Robotic Grasping Pose Detection For Unknown Object

The dexterous grasping ability is the key technology for robots to move from a structured environment to complex application scenarios.Currently,robotic grasping has been widely used in structured operation scenarios.However,in the face of unknown targets in complex unstructured environments,there is still the problem of difficult detection of grasping poses,which severely limits the task adaptability of robots.Therefore,this thesis focuses on the theoretical research and experimental verification of robot grasping posture detection for unknown objects.The main research work completed is as follows:(1)Eye-To-Hand calibration of robots based on small ball calibrators.To achieve the attitude conversion from camera coordinates to the robotic arm during robot grasping,an Eye-To-Hand hand-eye calibration platform is established in this thesis.In this paper,an Eye-To-Hand handeye calibration platform has been established.Using a small ball-based hand-eye calibration method,the Kronecker product and dual quaternion are used to solve the rotation and translation problem,and the hand-eye calibration of the robot grasping system is realized.(2)A data enhancement algorithm based on local transformation of key blocks.A data enhancement method based on dynamic block transformation of key regions,Trans-MixHidden,is proposed,which divides the image into key regions and non-key regions.Random dynamic blocks are introduced to perform local dynamic transformation or pixel erasure within a random region.Random dynamic blocks increase the difference in data information and effectively expand image data.The generalization performance is verified in the grasping detection algorithm,and the difference in training function is reduced from 1.87 to 0.52.The experiment shows that the data enhancement algorithm designed in this thesis has a good effect on alleviating the overfitting phenomenon of model training.(3)An algorithm for robot grasping posture estimation based on key point detection.Aiming at the difficulty of robot grasping pose estimation for unknown targets,a pose estimation method based on key point detection was proposed and a Center Net SPF grasping pose estimation network was designed.Classify and transform the grab angle into object category features to enhance the generalization ability of unknown target detection;Design a dual channel feature fusion feature extraction network structure to improve the prediction accuracy of the algorithm.After testing,the algorithm in this article was trained and tested in the Cornell dataset based on image segmentation and object segmentation,with accuracies of 98.1% and97.6%,respectively.The unknown object detection performance was tested in the Jacquard dataset,with an accuracy of 95.5%.(4)An experimental platform for robot object grasping has been built.Relevant experiments were carried out around the detection and grasping operation of unknown object grasping posture,and experiments were designed for unknown object single target,occluded target,and mixed multiple targets grasping.In the public dataset testing and object grabbing experiments,the accuracy of object grabbing posture detection reaches 95.3%,and the success rate of object grabbing reaches 88%,verifying the effectiveness of the relevant algorithms in this article.In summary,the research work carried out in this thesis around the detection technology of unknown target robot grasping posture has certain theoretical significance for enriching the theory of robot hand eye coordination control and has certain engineering application value in robot grasping operations in unstructured environments.