A method for predicting grasping position of parallel grippers of robotic arms based on improved PointNet++

Through the improved PointNet++ model, the gripping position prediction of the parallel jaws of the robot arm is solved, and the problem of large calculation volume and insufficient crawling success rate is achieved, more efficient three-dimensional point cloud data processing and higher crawling success rate are improved. The operation flexibility and accuracy of the robotic arm.

CN118596156BActive Publication Date: 2024-11-05HUAQIAO UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411045466.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-05
Estimated Expiration
2044-08-01

AI Technical Summary

Technical Problem

In the prior art, the parallel jaws of the robot arm have problems such as large calculation and unknown practical application effects when dealing with complex scenarios and diverse objects. Especially when processing large-scale point cloud data, the calculation and memory overhead are significant, and the existing depth is The learning method is insufficient in predicting success rate.

Method used

Using the improved PointNet++ model, point cloud data is featured through point set abstraction layer, multi-layer perceptron, channel attention mechanism and feature propagation layer, and combined with the furthest point sampling method and distance-weighted interpolation, the predicted position parameters and Grab the probability of success and optimize the model training using the combined loss function.

Benefits of technology

It improves the task flexibility and accuracy of the robotic arm in complex environments, accurately predicts the position and posture of objects, and at the same time improves the crawling success rate and operation efficiency, simplifies the model training process, and enhances the generalization ability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118596156B_ABST
    Figure CN118596156B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for predicting the grasping position of a parallel gripper of a robotic arm based on improved PointNet++, which relates to the fields of computer vision and robotics technology, and includes the following steps: constructing a position prediction model based on improved PointNet++ and training it using a training set; inputting the point cloud data of the object to be grasped into the trained position prediction model to obtain predicted position parameters. The position prediction model includes a number of feature extraction modules connected in sequence and a fully connected layer, wherein the feature extraction module uses PointNet++ with added channel attention to perform multiple feature extractions on the input point cloud data; the fully connected layer receives the global features output by the last feature extraction module and outputs the predicted position parameters. The present invention combines 3D point cloud technology and deep learning, which not only improves the accuracy of the robotic arm grasping, but also enables more stable and efficient grasping operations in complex environments.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of computer vision and robotics technology, and in particular to a method for predicting grasping positions of parallel grippers of a robotic arm based on improved PointNet++. Background Art

[0002] With the development of industrial automation and intelligent manufacturing, robotic arms are increasingly used in production lines. Traditional robotic arm target detection mainly relies on two-dimensional image processing technology, which often has limitations in dealing with complex scenes or under different lighting conditions. In addition, traditional methods often fail to meet the needs of high-precision operations in terms of spatial positioning accuracy of target objects. Therefore, developing a robotic arm target grasping method that can accurately locate objects in three-dimensional space is of great significance for improving the operating efficiency and flexibility of robotic arms. Parallel grippers are widely used in robotic arms due to their simple structure and easy control. They are used to complete tasks ranging from simple handling to complex assembly. Although parallel gripper robotic arms have been automated on many production lines, they still face a series of challenges when performing grasping tasks, especially when handling objects in diverse and unknown environments.

[0003] In the prior art, although some solutions based on 3D vision have been proposed, these solutions often require complex hardware support and have insufficient processing speed, making it difficult to meet the requirements of real-time operation. In addition, although the existing deep learning methods have made significant progress in the field of image recognition, their application in 3D point cloud data processing still faces challenges such as limited training data sets and insufficient model generalization ability. The Chinese invention application with publication number CN114211490A discloses a robot gripper posture prediction method based on the Transformer model, which applies the Transformer model to the field of robot gripper posture prediction technology. By calculating the global and local features of the point cloud and obtaining the point cloud features through a multi-head self-attention module, the prediction performance and efficiency of the robot gripper posture information are improved, and the robustness is enhanced; however, the Transformer model used in this method has a high computational complexity, especially when processing large-scale point cloud data, the computational and memory overheads are significantly increased; and this method mainly predicts the posture information of the gripper (such as gripper state, center point, rotation direction, width and depth), but does not directly consider the important indicator of grasping success rate, which may affect the grasping effect in practical applications. Summary of the invention

[0004] The purpose of the present invention is to solve the problem in the prior art that the amount of calculation is large and the actual application effect is unknown.

