Grabbing posture generation method based on enhanced single-view point cloud features

Abstract

This invention discloses a grasping posture generation method based on enhanced single-view point cloud features, belonging to the field of robot vision and manipulation technology. On one hand, this invention introduces a point cloud completion model to jointly model the completed point cloud with the original point cloud, supplementing the geometric information of invisible regions without relying on multi-view or precise models. Enhanced local point cloud features are obtained through a point cloud completion feature layer (PCF-Layer) and an attention-based feature fusion method, thereby improving feature integrity under single-view perception conditions. On the other hand, a pre-trained point cloud self-supervised encoder is introduced to extract shape prior features from the single-view object point cloud, used to characterize the overall shape structure and local geometric relationships of the object. A conditional variational autoencoder model is employed, where enhanced local geometric features and shape prior features serve as conditional variables to guide and constrain grasping posture generation, improving the structural consistency and stability of the generated grasping posture.

Description

Technical Field

[0001] This invention relates to the field of robot vision and manipulation technology, specifically to a method for generating grasping postures based on enhanced single-view point cloud features. Background Technology

[0002] With the widespread application of robotics in industrial assembly, warehousing and logistics, and home services, autonomous grasping has become a key capability for realizing intelligent robot operation. Six-DOF (6-DoF) grasping methods based on point clouds have gradually become a research hotspot due to their ability to provide rich three-dimensional geometric information and flexible grasping posture space. Traditional methods typically rely on two-dimensional point clouds (i.e., 2.5D point clouds) extracted from single-view depth images. These point clouds only reflect a single visible surface of the object and cannot provide the complete geometric shape of the object, thus easily leading to misjudgments of the object's true structure by the grasping algorithm. This is especially true for objects with reflective surfaces, transparent materials, or partial occlusion, where the grasping failure rate increases significantly. Furthermore, traditional feature extraction methods based on local point clouds often focus on modeling local geometric details, neglecting the overall topological structure and global spatial distribution of the object, making it difficult to comprehensively depict the object's true form. In addition, most existing grasping generation methods focus on directly regressing the grasping pose from the point cloud, often ignoring the kinematic constraints and workspace limitations faced by the robot during actual execution, further restricting their usability and generalization ability in real-world scenarios.

[0003] To overcome the incompleteness of single-view point clouds, an intuitive approach is to acquire complete 3D information of the object through multi-view acquisition or by deploying multiple cameras. However, such solutions are costly and complex to deploy, making them difficult to implement in dynamic or constrained environments. In recent years, with the development of point cloud completion technology, researchers have begun to explore generating approximately complete object point clouds by completing missing parts, and then using this as a basis for grasping planning. These methods input a portion of the observed point cloud into a trained completion network to predict the complete shape of the object, and then generate a grasping pose based on the completed point cloud. However, due to limited generalization ability, the performance of the completion model significantly decreases when the target object is an unseen object during training. This poses a significant limitation for high-precision tasks such as grasping. Therefore, previous completion-grasping methods that directly applied completed point clouds to grasping typically only achieve high success rates on the training dataset and have poor generalization ability for unseen objects.

[0004] How to extract richer geometric features from single-view point clouds, improve the consistency and stability of the captured structure, and increase its diversity is an urgent problem to be solved. To this end, a capture pose generation method based on enhanced single-view point cloud features is proposed. Summary of the Invention

[0005] The technical problem to be solved by this invention is: how to solve the problems of incomplete observation of single-view point cloud information, low stability of grasping and weak generalization ability caused by point cloud completion error in the existing technology, and to provide a grasping posture generation method based on enhanced single-view point cloud features.

[0006] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps:

[0007] S1: Enhanced Local Geometric Feature Acquisition

[0008] Visual information of the target object is acquired by a depth camera and converted into a raw point cloud. This raw point cloud is then input into a pre-trained PCN point cloud completion network to obtain a completed point cloud that does not overlap with the raw point cloud. The completed point cloud and the raw point cloud are then concatenated and processed by the attention-based point cloud completion feature module APCF. Multi-scale features are obtained using the PCF-Layer in the module, and then fused using attention-based multi-scale feature fusion to obtain fused features. Finally, the fused features are concatenated with the multi-scale features to obtain the enhanced local geometric features of the object point cloud.

