Research On Robot Object Grasp Method Based On Fusion Of Object Pose Estimation And Sampling Evaluation Strategy

In the complex and dynamic Home-Environments,grasping objects,as a crucial skill of the service robot,is essential for improving the service quality.At present,researchers have carried out a lot of research about robotic grasping and have proposed many effective methods,including perceptual planning and sampling evaluation.However,in the complex environments,various objects with different shape and interference from the background,other objects or sensor noise restrict the efficiency and precision of the robotic grasping significantly.For instance,there is a significant calculation error in perceptual planning,due to the influence of other objects or sensor noises,which cause lower grasp stability and adaptability for different scenarios.Although the sampling evaluation methods can achieve higher grasping stability and better flexibility for different situations with multiple objects or stacked items.The number of grasp candidates to be evaluated with the high-dimensional representation of sampling evaluation methods is too much,resulting in low grasping efficiency.As a result,we present a robotic grasping strategy based on object pose estimation and sampling evaluation in this work.The core idea of the proposed approach is to integrate these two methods mentioned above,so that they can have a complementary advantage with each other,and improve the grasping efficiency while ensuring the grasping stability.The main work of this paper is as follows:Firstly,the proposed method makes the fusion of the perceptual planning and sampling evaluation to achieve grasping objects effectively in the complex background.It reduces the long time-consuming for evaluating candidates by the guidance of object pose perception,improves the grasping accuracy by subsequent grasp quality calculation.The process of the proposed method includes object instance segmentation,object point cloud extraction,object pose estimation,grasping candidates sampling,grasping candidates evaluating and robotic grasping execution.To avoid the lack of local point cloud,we employ multi-perspective to obtain point cloud of the object and further improve the grasping performance.Secondly,the prior knowledge base of grasping objects in Home-Environment is constructed to improve sampling efficiency.A sampling method guided by object pose and grasp prior knowledge is utilized to reduce the computing cost for sample evaluation.During the process of generating samples,it adds different levels of gauss perturbation,which can effectively reduce the influence for the sampling precision caused by the error from pose measurement.A model-based method with Point Pair Feature is employed to measure the object pose efficiently.Moreover,a 3D-vision software for generating grasping prior knowledge is developed based on Open 3D,which makes it easier to build the prior knowledge base.Autodesk ReCap Photo is utilized to build the CAD model of objects and improve models development efficiency.Thirdly,to avoid the problem of data collection in traditional methods,we employ a self-supervised learning method guided by geometric analysis to prevent the long time-consuming of previous methods and improves the learning efficiency.The robotic grasping skill is learned in simulation and translated to the real world,which can effectively avoid the joints wear or breakdown of the physical robot,reduce the learning cost.For improving the development efficiency,a grasping objects dataset used in V-REP,including 160 daily objects,is presented to provide enough objects to be grasped in V-REP simulation.Finally,by comparative experiments,we demonstrate that the proposed grasping method based on fusion policy can not only adapt to grasp objects with different scenarios,but it also improves the execution efficiency and grasping precision.To test the performance of the proposed method,we build the grasping experiment platform and test the grasping performance with different objects.The results indicate that the grasping model learned in simulation can achieve high performance in the real-world with complex objects.It also prove that the proposed method has higher adaptability to complex environments.