Research On Robot Grasping Pose Detection Based On Deep Learning

Robots have been widely used in various industries and have greatly improved the productivity of society,robot grasping technology is one of the hot spots in robotics research.Currently,most industrial robots perform grasping tasks in a relatively fixed and structured environment.However,when external factors such as work scenarios,grasping tasks and target objects change,the robot is required to have good grasping detection performance,otherwise it will lead to grasping failure.Deep learning has good nonlinear fitting ability,knowledge transfer ability,and has achieved remarkable success in fields such as computer vision.In this paper,we use deep learning methods to focus on the robot grasping pose detection problem.Based on deep learning,a grasping pose detection network model is designed to improve the accuracy and generalization of grasping pose detection,so that the robot can still complete the grasping action in the face of unknown environment and unfamiliar target objects,and the model is experimentally validated.The main research contents of this paper are as follows:(1)Firstly,the robot grasping scheme is designed,and the main processes are object detection,pose detection and grasping execution.For the object detection problem,the image acquisition and processing method,depth camera imaging,and ranging principle are analyzed.For the grasping pose detection problem,we determine the grasping representation parameters and establish the grasping pose representation method in this paper.For the grasping execution problem,the pixel coordinate system to robot base coordinate system conversion relationship is derived to obtain the position of the object in the robot base coordinate system.(2)Secondly,different object detection methods are analyzed,a target detection network model is established,and the model is optimized using two schemes,improved scale detection and improved loss function,to improve the accuracy and rate of the model for object detection.The Cornell dataset is re-labeled so that the labels meet the requirements of the object detection task,and the object detection model is trained on the dataset.It is shown by testing that the improved object detection algorithm in this paper has good detection accuracy.(3)Then,the basic theory of convolutional neural network is investigated,and the residual network structure is used to establish the grasping pose detection model,and the feature pyramid mechanism is introduced to optimize the model.Based on the Cornell dataset,a self-built dataset is constructed as a supplement to the model training.The dataset is expanded by using data enhancement processing,and the grasping pose detection model is trained on the dataset.The experimental results show that the pose detection model in this paper can achieve good detection accuracy.(4)Finally,the robot grasping experimental platform is built and camera calibration and handeye calibration are completed.The robot grasping experiments are designed and the grasping experimental data are recorded.Analysis of the experimental results finally verifies that the grasping pose detection algorithm proposed in this paper has good accuracy and generalization.