A six-degree-of-freedom mechanical arm grabbing detection method in a complex scene

Abstract

The present application relates to a kind of complex scene under the mechanical arm six degrees of freedom grabbing detection method, belong to robot target object grabbing technical field.The point cloud cluster of target object and entire point cloud scene are obtained;The outline and the graspable area of point cloud are calculated;The centroid and feature vector of point cloud cluster are calculated;Pre-generated grasp pose on target object;The centroid and feature vector of point cloud cluster are used to filter grasp pose, and the retained grasp pose is the part of grasp pose that is close to the centroid of point cloud cluster and meets the requirements with the included angle of direction vector;The retained grasp pose is input into grasp pose generation network and is scored;Area point cloud analysis method scores the quality of retained grasp pose;Two scores are given different weights respectively, and final sorting is carried out, and the highest score corresponding grasp pose is output;Make the mechanical arm execute the grasp pose, and grab target object.It can be realized in complex environment to realize the stable grabbing of target object.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of robot target object grasping, and particularly relates to a mechanical arm six-degree-of-freedom grasping detection method in a complex scene, which combines regional point cloud analysis and grasping pose generation neural network on a target point cloud. BACKGROUND

[0002] With the rapid development of society, robots play an important role in industry, agriculture, seabed exploration and other fields. Among them, the grasping operation of the robot plays an important role, and the autonomy of the robot to realize the grasping function has great significance to improve the industrial production efficiency and improve the effect of human-computer interaction. Although in recent years, autonomous grasping of mechanical arms has made great progress in industrial production, but this type of industrial robot performs repeated grasping under the conditions of good lighting, fixed environment structure and known grasped objects, and the perception information ability of the surrounding environment and the target object is not high. However, it is still a great challenge for the mechanical arm to grasp new objects in an unstructured environment, mainly because the uncertainty of the new objects grasped and the surrounding scene greatly reduces the success rate of the mechanical arm grasping. Therefore, how to generate high-quality grasping poses in real time is a research hotspot in the field of robots.

[0003] With the rapid development of deep learning algorithms, data-driven mechanical arm grasping algorithms have been widely studied. This kind of method mainly uses the idea based on classification, uses CNN and other methods to learn the feature information of the target object from the two-dimensional image (2.5D depth map) or three-dimensional point cloud data obtained by the visual device, and then learns to generate an evaluation model for scoring the quality of the grasping pose, so as to output the optimal grasping pose in actual grasping. Grasping detection based on two-dimensional images (2.5D depth maps) has achieved rich results, but this method cannot calculate the interaction results between the mechanical gripper and the three-dimensional shape of the object, and has the limitation of grasping dimension. This "top-down" grasping method of the mechanical arm limits the grasping effect of the mechanical arm, making the grasping success rate of the mechanical arm in the unstructured environment lower. The grasping pose detection method based on three-dimensional point cloud can generate a 6D grasping pose, and the 6D grasping pose can calculate the three-dimensional position and three-dimensional pose angle of the mechanical gripper, so it can better grasp the target object.

[0004] Many existing point cloud-based grasp detection methods do not consider the actual distribution of objects, so the grasp poses generated in the actual multi-object environment are prone to failed grasping. This is mainly because the objects are only grasped by the single object during training, and the network cannot consider the distribution state of the target object during actual grasping. Some researchers introduce a collision detection module to filter out invalid grasps by taking the entire point cloud scene as the input of the neural network. Such networks often fail due to their large amount of computation and insufficient generalization to the scene. SUMMARY

[0005] The technical problem solved by the present application is:

[0006] In order to solve the problem that the neural network model cannot quickly analyze the regional point cloud scene during actual grasping, the present application proposes a six-degree-of-freedom grasp detection method for a robot arm in a complex scene, which combines a regional point cloud analysis method and a grasp pose generation network.

