| Generating grasp poses is a crucial component for any robot object manipulation task.However,most of the current grasping methods are mainly aimed at structured environments,and they rely heavily on prior knowledge such as known object models,contact information,and physical properties to plan grasping,resulting in relatively fixed application scenarios and grasping targets of these methods.In addition,the existing methods for grasping unknown objects mainly take 2D RGB images or 2.5D depth maps as input,while very few of them take the 3D geometric information of the target into consideration.Therefore,these methods can only constrain the grasp pose to be a three-degree-of-freedom(3-DOF)oriented rectangle which parallels to the image plane,limiting the variety of grasps.In this thesis,a 6-DOF robot grasp detection method based on point cloud features is proposed to address the challenging problem of directly obtaining the grasp pose of unknown objects in unstructured environments from the point cloud.This thesis mainly completes the research work in the following four aspects:(1)The overall framework of the grasp pose detection algorithm is designed,including three modules: grasp sampling,grasp optimization and grasp performance evaluation.All modules take 3D point clouds observed by a depth camera as input.According to the design scheme,the software and hardware platform of the grasp detection system are constructed.In order to facilitate the analysis,by simplifying the gripper model in the hardware platform to represent the 6-DOF grasp pose,the grasp pose detection problem is transformed into the establishment of the gripper coordinate system.(2)Considering that the current 6-DOF grasp sampling methods cannot generate stable grasp samples at the edge of the point cloud due to their dependence on the normal information of the object surface,this thesis proposes a grasp candidate generation method based on various geometric features of point cloud and qualitative grasping experience.Finally,a comparative experiment is designed to verify the effectiveness of this geometric heuristic sampling strategy from the sample coverage metric.The experimental results show that this sampling method can generate stable grasp candidates that are more abundant and uniformly distributed on the surface of the object.(3)Although failed grasps from grasping candidates can be initially eliminated by setting conditions,a large portion of the rejected grasps can be close to successful ones.Therefore,this thesis proposes three different grasp pose optimization methods based on the grasp performance evaluation indicators such as improved force closure and force balance to refine the original grasping samples,convert failed grasps into successful ones,and achieve high-quality sampling.Finally,the ablation experiments are used to verify that the proposed grasp optimization method can greatly improve the sample quality from the force closure rate and the robustness score.(4)In order to obtain the optimal grasp among candidates,this thesis designs an end-to-end grasp performance evaluation model based on the point cloud convolutional neural network Conv Point.This model takes the raw point cloud within the gripper closing area as input,does not need to convert the point cloud,and directly outputs the evaluation results corresponding to the performance of the grasp.Finally,the grasp with the highest score will be executed.Finally,the performance of the proposed method is measured with simulation experiments and practical test.The results show that the grasp evaluation network can capture the complex geometric structure of the contact area between the gripper and the object even if the point cloud is very sparse.What’s more,our proposed Conv GPD achieves 88.57% success rate on various commonly used objects,and generalizes well to other objects of unknown shape in unstructured environments.Code and video are available at https://github.com/quxiaochang/6-DOF-Conv GPD. |
PDF Full Text Request