[0009] S2: Shape Prior Feature Acquisition

[0010] Visual information of the target object is acquired by a depth camera and converted into a raw point cloud. The raw point cloud is then input into a pre-trained point cloud self-supervised encoding model. After block modeling and global association modeling, shape prior features are output to describe the overall shape structure and local geometric relationships of the target object.

[0011] S3: Creation and Generation Model Network Design and Processing

[0012] Based on the enhanced local geometric features obtained in step S1 and the shape prior features obtained in step S2, a conditional variational autoencoder model is used to generate the robotic arm's grasping posture. The enhanced local geometric features are used to improve the feature integrity under single-view perception conditions, while the shape prior features are used to provide the overall shape and structure of the target object and constrain the relationship between local contact geometry and the overall structure. Together, they serve as conditional variables to guide and constrain the grasping posture generation process.

[0013] Furthermore, in step S1, the process of acquiring the completed point cloud is as follows:

[0014] S11: Obtain visual information through a depth camera, and convert the visual information into local point cloud data of the scene in the camera coordinate system with three-dimensional spatial information according to the transformation relationship between pixel coordinates and world coordinates.

[0015] S12: Transform the local point cloud data of the scene in the camera coordinate system to the base coordinate system, which is the coordinate system where the robot arm base is located;

[0016] S13: Keeping the relative relationship between the target object and the robotic arm unchanged, filter out the local point cloud data of the object from the local point cloud data of the scene in the base coordinate system by setting a height threshold;

[0017] S14: Transform the local point cloud data of the object to the coordinate system of the end effector of the robotic arm, and use the FPS algorithm to perform noise reduction to obtain the original local point cloud data of the object.

[0018] S15: Input the original local point cloud data of the object into the pre-trained PCN point cloud completion network, and output the completed point cloud that does not contain the original point cloud.

[0019] Furthermore, in step S15, the specific processing procedure of the PCN point cloud completion network is as follows:

[0020] S151: Encode the raw point cloud data, perform feature mapping on each point in the point cloud through a shared multilayer perceptron, and extract the global feature vector using symmetric aggregation operations;

[0021] S152: Based on global feature vectors, a coarse point cloud of the target object is predicted by a decoding network to obtain a coarse complete point set representing the overall structure of the object;

[0022] S153: Based on the coarse completion point set, a two-dimensional parametric grid is introduced and combined with the global feature vector. Fine-grained completion points are generated through folding mapping to obtain the completion point cloud of the target object.

[0023] Furthermore, in step S1, the process of obtaining the enhanced local geometric features is as follows:

[0024] S16: The original point cloud and the completed point cloud are connected to form a mixed point cloud representing the entire object;

[0025] S17: Based on the hybrid point cloud, the attention-based point cloud completion feature module APCF is used for processing, and the point cloud completion feature layer PCF-Layer in the module is used to obtain the local geometric features of the neighborhood at different scales.

[0026] S18: The attention-based multi-scale feature fusion module APCF is used to fuse the local geometric features of the multi-scale neighborhood to generate fused features, which are then spliced ​​with the local geometric features of the multi-scale neighborhood to generate enhanced local geometric features for each point in the original point cloud.

[0027] Furthermore, in step S17, the point cloud completion feature layer PCF-Layer processing procedure is as follows:

[0028] S171: The sphere query method is adopted, with each point in the original point cloud as the query center, and the radius of the query area is determined according to the width of the robotic arm gripper.

[0029] S172: For each query region radius, select some nearest neighbor points from the hybrid point cloud to form multiple multi-scale neighborhoods;

[0030] S173: Input points from multiple neighborhoods at different scales into the corresponding number of PointNet++ style feature learning modules to obtain local geometric features of neighborhoods at different scales.

[0031] Furthermore, in step S18, the attention-based multi-scale feature fusion process is as follows:

[0032] S181: Channel mapping is performed on the local geometric features of the multi-scale neighborhood to map them to a unified dimensional space, so as to obtain the features of each scale represented by the unified dimension.

