Grasping Perception And Planning Method For Robotic Arm Based On RGB-D Images

In recent years,as the cornerstone of high-end intelligent equipment and the representative of emerging technologies,the rapid development of robot technology has received consistent attention from the academic and industrial circles at home and abroad.Among them,robot grasp technology has entered all aspects of medical,family,industry and other fields.And its ability is very important to the intelligent manufacturing industry.However,the existing grasping methods used by robot manipulators still have shortcomings.When facing objects with novel shapes or materials,the grasping accuracy decreases significantly.At the same time,most research methods use a single RGB-D camera to observe the scene,which can not realize the complete perception of it.This also brings additional challenges to the robot arm grasping.In addition,in the motion planning of manipulator,due to many joints and high degrees of freedom,there are problems of difficult to find the optimal solution,high time complexity and difficult to avoid obstacles accurately,which further increases the difficulty of solving the problem of robot grasping.Therefore,the research on accurate and efficient manipulator grasping technology is imminent.Aiming at the above problems,this paper studies the robot arm grasping perception and planning method based on RGB-D images.The completed work and innovations include the following points.Firstly,aiming at the grasping of novel objects by robotic arms,this paper proposes a domain adaptation method for novel object grasp.By introducing the idea of data mapping in transfer learning,all data are mapped to the regenerated Hilbert space through the Gaussian kernel function,so that the feature distributions of known objects and novel objects are as close as possible.In the grasp network,this paper generates a set of potential grasp poses in the input scene point cloud through a pure geometric algorithm and evaluates them,and then selects the most reliable grasp pose as the final target grasp.The experimental results show that compared with the existing methods,the domain adaptive grasping detection algorithm proposed in this paper improves the generalization of the network,and has higher accuracy and better grasping performance in the manipulator grasp task for novel objects.Secondly,aiming at the problem that the complete scene point cloud cannot be observed in single view RGB-D images,this paper proposes a capture detection method based on object symmetry prediction.Firstly,the scene point cloud is segmented,and the symmetry of the segmented region is used for grasping detection to obtain reliable grasping target points.Finally,the feasible jaw rotation range of grasping target points is determined by collision detection.Compared with the traditional capture detection algorithm,this method considers the lack of scene back information under single view observation.The experimental results show that the proposed grasping detection method based on object symmetry prediction has higher accuracy in the grasping process,lighter order and lower complexity than the traditional shape completion algorithm.Thirdly,aiming at the obstacle avoidance problem of the manipulator in complex scenes,this paper introduces the greedy strategy into the motion planning task.This paper proposes an improved two-way extended RRT algorithm integrating shape completion information,which enables the manipulator to perceive the unobserved field on the back,and alleviates the problem that the traditional RRT algorithm is easy to obtain the local optimal solution.The experimental results show that the proposed method saves time and cost,makes the manipulator motion planning efficient and progressive optimization,and further improves the reliability of obstacle avoidance.