[0005] The technical solution adopted by the present invention to solve the technical problem is: to provide a method for predicting the grasping position of a parallel gripper of a robotic arm based on improved PointNet++, comprising the following steps:

[0006] Collect the point cloud data of the target object as a training set. The point cloud data includes three-dimensional space coordinates and normal vectors.

[0007] Build a position prediction model based on improved PointNet++, receive point cloud data as the original point set, and output the predicted position parameters suitable for the grasping points of the parallel gripper of the robot arm;

[0008] Using the training set to train a location prediction model;

[0009] Input the point cloud data of the object to be grasped into the trained position prediction model to obtain the predicted position parameters;

[0010] The position prediction model includes several feature extraction modules connected in sequence and a fully connected layer. The feature extraction module adopts PointNet++ with added channel attention, including a point set abstraction layer, a multi-layer perceptron, a channel attention mechanism and a feature propagation layer; the feature extraction module implements the farthest point sampling method to select key points for the input point cloud data or the global features output by the previous feature extraction module, gathers local neighborhood points based on a preset radius, and then uses a multi-layer perceptron to independently extract local point features, and finally uses channel attention to output local features based on local point features to the feature propagation layer; the feature propagation layer uses distance weighted interpolation and residual connection to propagate local features back to the original point set to obtain global features, and outputs them to the next feature extraction module or the fully connected layer; the fully connected layer receives the global features output by the last feature extraction module, and outputs predicted position parameters and the probability of successful grasping.

[0011] Preferably, the number of the feature extraction modules is at least three, and the point set abstraction layers of different feature extraction modules use query spheres with different radii and different numbers of sampling points to extract local features of different scales.

[0012] Preferably, the implementation of the farthest point sampling method to select key points includes the following steps:

[0013] Choose an initial point p 0 As the first point of the sampling point set S, S = {p 0};

[0014] Calculate the shortest Euclidean distance d(p,S) from all unselected points p to all points in the current sampling point set S;

[0015] Select the point with the largest shortest Euclidean distance and add it to S. Repeat this step until the sampling set S reaches a predetermined size k or meets other termination conditions; where ∥ps∥ represents the Euclidean distance between points p and s, p next Indicates the next point.

[0016] Preferably, the other termination conditions include a convergence condition, which means that the coverage effect of the newly added points on the existing sampling set during the sampling process no longer significantly improves, as manifested by the increase in the shortest distance of the newly added points being less than a predetermined threshold.

[0017] Preferably, p 0 Use random selection or heuristic selection.

[0018] Preferably, the method of using channel attention to output local features to a feature propagation layer based on local point features comprises the following steps:

[0019] Perform global average pooling on the features of each channel, expressed as:

[0020]

[0021] Among them, x i,j,c The value of the cth channel at the (i, j) position of the feature map X representing the input channel attention, H and W represent the height and width of the feature map X, respectively, C represents the number of channels of the feature map X, and z c is the global average value of the c-th channel;

[0022] The output of the global average pooling is further learned and processed through a small fully connected network, and the channel weight is expressed as:

[0023] s1=σ(W 2 δ(W 1 z));

[0024] Where z represents the output of global average pooling, which consists of the global average of all channels; σ is the sigmoid activation function, which is used to generate weights in the range [0,1]; δ is the ReLU activation function; W 1 The size is The dimension reduction weight matrix, W 2 The size is The dimension-upgraded weight matrix of , where r is the scaling ratio; s1 is the generated channel weight set, containing the channel weight of each channel;

[0025] Multiply each channel of the original feature map X by the learned corresponding channel weight to complete the channel recalibration:

[0026]

[0027] in, Represents xi,j,c The recalibrated feature map, s c Represents the channel weight of the c-th channel.

[0028] Preferably, when training the location prediction model, a combined loss function including mean square error loss MSE and cross entropy loss Cross-Entropy is used, which is expressed as:

[0029] Combined Loss=α×MSE+β×Cross-Entropy;

[0030]

[0031]

[0032] Among them, α and β are weight factors used to balance the impact of the two loss functions; is the i-th predicted position parameter, y1 i is the i-th true location parameter, n is the total number of samples; is the probability of successful capture of the i-th sample, y2 i is the true captured label of the i-th sample, y2 i =0 or y2 i =1.