[0033] S182: Globally aggregate the mapped features at each scale to construct a scale description vector, and generate the weight coefficients corresponding to each scale through an attention network.

[0034] S183: Based on attention weights, features at each scale are weighted and fused to obtain fused features;

[0035] S184: The fused features are spliced ​​with features at each scale to obtain enhanced local geometric features.

[0036] Furthermore, in step S2, the process of obtaining the shape prior features is as follows:

[0037] S21: Obtain visual information through a depth camera, and convert the visual information into local point cloud data of the scene in the camera coordinate system with three-dimensional spatial information according to the transformation relationship between pixel coordinates and world coordinates;

[0038] S22: Transform the local point cloud data of the scene in the camera coordinate system to the base coordinate system;

[0039] S23: Keeping the relative relationship between the target object and the robotic arm unchanged, filter out the local point cloud data of the object from the local point cloud data of the scene in the base coordinate system by setting a height threshold;

[0040] S24: Transform the local point cloud data of the object to the coordinate system of the end effector of the robotic arm, and use the FPS algorithm to perform noise reduction processing to obtain the original local point cloud data of the object.

[0041] S25: The original point cloud data is divided into blocks using the FPS algorithm and the k-nearest neighbor algorithm to divide the point cloud into multiple local point cloud subsets;

[0042] S26: Calculate the geometric center point of each local point cloud subset, input the geometric center point into the position encoding function, and obtain the position encoding vector;

[0043] S27: Perform feature mapping processing on the point cloud data in each local point cloud subset. Using a weighted PointNet, an initial feature representation is obtained by using an MLP and a max pooling layer. This representation is then combined with the location encoding vector to form the input features of the local point cloud subset.

[0044] S28: Feed the input features into a pre-trained point cloud self-supervised encoder for feature encoding to obtain a high-dimensional feature representation corresponding to each local point cloud subset;

[0045] S29: By aggregating the high-dimensional feature representations of all local point cloud subsets through max pooling operations, shape prior features representing the overall shape structure and local geometric relationships of the target object are obtained.

[0046] Furthermore, in step S3, the specific process of capturing and processing the generative model network is as follows:

[0047] S31: A conditional variational autoencoder model is adopted, and the actual grasping posture and the corresponding conditional feature information are input into the encoder. The conditional feature information includes enhanced local geometric features and shape prior features.

[0048] S32: The encoder maps the actual grasping posture and conditional feature information, outputting the probability distribution parameters of the latent variables, including the mean vector μ and the variance vector σ, and constructing an approximate posterior distribution of the latent variables:

[0049] ;

[0050] in, As a latent variable, To accurately capture posture, For enhanced local geometric features, These are shape prior features;

[0051] S33: Based on the mean vector μ and variance vector σ, a reparameterization mechanism is used to sample the latent variables and output the latent variables. :

[0052] ;

[0053] in, Let be a random variable that follows a standard normal distribution;

[0054] S34: Decoder for latent variables The conditional feature information is processed to output the predicted grasping posture. :

[0055] ;

[0056] in, This indicates the decoder.

[0057] Furthermore, in step S31, the loss function of the conditional variational autoencoder model... The design is as follows:

[0058] ;

[0059] in, These are weight hyperparameters; For reconstruction loss; The loss is a posterior distribution constraint.

[0060] By minimizing the loss function The parameters of the encoder and decoder are optimized to enable the conditional variational autoencoder model to model the ambiguity and uncertainty of the grasping posture given enhanced local geometric features and shape prior features.

[0061] The present invention has the following advantages over the prior art:

[0062] 1. This invention introduces a point cloud completion model to generate a completed point cloud and jointly model the original point cloud. This allows for the supplementation of geometric information in invisible areas without relying on multi-view or accurate models. Multi-scale features are obtained through the point cloud completion feature layer PCF-layer, thereby improving the feature integrity under single-view perception conditions.

[0063] 2. This invention introduces a feature fusion method based on an attention mechanism to dynamically assign appropriate weights to features of different scales, thereby enhancing feature discrimination and improving the quality and stability of the generated grasping posture.

