Research On Target Pose Estimation For Robot Grasping Based On Convolutional Neural Networks

With the continuous progress of technology and rising labor costs,robots have been widely used as substitutes for manual labor.However,most robots currently face difficulties in handling the grasping task of unknown target object pose in complex environments,as they lack flexibility in dealing with changes in the environment,target state,or grasping task.This article takes the UR3 robotic arm as the research object,combines deep learning technology with color image data of the target object,and conducts research on robot grasping target pose estimation technology based on convolutional neural networks to achieve pose estimation and grasping of the target object in complex background.The primary focus of the research and its findings are summarized as follows:(1)In order to simplify the target object model,facilitate computation,and reduce computational complexity,a bounding box based method was adopted to approximate complex geometric objects.By combining the advantages of axis aligned bounding boxes and directional bounding boxes,a bounding box is designed to tightly wrap the target object,and a virtual coordinate system is established based on the key points of the target object to describe the position and direction of the object in three-dimensional space.(2)In response to the problem caused by partial occlusion in the process of capturing target objects,utilizing the advantages of convolutional neural networks in image processing,based on the YOLOv7 feature extraction network as the backbone architecture,in order to obtain more complete contextual information,multi-level pyramid pooling is introduced,and complex feature representations are automatically learned to achieve efficient and accurate feature point extraction of 3D objects,And combined with the Pn P algorithm to solve pose information.Through extensive testing of the pose estimation algorithm designed based on convolutional neural networks on a self created synthetic dataset,it was found that the average 2D reprojection accuracy(2D reprojection)reached 90.23%,and the average distance(ADD)index reached 55.73%.Compared to the Yo Lo-6D algorithm,the designed pose estimation method has improved by 1.43% and 4.24%,respectively,indicating its superiority in solving the occlusion problem of target objects during the grasping process.(3)In response to the difficulty in producing a 6D pose estimation dataset,a synthetic data generation method based on Unity-3D software was adopted to obtain diverse and controllable synthetic data,effectively reducing the difficulty,cost,and time of dataset production.By randomizing the parameters of the synthesized dataset,it is possible to cover a wider range of instances and increase the diversity of the dataset.By utilizing the automatic annotation function,accurate annotation labels such as bounding boxes and semantic segmentation can be automatically generated.This method can not only significantly reduce the difficulty of dataset construction,but also improve the quality and annotation effect of the dataset.(4)A system structure for robot target pose estimation and grasping system was designed.During the system construction process,Zhang Zhengyou calibration algorithm was used for camera internal calibration to eliminate the impact of camera distortion on measurement accuracy.Subsequently,based on the actual working method,the "eye in hand" installation method was selected,and precise hand eye calibration was performed on the UR3 robotic arm using Ar Uco calibration code.(5)Construct a robot target pose estimation and grasping experimental system.An experimental system was constructed on the UR3 robotic arm working platform,and the trained model was deployed to the system for target pose estimation and grasping experiments to verify the pose estimation algorithm and its performance in real-world scenarios.In both single target and multi target mixed scenes,experimental results show that the pose estimation algorithm combining RGB image information of the target object performs well in both pose estimation and actual grasping,with success rates of 92.5% and 91.7%,respectively.It successfully achieves pose estimation and grasping operations for the target object.