[0033] Preferably, the predicted position parameter output by the fully connected layer is a six-dimensional vector, and the six dimensions include:

[0034] The three-dimensional coordinates of the grasping position determine the exact point where the robot arm gripper lands in physical space, which serves as the geometric center of the grasping operation;

[0035] The three-dimensional rotation angle of the grasping posture guides the parallel gripper to adjust to the direction suitable for grasping the target object.

[0036] The present invention has the following beneficial effects:

[0037] (1) The present invention is directed to a visual perception system for parallel gripper grasping by a robotic arm. By using advanced processing of 3D point cloud data and deep learning models, the system can accurately grasp the three-dimensional spatial features of the target object, thereby improving the flexibility and accuracy of the robotic arm in performing tasks in complex environments.

[0038] (2) The present invention can more accurately extract important features about the position and shape of objects from complex three-dimensional point cloud data through multi-layer point set abstraction and feature propagation modules;

[0039] (3) The present invention not only accurately predicts the grasping position and posture, but also predicts the probability of successful grasping, so that the robot arm can avoid grasping attempts with high uncertainty in actual operation, thereby improving the overall operation success rate and efficiency;

[0040] (4) The present invention optimizes the grasping parameter prediction by using a combined loss function, allowing the model to simultaneously consider the position, posture accuracy and grasping success rate during the training phase, thereby simplifying the model training process and enhancing the model's generalization ability for actual grasping tasks.

[0041] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A method step diagram of an embodiment of the present invention;

[0043] Figure 2 is a structural diagram of a location prediction model according to an embodiment of the present invention, wherein: Figure 2 (a) is the overall schematic diagram of the location prediction model. Figure 2 (b) is a schematic diagram of the feature extraction module;

[0044] Figure 3 This is a structural diagram of the channel attention mechanism of an embodiment of the present invention;

[0045] Figure 4 The figure is a flow chart of grasping using the position prediction model trained in the embodiment of the present invention. DETAILED DESCRIPTION

[0046] See also Figure 1 FIG. 1 is a diagram showing steps of a method according to an embodiment of the present invention, comprising the following steps:

[0047] S101, using a depth camera to collect point cloud data of a target object as a training set, where the point cloud data includes three-dimensional space coordinates and normal vectors;

[0048] S102, constructing a position prediction model based on the improved PointNet++, receiving point cloud data as an original point set, and outputting predicted position parameters suitable for grasping points of the parallel gripper of the robot arm;

[0049] S103, training a location prediction model using the training set;

[0050] S104, inputting the point cloud data of the object to be grasped into the trained position prediction model to obtain predicted position parameters;

[0051] For details, see Figure 2 As shown, it is a position prediction model of an embodiment of the present invention, wherein: Figure 2(a) is a schematic diagram of the overall location prediction model, including several feature extraction modules connected in sequence and a fully connected layer; Figure 2 (b) is a schematic diagram of the feature extraction module, including a point set abstraction layer, a multi-layer perceptron, a channel attention mechanism (SENet, Squeeze-and-Excitation Net) and a feature propagation layer; the feature extraction module uses PointNet++ with added channel attention, and selects key points by using the farthest point sampling method on the input point cloud data or the global features output by the previous feature extraction module through the point set abstraction layer, gathers local neighborhood points based on a preset radius, and then uses a multi-layer perceptron to independently extract local point features. Finally, the channel attention is used to enhance the model's attention to local point features, and the local features are output to the feature propagation layer; the feature propagation layer uses distance weighted interpolation and residual connections to propagate local features back to the original point set to obtain global features, which are output to the next feature extraction module or the fully connected layer; the fully connected layer receives the global features output by the last feature extraction module and outputs the predicted position parameters.

[0052] Specifically, the embodiment of the present invention adopts three feature extraction modules, and the point set abstraction layers of different feature extraction modules adopt query spheres with different radii and different numbers of sampling points to extract local features of different scales.