[0064] 3. This invention extracts prior shape features from a single-view point cloud and constrains the relationship between local contact geometry and overall structure during the grasping generation process, thereby reducing unreasonable grasping actions caused by relying solely on local information. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating the grasping posture generation method based on enhanced single-view point cloud features in an embodiment of the present invention.

[0066] Figure 2 This is an example of the original point cloud of an object in an embodiment of the present invention;

[0067] Figure 3 This is a schematic diagram of the PCN point cloud completion network structure in an embodiment of the present invention;

[0068] Figure 4 This is an example of object completion point cloud in an embodiment of the present invention;

[0069] Figure 5 This is an example of a mixed point cloud of objects in an embodiment of the present invention;

[0070] Figure 6 This is a flowchart illustrating the enhanced local geometric feature extraction process of the attention-based point cloud completion feature module APCF in an embodiment of the present invention.

[0071] Figure 7 This is a flowchart of shape prior feature extraction in an embodiment of the present invention;

[0072] Figure 8 This is an example of grasping posture generation in an embodiment of the present invention. Detailed Implementation

[0073] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0074] like Figure 1 As shown, this embodiment provides a technical solution: a grasping pose generation method based on enhanced single-view point cloud features, including the following steps:

[0075] Step 1: Enhanced Local Geometric Feature Acquisition

[0076] Visual information of the target object is acquired by a depth camera and converted into a raw point cloud. This raw point cloud is then input into a pre-trained PCN point cloud completion network to obtain a completed point cloud that differs from the raw point cloud. The completed point cloud and the raw point cloud are then connected in a hybrid manner and processed by the Attention-Based Point Cloud Completion Feature Module (APCF). The module utilizes the PCF-Layer to obtain multi-scale features, which are then fused using attention-based multi-scale feature fusion to obtain a fused feature. Finally, the fused feature is concatenated with the multi-scale feature to obtain the enhanced local geometric features of the object's point cloud.

[0077] In this embodiment, the point cloud acquisition process is as follows:

[0078] Visual information is acquired through a depth camera, and then converted into point cloud data with three-dimensional spatial information based on the transformation relationship between pixel coordinates and world coordinates. This yields local point cloud data of the scene in the camera coordinate system.

[0079] The scene local point cloud data in the camera coordinate system is transformed to the base coordinate system to obtain the scene local point cloud data in the base coordinate system, where the base coordinate system is the coordinate system of the robot arm base. The relative relationship between the target object and the robot arm remains unchanged. The object local point cloud data is filtered out from the scene local point cloud data in the base coordinate system by setting a height threshold. The object local point cloud data is then transformed to the coordinate system of the robot arm tool end effector and processed using the FPS algorithm to remove noise, resulting in the denoised object local original point cloud P. The number of points in the cloud is fixed at 1024. Figure 2 As shown;

[0080] PCN point cloud network completion, such as Figure 3 As shown, the obtained original point cloud P is first processed by a shared multilayer perceptron to perform feature mapping on each point in the point cloud, resulting in intermediate features G. After concatenation with the intermediate features F, the point cloud is again processed by a shared multilayer perceptron to perform feature mapping on each point, resulting in a global feature vector V. The global feature vector V is then used to generate a coarse point cloud for predicting the target object through a fully connected network. A two-dimensional parametric mesh is introduced and combined with the global feature vector V. Through folding mapping, the two-dimensional mesh is mapped to three-dimensional space, generating more detailed points around the coarsely completed point cloud, thus obtaining the final finely completed point cloud. The number of point clouds is 1024.

[0081] like Figure 4 As shown, the obtained complete point cloud For points not included in the original point cloud, the enhanced local geometric feature acquisition process based on the obtained completed point cloud is as follows:

[0082] like Figure 5 As shown, the obtained complete point cloud Connect the original point cloud P to form a hybrid point cloud. :

[0083]

[0084] Hybrid point cloud After such Figure 6 The attention-based point cloud completion feature module APCF processing is shown below:

[0085] In the point cloud completion feature layer PCF-Layer, the sphere query method is adopted, with each point in the original point cloud as the query center. The query region radius is determined by the width of the gripper, taking one-third, one-half of the maximum opening and closing width of the gripper, and the maximum opening and closing width of the gripper as the three query region radii.