[0007] To solve the above technical problems, the technical scheme adopted by the present application is:

[0008] A six-degree-of-freedom grasp detection method for a robot arm in a complex scene, characterized by the following steps:

[0009] Step 1: Obtain the point cloud cluster of the target object and the entire point cloud scene;

[0010] Step 2: Calculate the contour and graspable region of the point cloud using the regional point cloud analysis method, and segment each individual graspable region;

[0011] Step 3: For each individual graspable region, calculate the centroid and feature vector of the point cloud cluster using PCA;

[0012] Step 4: Pre-generate grasp poses on the target object using the neural network model GPG;

[0013] Step 5: Filter out the grasp poses generated by GPG in the non-graspable region;

[0014] Step 6: Filter the grasp poses using the centroid and feature vector of the point cloud cluster, and retain the grasp poses that are close to the centroid of the point cloud cluster and have an angle with the direction vector that meets the requirements;

[0015] Step 7: Input the retained grasp poses into the grasp pose generation network PointNetGPD for scoring, and obtain the score score1;

[0016] Step 8: Score the quality of the retained grasp poses using the regional point cloud analysis method, and obtain the score score2;

[0017] Step 9: Give different weights to the scores score1 and score2 obtained by each grasping posture respectively, and perform the final sorting, output the grasping posture corresponding to the highest score; let the robot arm execute the grasping posture to grasp the target object.

[0018] Further technical solutions of the application: in step 1, the camera carried by the robot arm acquires the point cloud cluster and the whole point cloud scene of the target object, the camera is a depth camera or a binocular camera, and is fixed at the end of the robot arm to follow the movement of the robot arm or is fixed at a position outside the robot arm body.

[0019] Further technical solutions of the application: in step 2, specifically:

[0020] Suppose that P is a point cloud point in the point cloud set, the point cloud in the neighborhood of P is projected to the tangent plane of the point, and the point corresponding to P after projection is p; a series of angles are formed by connecting the surrounding points with point p as the center, forming an angle set:

[0021] Θ={θ1,θ2,…,θ n}

[0022] The maximum angle θ max =max(Θ) in the set is found, when θ max is greater than the first set threshold, it is determined that P is a boundary point, otherwise it is not a boundary point; after the outline of the point cloud cluster is found, comparison is made from bottom to top along the outline of the target point cloud cluster, and the distance between the maximum opening width of the gripper and the point cloud outline is compared to calculate the graspable region.

[0023] Further technical solutions of the application: in step 3, specifically:

[0024] Suppose that the target point cloud cluster is C, the set containing n point cloud points is {c1,c2,c3,…,c n}, and the centroid of the point cloud cluster is P cen :

[0025]

[0026] A covariance matrix of the point cloud cluster is constructed, and eigenvalues λ1, λ2, λ3 of the point cloud cluster are calculated by using the covariance matrix, and eigenvectors u1, u2, u3 corresponding to each eigenvalue are calculated; wherein, the eigenvalues are set as λ1≥λ2≥λ3, and the eigenvector corresponding to the smallest eigenvalue is set as the direction vector of the point cloud cluster.

[0027] Further technical solutions of the application: the graspable region is calculated by comparing the maximum opening width of the gripper and the distance between the point cloud outline, specifically:

[0028] When the distance between the point cloud contours is less than the maximum opening width of the gripper, the position is a graspable position; otherwise, it is an ungraspable position.

[0029] Further technical solutions of the present application: the grasp pose in step 4 only considers the target object and does not consider the supporting desktop and surrounding objects; represented as g i =(p,r), wherein p=(x,y,z) is a position vector, and r=(r x ,r y ,r z ) is an attitude angle vector.

[0030] Further technical solutions of the present application: step 6 uses the centroid and feature vector of the point cloud cluster to filter the grasp pose, specifically:

[0031] Calculate the distance between the position vector of the grasp pose and the centroid of the point cloud cluster, and when the distance is less than a second set threshold, it is retained;

[0032] Calculate the included angle between the approach vector of the grasp pose and the direction vector of the point cloud cluster, and when the included angle is less than a third set threshold, it is considered that the grasp pose will approach the target object in a perpendicular direction; when the direction vector and the approach vector are in the same direction, that is, the grasp pose approaches the object from below the supporting desktop, a collision will occur; therefore, the approach vector is negated, and the included angle between the approach vector of the grasp pose and the direction vector of the point cloud cluster is calculated again.

