Research On Dynamic Path Planning And Multi-objective Recognition For The Manipulator In Cluttered Scenes

As the basic core equipment for the development of national industrial automation,industrial robotic manipulator plays an important role in the automobile manufacturing,aerospace,3C home appliances and logistics express delivery industries.In the context and background of smart manufacturing,the future factory must be highly intelligent,informative and Human-Robot Collaborative(HRC),and the working scenes in which the robotic manipulator is faced with must be dynamic and uncertain.However,the current robotic manipulator has insufficient ability of intelligent active perception and safe autonomous interaction in above application scenarios.There are no feasible and better solutions in relevant research fields at home and abroad,and there are still many key issues that need to be solved or tackled urgently.This thesis takes industrial manipulator as the research object,and focuses on dynamic path planning for robotic autonomous obstacle avoidance and multi-objective visual recognition in cluttered scenes.Dynamic path planning for robotic autonomous obstacle avoidance and multi-objective visual recognition as well as 3D pose estimation are gradually studied and researched in depth,which aims to provide novel ideas for solving some key technical problems involved in the autonomous operation for the robotic manipulator and provide practical guidance for industrial application.In terms of the existing Rapidly-exploring Random Tree(RRT)planning algorithm cannot well solve the dynamic path planning issue of autonomous obstacle avoidance for the robotic manipulator in dynamic environment.A path planning method called Smoothly-RRT for static obstacle avoidance is proposed,which improves the basic RRT algorithm from node extension and trajectory optimization smoothing.The Bi-RRT-Star dynamic path planning algorithm based on the object direction improving sampling function is presented,which improves from connection strategy,heuristic dense sampling and adjacent nodes extension.Furthermore,a heuristic greedy multi-step extension algorithm is employed.The simulation and experimental results show that,compared with the Basic-RRT and Bi-RRT algorithms,the searching speed and searching efficiency are significantly improved,the average planning time is shorter,the planning success rate is higher,the generation path is smoother,and has the ability to avoid both regular and irregular moving obstacles dynamically in the global environment.In terms of the lack of pose recognition research at home and abroad in cluttered scenes where multi-objective coexistence of large and small objects is common,as well as the lack of depth resolution and field of view of RGB-D sensors inherently,a pose recognition method of hybrid configuration combining RGB-D sensor and hand-eye camera is proposed.The hand-eye camera is fixed on the end effector of the robotic manipulator,which captures the object image of the scene.An approach to visual recognition and 3D pose estimation for large object with weak texture surface is proposed based on Kinect V2 sensor,which can quickly and robustly identify and locate different types of large objects in cluttered scenes.For the issue that the small object in the field of view cannot be recognized,Kinect is used to perceive the large object next to the small object,which can guide the robotic manipulator from the initial pose to reach around the small object and apply the hand-eye camera of robotic end effector to perceive the small object.The proposed method and two mainstream pose recognition methods are compared to verify that the proposed method has smaller pose estimation error and higher recognition rate,which achieves accurate identification and automatic sorting for large and small objects in cluttered scenes.In terms of some shortcoming in the field of pose recognition for small objects with weak texture feature in cluttered scenes at home and abroad,in order to solve the issue that Iterative Closest Point(ICP)algorithm fails to obtain global optimal solution since the deviation from scene point cloud to object CAD model is huge in nature.An automatic pose recognition method for small cluttered objects is proposed based on low-cost monocular vision 3D reconstruction.The scene point cloud obtained by 3D reconstruction is sequentially denoised and filtered,improved normal vector estimation,Color signatures of histogram of orientations(Color-SHOT)local feature descriptor generation,two coarse matching online between the scene point cloud and CAD model based on Random sample consensus(RANSAC)algorithm and proposed Remote Closest Point(RCP)algorithm,and Levenberg Marquardt(LM)-ICP fine registration.Finally,the accurate pose estimation of small cluttered objects is obtained.The comparative experiments between the proposed method and two mainstream pose recognition methods are conducted.The results show that the proposed pose recognition method has smaller pose estimation error and higher recognition rate,which realizes pose estimation and automatic classification for small cluttered objects with weak texture in cluttered scenes using robotic manipulator based on monocular camera.An integrated application management software is designed and developed for multi-objective identification sorting and dynamic obstacle avoidance path planning in cluttered scenes.Based on a brief analysis of the software requirements,the overall architecture and functional sub-modules of the integrated software management platform are designed in detail.And a practical engineering application example of the multi-objective identification sorting and dynamic safety obstacle avoidance is made up of three kinds of water pipe joints,connecting parts,gaskets,bolts and human-like obstacles as well as human dynamic arm interference.It shows that the proposed method can accurately identify the poses of the above-mentioned target parts,and at the same time realize safely avoidance of both static and dynamic obstacles in the process of identification,grasping and sorting.It can further demonstrate practicality and technical superiority of the multi-objective automatic pose recognition and dynamic path planning method for robotic autonomous obstacle avoidance proposed in this paper.