[0086] For each query region radius, select 64, 64, or 128 nearest neighbor points from the mixed point cloud to form a structure like... Figure 6 middle , , Multi-scale neighborhood;

[0087] Points in neighborhoods at different scales are input into three PointNet++ style feature learning modules to obtain local geometric features of neighborhoods at different scales. , , Their dimensions are 64, 128, and 128 respectively;

[0088] Channel mapping is performed on the local geometric features of the multi-scale neighborhood to map them to a feature space of uniform dimension (the default feature dimension is set to 256), resulting in the mapped multi-scale features. :

[0089]

[0090] in, It is a channel mapping function;

[0091] For the mapped multi-scale features The data is fed into an attention-based multi-scale feature fusion module for global aggregation, constructing a scale description vector, and generating weight coefficients for each scale through an attention network.

[0092]

[0093] Among them, the weighting coefficient include , , Three coefficients, and satisfying:

[0094]

[0095] Based on attention weights The features at each scale are weighted and fused to obtain 256-dimensional fused features. :

[0096]

[0097] Fusion features With multi-scale local geometric features , , By stitching the data together, we obtain 576-dimensional enhanced local geometric features for each point in the original point cloud. .

[0098] Step 2: Obtaining shape prior features

[0099] Visual information of the target object is acquired by a depth camera and converted into raw point cloud. The raw point cloud data is then input into a pre-trained point cloud self-supervised coding model to perform block modeling and global correlation modeling on the point cloud, and output shape prior features to describe the overall shape structure and local geometric relationships of the target object.

[0100] In this step, the original point cloud P is obtained using the same method as in step one, and the farthest point sampling (FPS) method and the K-nearest neighbor (KNN) algorithm are used to divide the point cloud into blocks. Specifically, given a set of 1024 points... The original point cloud P is sampled from the center point CT using the farthest point sampling (FPS) method, as shown in the formula below, selecting 64 points;

[0101]

[0102] Based on the selected 64 center points CT, the k-nearest neighbor (KNN) algorithm is used, as shown in the formula below, to select 32 nearest neighbor points from the original point cloud P to form the corresponding local point cloud subset patch;

[0103]

[0104] For each local point cloud subset patch, on the one hand, each point in the local point cloud subset patch is represented by normalized coordinates relative to its center point CT. The normalized point patch is input into a weighted PointNet, and an MLP layer and max pooling are used to output the initial feature representation. On the other hand, the position encoding vector is obtained through the position encoding function. :

[0105]

[0106] in, Here, S is the coordinate of the i-th point in the local point cloud subset, and S is a hyperparameter. It is the input tensor;

[0107] like Figure 7 The flowchart shown is for extracting prior features of the shape. The input is fed into the point cloud self-supervised encoder (Transformerencoder) and combined with the position encoding vector. The standard transformer module processes the data to obtain a high-level vector representation of each patch. :

[0108]

[0109] Where c is the dimension size (the default value is 384).

[0110] Max pooling is used to extract high-dimensional features from all local point cloud patches. Aggregation is performed to obtain 384-dimensional shape prior features for characterizing the overall structure of the target object. .

[0111] Step 3: Design and Processing of the Generative Model Network

[0112] Based on the enhanced local geometric features obtained in step one, combined with the shape prior features obtained in step two, a conditional variational autoencoder model is used to generate the gripping posture of the robotic arm. The enhanced local geometric features are used to improve the feature integrity under single-view perception conditions, while the shape prior features are used to provide the overall shape structure of the target object and the relationship between the local contact geometry and the overall shape structure. Together, they serve as conditional variables to guide and constrain the gripping posture generation process.

