Research On Robotic Object Detection And Grasp Planning Algorithm

Robots play a crucial role in modern industrial production,healthcare,warehousing and logistics,space exploration,and other fields.Grasping is a fundamental operational task for robots,including target classification and detection,grasping detection,and grasping path planning,enabling active perception,real-time interaction,hand-eye coordination,intelligent decision-making,and precise grasping of the environment and target objects.In recent years,although there have been numerous research results on robot grasping,further research and exploration are still needed to promote its widespread application.Its research has important theoretical significance and practical application value.This paper investigates a lightweight target detection model,an intelligent grasping detection algorithm and manipulator path planning algorithm,and conducts simulation verification through ROS system.The main contents include:(1)For the UR5 six-degree-of-freedom manipulator,a forward and inverse kinematics model based on the DH parameter method is constructed,and simulation verification is performed using the Matlab Robotics toolbox.At the same time,the ‘eye to hand’ planar grasping work mode of the manipulator is selected based on its grasping method and handeye combination method.(2)MN-YL-CA target detection model is proposed.Its main ideas are as follows.The YOLOv5 is selected by analyzing existing mainstream target detection algorithms based on the VOC2007,VOC2012,and COCO public datasets;In order to improve the detection speed,the lightweight network MobileNetV3 is applied to replace the backbone network of YOLOv5,which reduces network parameters and computational complexity;The CA attention mechanism is introduced to improve the detection accuracy of the lightweight network.An experimental dataset of eight graspable objects is constructed,and the above model is verified based on it.The results show that compared with YOLOv5,the MN-YLCA algorithm improves 0.14% mAP and 22.53 FPS,reduces 55.7% parameter volume and73.5% computational complexity,and can effectively obtain the category and location of the target object.(3)A grasp detection algorithm based on the GG-CNN is given,which is verified based on the re-labeled Cornell grasp dataset.The experimental results show that the above algorithm can achieve an accuracy of 80% and a detection speed of 40 FPS,and can effectively obtain the grasp pose of the target object.(4)The Improved RRT-connect(IRRTC)path planning algorithm is proposed for the manipulator grasping task,which is respectively verified under the cases of 2D and 3D environments with multiple obstacles.The RRTC algorithm has the shortest planning time and the longest path than RRT and RRT~*,which are discovered by performing them in 2D and 3D space with central and global obstacles.To shorten the path length,the IRRTC algorithm is proposed by integrating the target offset strategy,greedy algorithm,and cubic B-spline.The simulation results show that compared with RRTC,the IRRTC algorithm reduces the path length by 20.1% and 15.5% under the 2D space with the center and global obstacles,and by 32.4% and 33.9% under the 3D space with the center and global obstacles,respectively.(5)The UR5 manipulator system and grasp simulation environment are built based on the ROS system,and the above models and algorithms are deployed into them.Visualguided obstacle avoidance path planning experiments are conducted using the octree map built from point cloud information to verify the feasibility of the IRRTC algorithm on the manipulator.Through the simulation grasp environment of the manipulator,grasp and placement experiments are conducted.The results show that the recognition rate is 92.5%,the grasping rate is 93.75%,and the obstacle avoidance rate is 100%.