[0053] Specifically, the point set abstraction layer uses subsets sampled from the three-dimensional point cloud to construct local point group features; each point set abstraction layer extracts local features of different scales by applying query balls of different radii and different numbers of sampling points; these layers generate feature descriptions of each point by aggregating point features within the local neighborhood and using a variety of convolutional neural networks and nonlinear activation functions; the feature propagation layer is mainly used to propagate and restore feature information in the point set after multi-scale abstraction, allowing the model to transform from local features to global features; the feature propagation layer uses upsampling and interpolation methods, combined with the features passed down from the point set abstraction layer, to reconstruct detailed features of higher resolution point sets, ensuring that the model can integrate and utilize deep and shallow features at different levels.

[0054] Specifically, both the point set abstraction layer and the feature propagation layer are equipped with specially designed neural network structures, including multiple convolutional layers, batch normalization layers, and ReLU activation layers, which work together to optimize the feature extraction and transmission process.

[0055] Specifically, the farthest point sampling method comprises the following steps:

[0056] First, choose an initial point p 0 As the first point of the sampling point set S, usually p 0 It can be randomly selected or based on some heuristic selection, S = {p 0};

[0057] Secondly, at each step, calculate the shortest Euclidean distance d(p,S) from all unselected points p to all points in the current sampling point set S, and then select the point with the largest distance to add to S, expressed as:

[0058]

[0059] Among them, ∥ps∥ represents the Euclidean distance between points p and s, p next Indicates the next point;

[0060] Finally, repeat the previous step until the sampling set S reaches the predetermined size k, or meets other termination conditions, such as the convergence condition: if the coverage effect of the newly added points on the existing sampling set during the sampling process is no longer significantly improved (that is, the increase in the shortest distance of the newly added points is less than a predetermined threshold), the sampling process will be terminated early.

[0061] The farthest point sampling is iterated by selecting the point with the largest farthest distance from all points in the current sampling set S each time, ensuring that the newly added points can cover the unselected spatial area as much as possible.

[0062] Specifically, the channel attention mechanism allows the model to perform adaptive feature recalibration, which enhances the model's sensitivity to useful features while suppressing less important information, ultimately improving overall performance. Figure 3 As shown, the channel attention operation includes the following steps:

[0063] First, the features of each channel are globally averaged pooled to compress the spatial dimension of each channel into a real number; its purpose is to extract global spatial features and generate a descriptive statistic for each channel; given a feature map X of size H×W×C, where H and W are height and width respectively, and C is the number of channels, the output z of global average pooling can be expressed as follows:

[0064]

[0065] Among them, x i,j,c is the value of the cth channel of the feature map X at position (i, j), z c is the global average value of the c-th channel;

[0066] Secondly, the SE block further learns and processes the output of the previous step through a small fully connected network. This network usually consists of two layers. The first layer is used for dimensionality reduction (reducing parameters and complexity), followed by ReLU activation; the second layer is used to restore the dimension and use the sigmoid activation function to generate the weight of each channel. This can be expressed as:

[0067] s1=σ(W 2δ(W 1 z));

[0068] Among them, σ is the sigmoid activation function, which is used to generate weights in the range [0,1]; δ is the ReLU activation function; W 1 The size is The dimension reduction weight matrix, W 2 The size is The dimension-upgraded weight matrix of , where r is the scaling ratio; s1 is the generated channel weight;

[0069] Finally, each channel of the original feature map X is multiplied by the learned corresponding channel weight to complete the channel recalibration:

[0070]

[0071] in, is the recalibrated feature map.

[0072] Specifically, the grasping parameter prediction layer outputs a six-dimensional vector specifically suitable for the parallel gripper of the robot arm, including three-dimensional coordinates and three-dimensional rotation angles, and is particularly used to describe the grasping position and grasping posture of the parallel gripper. The parallel gripper usually includes two symmetrical grasping fingers that can move in parallel; this structure allows the gripper to exert direct and uniform force on the object, thereby grasping the object stably; because the two grasping surfaces remain symmetrical, the parallel gripper can maintain the stability of the object when grasping. Therefore, the grasping parameter prediction layer uses the high-dimensional features obtained by the previous multi-layer feature abstraction and feature propagation modules, processes these features through a series of fully connected layers, and finally outputs the grasping parameters of the six-dimensional vector, including:

