposed cascaded method,that is target detection,depth estimation and key point estimation.An average precision of 0.8224 was obtained on the grasp detection data set.The depth estimation algorithm designed for the characteristics of the part grasp scene achieves the root mean square error of 0.0482.The keypoint estimation algorithm for the RGB-D input has a root mean square error of 0.0056 on the grasp keypoint dataset.Using the above algorithm cascade,this dissertation achieves a 70% capture success rate on our part grasp task.