Research On Robot Multi-Object Grasp Based On Deep Learning

Grasping operations is a common task in robotics.At present,most mature robotic grasping systems work in a structured environment.Once the environment,grasping tasks,and target conditions change,they need to be reset,lacking robustness and flexibility.However,in an unstructured environment,when facing the unpredictability of grasping targets,the randomness of the poses of the grasping targets and the stacking of multiple grasping targets,robot's grasping becomes difficult.This thesis intensive studies of robotic grasping scene understanding,grasping pose estimation,etc.This implement a method for multi-target grasping tasks,and use physical grasping Take experiments for verification.The main contents are as follows:(1)Describes the framework of the entire robotic grasping system.The depth camera and robotic arm are the basis of the system.In view of the lack of holes in the depth map obtained by the depth camera,an image repair algorithm based on a fast-moving algorithm was used to repair the depth map.In vision-guided robotic arm technology,hand-eye calibration is an important prerequisite for grasping.This article uses Eye-to-Hand hand-eye model and HALCON vision software to realize hand-eye calibration of camera and robotic arm.(2)Research on multi-objective pixel-level semantic segmentation based on full convolutional networks.Aiming at the particularity of the multi-object capture scene,a multi-object capture data set is established to train the segmentation network.The segmentation algorithm of improved DeepLabV3 + algorithm is proposed.Compared with the original algorithm,it has better segmentation effect on small targets,stronger network space perception ability,and better segmentation of stacked scenes.Through comparative experiments,the performance of the proposed algorithm is verified.(3)Research on robot grasping pose estimation method based on deep learning.Describes robot detection problems.RG-D information is used as input,Faster R-CNN is used as the basic framework,and multi-scale feature maps are obtained by fusing the FPN network to improve the detection of small targets.ROI Pooling is replaced by ROI Align to improve the positioning accuracy of the grab box.It is verified through comparative experiments,and the proposed algorithm achieves 96.9% accuracy on Cornell's captured dataset.(4)The tower builds a robotic grabbing system based on the Kinect depth camera and Kinova robotic arm.Under the ROS framework,it realizes the information transfer of each sub-module and drives the robotic arm to complete different grabbing tasks.This dissertation proposes a multi-object grasping method based on deep learning,which realizes multi-object segmentation and grasping pose estimation in the grasping scene,and the robot grabbing system has a success rate of 91% for a single object,while the success rate of multiple objects placed randomly is 88%,and the success rate of multiple objects placed in a stack is 78%.