[0033] Further technical solutions of the present application: step 7 inputs the retained grasp pose into the grasp pose generation network PointNetGPD for scoring, including:

[0034] Offline training of the PointNetGPD model: generating and scoring the grasp pose of the target object; inputting the point cloud inside the grasp pose and the corresponding score into the neural network model PointNet for classification learning to generate a classification model of the grasp pose;

[0035] The trained PointNetGPD model is used to score the input grasp pose.

[0036] Further technical solutions of the present application: step 8 uses the regional point cloud analysis method to score the quality of the retained grasp pose, and the scoring formula is as follows:

[0037]

[0038] L d1 (g i ,O)=||p i -p O ||

[0039]

[0040]

[0041] wherein, L d1 represents the distance between the position vector of the grasping posture and the cluster center of the point cloud, the greater the distance, the smaller the corresponding score; O near represents the average distance between the position vector of the grasping posture and the surrounding point cloud set obtained by taking the cluster center of the target point cloud as the center of the sphere and a certain size as the radius, the smaller the average distance, the closer the grasping posture to the middle position, and the less likely to collide with the surrounding objects; L θ is the included angle between the approach vector and the direction vector of the grasping posture, the larger the included angle, the lower the corresponding score; λ1', λ2', λ3' are balance coefficients.

[0042] The beneficial effects of the present application are:

[0043] The six-degree-of-freedom grasping detection method for a robot arm in a complex scene provided by the present application can realize stable grasping of a target object in a complex environment. The regional point cloud analysis method can analyze the target point cloud cluster and the surrounding point cloud scene in real time, so as to filter out the grasping posture that collides with the surrounding environment. Combined with a neural network, the grasping posture is scored together to generate the optimal grasping posture. Therefore, the method of the present application can help the robot arm complete the analysis of the target scene in a complex scene, and to a great extent, help the robot arm complete the target grasping task in a non-structured environment. BRIEF DESCRIPTION OF DRAWINGS

[0044] The accompanying drawings are included to provide a further understanding of the embodiments, and are incorporated in and constitute a part of this specification, illustrate embodiments of the application, and together with the description serve to explain the principles of the application, and should not be considered limiting of the present application in any way. In the drawings:

[0045] Figure 1 is the overall flowchart of the method of the present application.

[0046] Figure 2 is the feature vector and cluster center generated by the PCA algorithm for point cloud cluster calculation.

[0047] Figure 3 is the angle threshold-based method, and the finally calculated point cloud contour.

[0048] Figure 4 is a schematic diagram of grasping posture screening.

[0049] Figure 5 is the entire grasping posture generation and selection process.

[0050] Figure 6 Flow chart for the mechanical arm to grab the actual object in the present application. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0052] As shown in Figure 1 The present application provides a six-degree-of-freedom mechanical arm grasping detection method in a complex scene, specifically a method based on regional point cloud analysis and grasp pose generation network combination, including the following steps:

[0053] Step 1, the camera carried by the mechanical arm acquires the point cloud cluster of the target object and the entire point cloud scene, the camera is a depth camera or a binocular camera, and the fixing method has two kinds. It can be fixed at the end of the mechanical arm to follow the movement of the mechanical arm, or it can be fixed at a position outside the mechanical arm body.

[0054] In this embodiment, the point cloud cluster of the target object and the entire point cloud scene are obtained by a depth camera;

[0055] Step 2, the regional point cloud analysis method calculates the contour and graspable region of the point cloud, so as to segment each individual graspable region;

[0056] Step 3, for each individual graspable region, the PCA (Principal Component Analysis method) is used to calculate the centroid and feature vector of the point cloud cluster;

[0057] Step 4, a large number of grasping poses are generated for the target point cloud cluster;

[0058] In this embodiment, a large number of grasping poses are generated in advance on the target object by a neural network model (Grasp Pose Generation, GPG), and these grasping poses only consider the target object and do not consider the supporting desktop and surrounding objects;

[0059] Step 5, the regional point cloud analysis method filters the grasping poses generated by GPG, and retains the collision-free and high-quality grasping poses with the desktop;

[0060] Including the regional point cloud analysis method filters out the grasping poses generated by GPG in the ungraspable region according to the calculated graspable region;

[0061] The step 2 is calculated in each graspable region, the grasp pose is filtered by using the centroid and feature vector of the point cloud cluster, and the retained grasp pose is the part of the grasp pose close to the centroid of the point cloud cluster and the smaller angle with the direction vector;