[0113] In this embodiment, a conditional variational autoencoder model is used to model the generation process of the grasping posture as the following conditional probability distribution:

[0114]

[0115] Where g represents the grasping posture, which includes a matrix, This represents the enhanced local geometric feature of the i-th object point obtained in step one. This represents the prior features of the target object shape obtained in step two. Representing latent variables, which follow a standard normal distribution:

[0116]

[0117] Specifically, during the training phase, a conditional encoder is constructed based on each known crawled training sample to approximate the posterior distribution:

[0118]

[0119] Given the true grasping pose g and the enhanced local geometric features and shape prior features The conditional probability distribution is learned, and the mean and variance of the latent variable distribution are output:

[0120]

[0121] Latent variables were obtained through sampling using reparameterization techniques. :

[0122]

[0123] Latent variables With conditional features and Input to decoder to generate predicted grasping pose :

[0124]

[0125] In this embodiment, the loss function is designed as follows, wherein the reconstruction loss... The constrained grasping pose is close to the actual grasping pose, and the KL divergence is... Constraint latent variable distribution The distribution approximates a standard normal distribution p(z)=N(0,I), where λ is a weight hyperparameter used to adjust the ratio of the two parts of the loss.

[0126]

[0127] By minimizing the loss function The parameters of the conditional encoder and decoder are optimized to enable the model to effectively model the ambiguity and uncertainty of the grasping posture given enhanced local geometric features and global shape priors.

[0128] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A grasping pose generation method based on enhanced single-view point cloud features, characterized in that, Includes the following steps: S1: Enhanced Local Geometric Feature Acquisition Visual information of the target object is acquired by a depth camera and converted into a raw point cloud. This raw point cloud is then input into a pre-trained PCN point cloud completion network to obtain a completed point cloud that does not overlap with the raw point cloud. The completed point cloud and the raw point cloud are then concatenated and processed by the attention-based point cloud completion feature module APCF. Multi-scale features are obtained using the PCF-Layer in the module, and then fused using attention-based multi-scale feature fusion to obtain fused features. Finally, the fused features are concatenated with the multi-scale features to obtain the enhanced local geometric features of the object point cloud. S2: Shape Prior Feature Acquisition Visual information of the target object is acquired by a depth camera and converted into a raw point cloud. The raw point cloud is then input into a pre-trained point cloud self-supervised encoding model. After block modeling and global association modeling, shape prior features are output to describe the overall shape structure and local geometric relationships of the target object. S3: Creation and Generation Model Network Design and Processing Based on the enhanced local geometric features obtained in step S1 and the shape prior features obtained in step S2, a conditional variational autoencoder model is used to generate the robotic arm's grasping posture. The enhanced local geometric features are used to improve the feature integrity under single-view perception conditions, while the shape prior features are used to provide the overall shape and structure of the target object and constrain the relationship between local contact geometry and the overall structure. Together, they serve as conditional variables to guide and constrain the grasping posture generation process.

2. The grasping pose generation method based on enhanced single-view point cloud features according to claim 1, characterized in that, In step S1, the process of acquiring the point cloud completion is as follows: S11: Obtain visual information through a depth camera, and convert the visual information into local point cloud data of the scene in the camera coordinate system with three-dimensional spatial information according to the transformation relationship between pixel coordinates and world coordinates. S12: Transform the local point cloud data of the scene in the camera coordinate system to the base coordinate system, which is the coordinate system where the robot arm base is located; S13: Keeping the relative relationship between the target object and the robotic arm unchanged, filter out the local point cloud data of the object from the local point cloud data of the scene in the base coordinate system by setting a height threshold; S14: Transform the local point cloud data of the object to the coordinate system of the end effector of the robotic arm, and use the FPS algorithm to perform noise reduction to obtain the original local point cloud data of the object. S15: Input the original local point cloud data of the object into the pre-trained PCN point cloud completion network, and output the completed point cloud that does not contain the original point cloud.

3. The grasping posture generation method based on local geometric features and shape prior features according to claim 2, characterized in that, In step S15, the specific processing procedure of the PCN point cloud completion network is as follows: S151: Encode the raw point cloud data, perform feature mapping on each point in the point cloud through a shared multilayer perceptron, and extract the global feature vector using symmetric aggregation operations; S152: Based on global feature vectors, a coarse point cloud of the target object is predicted by a decoding network to obtain a coarse complete point set representing the overall structure of the object; S153: Based on the coarse completion point set, a two-dimensional parametric grid is introduced and combined with the global feature vector. Fine-grained completion points are generated through folding mapping to obtain the completion point cloud of the target object.

