Parts Pose Inference And Grasp Planning For Programming By Demonstration In Industrial Assembly

With the development of robotics technology,the application of industrial robots is gradually moving from traditional manufacturing industry to emerging manufacturing industry such as 3C and hardware.Many products produced in 3C electronics industry have the characteristics of variety,small batch and short cycle,which lead to higher requirements for the convenience and efficiency of both programming and deployment.Programming by demonstration technology can reduce programming difficulty and shorten development time by learning and understanding the activities of human operators,which makes it easier to use industrial robots.This thesis carries out a research in the background of the programming by demonstration in industrial assembly tasks,which involves three aspects:inference of assembly relations and parts poses,gripper selection and grasp planning.The main contributions are as follows:1.An assembly relations and poses inference method for small parts based on integrating visual observation and CAD models is proposed.Considering the limitations of visual observations and CAD models,the knowledge of CAD is analyzed,then a deterministic template assembly graph is constructed.Observed parts are matched with CAD models according to composite evaluation criterias.Matched parts are corrected with the prior knowledge existing in the CAD.In the actual production,the assembly targets are often changed and inconsistent with CAD models.Take this into consideration,for those cannot be matched with CAD mod-els,probabilistic assembly graph is employed.Experimental results verify the feasibility,accuracy and flexibility of the proposed method.2.A method for selecting gripper types of unknown parts based on deep convolutional neural networks is proposed.The gripper type refers to the category of end-effector,such as par-aller girpper and vacuum suction.A model based on gradient features is proposed as the baseline and replaced by the fully convolutional neural networks.A large-scale simulation dataset is generated by rendering and automatical labeling owning to the lack of open source gripper selection dataset.In order to generalize the model from the simulation to reality,data transformation based on the point cloud projection is proposed,which can achieve high performance without using any real data.The accuracy of gripper selection is 94.3%in the simulation dataset and 94₁%in the real dataset.Experimental results indicate that the pro-posed method is flexible,reliable and can generalize well from simualtion to reality without trainning with real dataset.3.A cross complementarity method for simultaneous gripper selection and grasp planning is proposed.A location heat map is generated from the gripper selection network which im_plictly learns location information.Grasp candidates can be generated from the heat map and evaluated to obtain the final planning result.An optimal procedure inlcuding grasp sampling and grasp evaluation is adopted to make results more stable.The evaluation stage with a fusion model can simultaneously evaluate the stability of multiple end-effectors and perform better than single model.As for the sampling process,the traditional cross entropy method is initialized with a uniform distribution,while the proposed method takes the heat map as the initial distribution and can achieve higher accuracy with shorter convergence time.An intelligent grasp experimental platform is built to test gripper selection and grasp planning for unknown parts by a robot arm.The overall success rate of gripper selection and grasp planning is 97.06%.The results of simulation and realistic experiments show that although the model is completely learned from the simulation dataset,it can generalize well to the realistic data with a high planning and execution accuracy.