[0073] The 3D coordinates of the grasping position, which determine the exact location in 3D space where the robot gripper should grasp the target object; these coordinates are predicted from the input point cloud data by a deep learning model, which can predict the exact point in physical space where the parallel gripper of the robot arm will land; this is the geometric center of the grasping operation and is usually located at the best position on the surface of the object;

[0074] The three-dimensional rotation angle of the grasping posture, which includes the rotation angles around the x-axis, y-axis and z-axis, which guide the parallel gripper to adjust to the direction suitable for grasping the target object; this includes the rotation angle of the gripper relative to the grasping point, ensuring that the two grasping surfaces of the gripper are parallel to the surface of the object when the gripper contacts the object; by precisely controlling the rotation angle of the gripper, stability can be ensured during the grasping process, especially when grasping objects with complex shapes or specific positions.

[0075] Specifically, the combined loss function includes a mean square error loss for optimizing position and posture prediction and a cross entropy loss for evaluating the probability of successful grasping.

[0076] The mean square error loss (MSE) is used to optimize the prediction of the grasping position and posture of the robot arm; this loss function calculates the difference between the grasping position and posture (including 3D coordinates and rotation angles) predicted by the model and the true label; by minimizing this difference, the loss function guides the model to adjust its parameters during training to improve the precision and accuracy of the prediction; the mean square error loss (MSE) formula is as follows:

[0077]

[0078] in, is the i-th grasping parameter (including position and posture) predicted by the model, y i is the corresponding true parameter, and n is the total number of samples.

[0079] Cross entropy loss is used to evaluate the probability of whether the grasping behavior predicted by the model is successful. This loss function processes the consistency between the probability distribution of the model output and the actual grasping result, helping the model to learn and predict more accurately in the classification task (whether the grasping is successful or not). The cross entropy loss formula is as follows:

[0080]

[0081] where y i is the true label (0 or 1) of the i-th sample, is the probability of successful grasping predicted by the model, and m is the number of samples.

[0082] By combining the mean square error loss and the cross entropy loss, the loss function of the present invention not only focuses on the single prediction accuracy, but also covers the overall evaluation of the grasping success rate; this combination provides a comprehensive optimization path, prompting the model to achieve a balance between the spatial accuracy of the grasping position and the success probability of the grasping behavior; the formula of the combined loss function is as follows:

[0083] Combined Loss=α×MSE+β×Cross-Entropy;

[0084] Among them, α and β are weight factors, which are used to balance the influence of the two loss functions and are adjusted according to different application scenarios; this combination ensures that the model can not only accurately predict the grasping position and posture of the parallel gripper, but also effectively predict the success probability of grasping.

[0085] Specifically, when grasping using the position prediction model trained by the embodiment of the present invention, see Figure 4 As shown, it can be divided into:

[0086] The data acquisition module 101 receives point cloud data of the target object collected from the depth camera, where each point contains three-dimensional space coordinates and a normal vector, performs denoising, filtering and downsampling on the point cloud data, and then inputs the data into the visual perception module;

[0087] The visual perception module 102 uses the position prediction model of the embodiment of the present invention to extract features, thereby predicting grasping parameters, generating grasping parameters specifically used to predict the parallel gripper of the robot arm, and inputting the grasping parameters into the control module;

[0088] The control module 103, the robot arm controller receives the grasping parameters, performs hand-eye calibration, and finally enables the robot arm to grasp the target object.

[0089] The present invention can significantly improve the accuracy and efficiency of the parallel gripper of the robot arm when performing complex grasping tasks, improve operational flexibility, enhance the ability to adapt to different grasping tasks, and reduce the operational error rate.