[0062] Step 6, the retained grasp pose is input into the trained neural network model PointNetGPD for scoring, so that the score score1 corresponding to the grasp pose is obtained;

[0063] Step 7, the region point cloud analysis method also scores the grasp pose, and outputs the score score2 corresponding to the grasp pose;

[0064] Step 8, the scores corresponding to the grasp poses obtained in steps 6 and 7 are combined by weighting, and the grasp pose with the highest score is selected;

[0065] Step 9, the grasp pose is converted in coordinates, the grasp pose in the mechanical arm coordinate system is obtained, and then the mechanical arm is planned to execute grasping.

[0066] In actual operation, the devices required by the present application include a depth camera, a six-degree-of-freedom mechanical arm, a notebook computer and communication equipment therebetween.

[0067] The above steps are specifically as follows:

[0068] Step 1, the depth camera is mounted at the end of the mechanical arm, and the point cloud cluster of the target object and the whole point cloud scene are obtained after initialization.

[0069] Step 2, based on the point cloud cluster of the target object, the contour points of the point cloud cluster are obtained by using the angle threshold method. As shown in Figure 3 (a), when the target point P is the contour point of the target object, the angle formed by the point and the surrounding points has a large value; when the target point P is the point cloud inside the point cloud cluster, the angle formed by the point and the surrounding points is generally small. Therefore, the contour of the point cloud cluster can be obtained by the following steps:

[0070] (1) let the current point be P, project P and the surrounding point cloud onto the tangent plane of the point, and let the projected point be p.

[0071] (2) take p as the center point, connect the surrounding points with the point two by two to form a series of angles, and form a new set Θ={θ1,θ2,…,θ n}.

[0072] (3) find the maximum angle θ max =max(Θ) in the set.

[0073] (4) by setting a threshold value ο, when θ maxIf the distance is greater than the threshold, it is determined that P is a boundary point, otherwise it is not a boundary point.

[0074] As shown in Figure 3 (b), when the target object is a hammer, the algorithm can well obtain the contour points of the point cloud cluster. After obtaining the contour points, calculation is performed from bottom to top along the axis of the target object, and the distance between contour points at a certain position is calculated. When the distance is less than the maximum opening value of the gripper, it is considered that the point is a graspable point. The graspable region of the point cloud cluster is obtained according to this principle. When the shape of the target object is relatively complex, the graspable region of the point cloud is generally a relatively simple region such as a cube or a cylinder. When the target object is relatively simple, the entire point cloud cluster is a graspable region or a non-graspable region, so subsequent calculation can be directly performed.

[0075] Step 3, PCA calculation is performed on the point cloud cluster of the target object or the segmented graspable region, so that the centroid and the characteristic vector of the point cloud part can be obtained, as shown in Figure 2 .

[0076] Suppose the target point cloud cluster is C, and the set of n point clouds is {c1, c2, c3, …, cn}. Therefore, the centroid of the point cloud cluster is P n : cen

[0077]

[0078] Then the covariance matrix of the point cloud cluster is constructed:

[0079]

[0080] The eigenvalues λ1, λ2, λ3 of the point cloud cluster are obtained by using the matrix, and the characteristic vectors u1, u2, u3 corresponding to each eigenvalue are calculated. Among them, the size of the set characteristic is λ1≥λ2≥λ3. The characteristic vector corresponding to the smallest eigenvalue is set as the direction vector of the point cloud cluster.

[0081] Step 4, the GPG algorithm generates a large number of grasping poses for the target point cloud, and the number of pre-generated grasping poses is set to 100. Each grasping pose is represented as g i =(p, r), wherein p is a position vector and r is an attitude angle vector. Therefore, p=(x, y, z) and r=(r x , r y , r z ). As shown in Figure 5 (a), most of these grasping poses are prone to collision with the supporting table, and many grasping qualities are not very high, so directly inputting these grasping poses into the neural network for scoring is not only time-consuming and inefficient, but also may fail to grasp. ​

[0082] Step 5, the GPG generated grasp poses on the non-graspable area due to lack of generalization are filtered out. Then the graspable area is segmented out, which is generally a common cuboid or cylinder, etc. As shown in Figure 4 The region point cloud analysis method mainly filters and screens the grasp poses generated by the GPG from two aspects:

[0083] (1) For the grasp pose g i = (p i , r i ), first, the position is screened. The centroid of the point cloud cluster can be obtained as P cen = (x, y, z), and for the pre-generated grasp pose g i , i ∈ {1, 2, 3, …, n}. The distance d i between the position vector of each grasp pose and the centroid of the point cloud cluster can be obtained. When the distance is less than the set threshold T h , it is considered that the grasp pose is closer to the centroid of the point cloud cluster (the centroid of the object). The calculation formula is as follows:

[0084]

[0085] d ci ≤ T h

[0086] (2) Secondly, the screenable grasp pose is screened, as shown in Figure 4 (b), the angle θ i between the approach vector n a of the grasp pose g i and the direction vector u3 of the point cloud is calculated.

[0087]

[0088] θ i ≤ T θ

[0089] When the angle θ i is less than the set threshold T θ , it is considered that the grasp pose approaches the target object at a near-perpendicular angle, which has a higher success rate of grasping. At the same time, if the approach vector of the generated grasp pose and the direction vector of the object u3 are in the same direction, the mechanical gripper will approach the target object from below the supporting table, and therefore will collide with the supporting table. Therefore, the approach vector needs to be inverted, so as to better grasp the target object.

[0090] After the above two aspects of screening, high-quality and collision-free grasp poses can be obtained, as shown in Figure 5 (b).

[0091] Step 6, filter out the large number of grasp poses from step 5, and transport the remaining grasp poses to the offline trained PointNetGPD model to score the grasp poses and obtain the score score1. During offline training, first generate grasp poses for the target object and score them. Then input the point cloud inside the grasp pose and the corresponding score into the neural network model PointNet for classification learning, and finally generate a classification model for the grasp pose. This model can score the input grasp pose and output the optimal grasp pose.

[0092] Step 7, the regional point cloud analysis method scores the retained grasp poses, mainly based on the analysis of the actual object distribution and the surrounding object situation, to obtain the score score2. The specific calculation formula is as follows:

[0093]

[0094] L d1 (g i ,O)=||p i -p O ||

[0095]

[0096]

[0097] where λ1', λ2', λ3' are balance coefficients. O represents the target object point cloud cluster, then represents the distance between the position of the grasp pose and the centroid of the point cloud cluster, and the larger the distance, the lower the corresponding score. O near represents the part of the point cloud of the surrounding object searched with the centroid of the target point cloud cluster as the center and a certain distance as the radius. represents the average distance between the position of the grasp pose and the surrounding point cloud, and the smaller the distance, the closer the grasp pose to the middle position, and the higher the corresponding score. L θ represents the angle between the direction vector of the grasp pose and the target object, and the smaller the angle, the higher the corresponding score.

[0098] Step 8, the scores obtained in steps 6 and 7 are weighted and averaged:

[0099] S=αscore1+β·score2

[0100] where α and β are weighting coefficients, and finally the filtered grasp pose is obtained, and the grasp pose with the highest score is selected. The final grasp pose is shown in Figure 5 (c).

[0101] Step 9, after the selection of the grasping posture is completed, the coordinates and posture of the grasping posture are converted through coordinate conversion, and the posture of the grasping posture in the polar coordinate system of the robot arm is obtained. The robot arm moves to the target position through inverse kinematics solution, and completes the grasping of the target object. Figure 6 The flow of the robot arm grasping the target object is shown, Figure 6 (a) initialization of the state of the robot arm, Figure 6 (b) approaching the target object, Figure 6 (c) reaching the target posture, Figure 6 (d) closing the mechanical gripper to complete the grasping task.

[0102] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application.