[0090] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A robot arm parallel gripper grasping position prediction method based on improved PointNet++, characterized in that: The following steps are involved: Collect the point cloud data of the target object as a training set. The point cloud data includes three-dimensional space coordinates and normal vectors. Build a position prediction model based on improved PointNet++, receive point cloud data as the original point set, and output the predicted position parameters suitable for the grasping points of the parallel gripper of the robot arm; Using the training set to train a location prediction model; Input the point cloud data of the object to be grasped into the trained position prediction model to obtain the predicted position parameters; The position prediction model includes several feature extraction modules connected in sequence and a fully connected layer. The feature extraction module adopts PointNet++ with added channel attention, including a point set abstraction layer, a multi-layer perceptron, a channel attention mechanism and a feature propagation layer. The feature extraction module selects key points by using the farthest point sampling method for the input point cloud data or the global features output by the previous feature extraction module, gathers local neighborhood points based on a preset radius, and then uses a multi-layer perceptron to independently extract local point features. Finally, the channel attention is used to output local features to the feature propagation layer based on local point features. The feature propagation layer uses distance weighted interpolation and residual connection to propagate local features back to the original point set to obtain global features, which are then output to the next feature extraction module or fully connected layer. The fully connected layer receives the global features output by the last feature extraction module and outputs the predicted position parameters and the probability of successful grasping. The implementation of the farthest point sampling method to select key points includes the following steps: Select an initial point p0 as the first point of the sampling point set S, S = {p0}; Calculate the shortest Euclidean distance d(p, S) from all unselected points p to all points in the current sampling point set S; Select the point with the largest shortest Euclidean distance and add it to S. S=S∪{p next }; Repeat this step until the sampling set S reaches a predetermined size k or meets other termination conditions; where ||ps|| represents the Euclidean distance between points p and s, p next Indicates the next point; The method of using channel attention to output local features to a feature propagation layer based on local point features includes the following steps: Perform global average pooling on the features of each channel, expressed as: Among them, x i,j,c The value of the cth channel at the (i, j) position of the feature map X representing the input channel attention, H and W represent the height and width of the feature map X, respectively, C represents the number of channels of the feature map X, and z c is the global average value of the c-th channel; The output of the global average pooling is further learned and processed through a small fully connected network, and the channel weight is expressed as: s1=σ(W2δ(W1z)); Where z represents the output of global average pooling, which consists of the global average of all channels; σ is the sigmoid activation function, which is used to generate weights in the range [0, 1]; δ is the ReLU activation function; W1 is a The dimension reduction weight matrix W2 is of size The dimension-upgraded weight matrix of , where r is the scaling ratio; s1 is the generated channel weight set, containing the channel weight of each channel; Multiply each channel of the original feature map X by the learned corresponding channel weight to complete the channel recalibration: in, Represents x i,j,c The recalibrated feature map, s c Represents the channel weight of the c-th channel.

2. The method for predicting the grasping position of a parallel gripper of a robotic arm based on improved PointNet++ according to claim 1 is characterized in that: The number of the feature extraction modules is at least three, and the point set abstraction layers of different feature extraction modules use query balls with different radii and different numbers of sampling points to extract local features of different scales.

3. The method for predicting the grasping position of a parallel gripper of a robotic arm based on improved PointNet++ according to claim 1, characterized in that: The other termination conditions include a convergence condition, which means that the coverage effect of the newly added points on the existing sampling set during the sampling process no longer significantly improves, as shown by the increase in the shortest distance of the newly added points being less than a predetermined threshold.

4. The method for predicting the grasping position of a parallel gripper of a robotic arm based on improved PointNet++ according to claim 1, characterized in that: p0 is selected randomly or heuristically.

5. The method for predicting the grasping position of a parallel gripper of a robotic arm based on improved PointNet++ according to claim 1, characterized in that: When training the location prediction model, a combined loss function including mean square error loss MSE and cross entropy loss Cross-Entropy is used, which is expressed as: Combined Loss=α×MSE+β×Cross-Entropy; Among them, α and β are weight factors used to balance the impact of the two loss functions; is the i-th predicted position parameter, y1 i is the i-th true location parameter, n is the total number of samples; is the probability of successful capture of the i-th sample, y2 i is the true captured label of the i-th sample, y2 i =0 or y2 i =1.

6. The method for predicting the grasping position of a parallel gripper of a robotic arm based on improved PointNet++ according to claim 1, characterized in that: The predicted position parameters output by the fully connected layer are six-dimensional vectors, which include: The three-dimensional coordinates of the grasping position determine the exact point where the robot arm gripper lands in physical space, which serves as the geometric center of the grasping operation; The three-dimensional rotation angle of the grasping posture guides the parallel gripper to adjust to the direction suitable for grasping the target object.