Claims

1. A method for detecting a six-degree-of-freedom grasping of a robot arm in a complex scene, characterized in that, The steps are as follows: Step 1: Obtain the point cloud cluster of the target object and the entire point cloud scene; Step 2: Calculate the contour and graspable area of the point cloud by the regional point cloud analysis method, and segment each individual graspable area; Specifically: Assume a point P For a point cloud point in a point cloud set, project P the point cloud in its surrounding neighborhood to the tangent plane of the point, and after the projection P the corresponding point is p ; With the point p as the center, the surrounding points are connected with the point p one by one to form a series of angles, forming an angle set: Finding the maximum included angle in the set When is greater than the first set threshold, it is determined P that the point is a boundary point, otherwise it is not a boundary point; after the contour of the point cloud cluster is obtained, comparison is made from bottom to top along the contour of the target point cloud cluster, and the distance between the maximum opening width of the gripper and the point cloud contour is calculated to obtain the graspable region; Step 3: For each individual graspable area, calculate the centroid and feature vector of the point cloud cluster using the principal component analysis method PCA; Specifically: Let the target point cloud cluster be C , the set of n point cloud points is , and the centroid of the point cloud cluster is calculated as : A covariance matrix of the point cloud cluster is constructed, and an eigenvalue of the point cloud cluster is solved by using the covariance matrix , , , a feature vector corresponding to each eigenvalue is calculated , , ; wherein the size of the set feature is , and the feature vector corresponding to the smallest eigenvalue is set as the direction vector of the point cloud cluster; Step 4: Generate a pre-generated grasp pose on the target object by the neural network model GPG; Step 5: According to the graspable area, the regional point cloud analysis method filters out the grasp pose generated by GPG on the non-graspable area; Step 6: Filter the grasp pose using the centroid and feature vector of the point cloud cluster, and the retained grasp pose is the part of the grasp pose that is close to the centroid of the point cloud cluster and has an angle with the direction vector that meets the requirements; Specifically: Calculate the distance between the position vector of the grasp pose and the centroid of the point cloud cluster, and when the distance is less than a second set threshold, it is retained; Calculate the angle between the approach vector of the grasp pose and the direction vector of the point cloud cluster, and when the angle is less than a third set threshold, it is considered that the grasp pose will approach the target object in a perpendicular direction; When the direction vector and the approach vector are in the same direction, i.e. the grasp pose approaches the object from under the supporting table, a collision will occur; Therefore, the approach vector is negated, and the angle between the approach vector of the grasp pose and the direction vector of the point cloud cluster is calculated again. Step 7: input the reserved grasp pose into the grasp pose generation network PointNetGPD for scoring to obtain a score ; Step 8: Score the quality of the retained grasp pose from the grasp, with a score of ; the scoring formula is as follows: wherein, represents the distance between the position vector of the grasp pose and the center of the point cloud cluster, the larger the distance, the smaller the corresponding score; represents the point cloud set of the surrounding object obtained by taking the center of the target point cloud cluster as the center of the sphere and a certain size as the radius of the sphere, represents the average distance between the position vector of the grasp pose and the surrounding point cloud set, the smaller the average distance, the closer the grasp pose to the intermediate position, and the less likely to collide with the surrounding object; is the angle between the approach vector and the direction vector of the grasp pose, the larger the angle, the lower the corresponding score; is a balance coefficient, represents the target object point cloud cluster, is the approach vector; Step 9: score each grasp pose and are given different weights, respectively, and a final ranking is performed, outputting the grasp pose corresponding to the highest score; the robot arm is instructed to execute this grasp pose to grasp the target object.

2. The method of claim 1, wherein, In step 1, the camera carried by the mechanical arm obtains the point cloud cluster and the entire point cloud scene of the target object, the camera is a depth camera or a binocular camera, and is fixed on the end of the mechanical arm to follow the movement of the mechanical arm or is fixed on the body of the mechanical arm.

3. The method of claim 1, wherein, The graspable area is calculated by comparing the maximum opening width of the gripper and the distance between the point cloud contours, specifically: When the distance between the point cloud contours is less than the maximum opening width of the gripper, the position is a graspable position; Otherwise, it is a non-graspable position.

4. The method of claim 1, wherein, The grasp pose described in step 4 only considers the target object and does not consider the supporting table and surrounding objects; the grasp pose is represented as where, is a position vector, is an angle vector.

5. The method of claim 1, wherein, Step 7 inputs the retained grasp pose into the grasp pose generation network PointNetGPD for scoring, including: Offline training of PointNetGPD model: generate and score the grasp pose of the target object; input the point cloud inside the grasp pose and the corresponding score into the neural network model PointNet for classification learning to generate a classification model of the grasp pose; Use the trained PointNetGPD model to score the input grasp pose.

Images