4. The grasping pose generation method based on enhanced single-view point cloud features according to claim 1, characterized in that, In step S1, the process of obtaining the enhanced local geometric features is as follows: S16: The original point cloud and the completed point cloud are connected to form a mixed point cloud representing the entire object; S17: Based on the hybrid point cloud, the attention-based point cloud completion feature module APCF is used for processing, and the point cloud completion feature layer PCF-Layer in the module is used to obtain the local geometric features of the neighborhood at different scales. S18: The attention-based multi-scale feature fusion module APCF is used to fuse the local geometric features of the multi-scale neighborhood to generate fused features, which are then spliced ​​with the local geometric features of the multi-scale neighborhood to generate enhanced local geometric features for each point in the original point cloud.

5. The grasping pose generation method based on enhanced single-view point cloud features according to claim 4, characterized in that, In step S17, the point cloud completion feature layer PCF-Layer processing procedure is as follows: S171: The sphere query method is adopted, with each point in the original point cloud as the query center, and the radius of the query area is determined according to the width of the robotic arm gripper. S172: For each query region radius, select some nearest neighbor points from the hybrid point cloud to form multiple multi-scale neighborhoods; S173: Input points from multiple neighborhoods at different scales into the corresponding number of PointNet++ style feature learning modules to obtain local geometric features of neighborhoods at different scales.

6. The grasping pose generation method based on enhanced single-view point cloud features according to claim 5, characterized in that, In step S18, the attention-based multi-scale feature fusion process is as follows: S181: Channel mapping is performed on the local geometric features of the multi-scale neighborhood to map them to a unified dimensional space, so as to obtain the features of each scale represented by the unified dimension. S182: Globally aggregate the mapped features at each scale to construct a scale description vector, and generate the weight coefficients corresponding to each scale through an attention network. S183: Based on attention weights, features at each scale are weighted and fused to obtain fused features; S184: The fused features are spliced ​​with features at each scale to obtain enhanced local geometric features.

7. The grasping pose generation method based on enhanced single-view point cloud features according to claim 1, characterized in that, In step S2, the process of obtaining the shape prior features is as follows: S21: Obtain visual information through a depth camera, and convert the visual information into local point cloud data of the scene in the camera coordinate system with three-dimensional spatial information according to the transformation relationship between pixel coordinates and world coordinates; S22: Transform the local point cloud data of the scene in the camera coordinate system to the base coordinate system; S23: Keeping the relative relationship between the target object and the robotic arm unchanged, filter out the local point cloud data of the object from the local point cloud data of the scene in the base coordinate system by setting a height threshold; S24: Transform the local point cloud data of the object to the coordinate system of the end effector of the robotic arm, and use the FPS algorithm to perform noise reduction processing to obtain the original local point cloud data of the object. S25: The original point cloud data is divided into blocks using the FPS algorithm and the k-nearest neighbor algorithm to divide the point cloud into multiple local point cloud subsets; S26: Calculate the geometric center point of each local point cloud subset, input the geometric center point into the position encoding function, and obtain the position encoding vector; S27: Perform feature mapping processing on the point cloud data in each local point cloud subset. Using a weighted PointNet, an initial feature representation is obtained by using an MLP and a max pooling layer. This representation is then combined with the location encoding vector to form the input features of the local point cloud subset. S28: Feed the input features into a pre-trained point cloud self-supervised encoder for feature encoding to obtain a high-dimensional feature representation corresponding to each local point cloud subset; S29: By aggregating the high-dimensional feature representations of all local point cloud subsets through max pooling operations, shape prior features representing the overall shape structure and local geometric relationships of the target object are obtained.

8. The grasping pose generation method based on enhanced single-view point cloud features according to claim 1, characterized in that, In step S3, the specific process of designing and processing the capture generative model network is as follows: S31: A conditional variational autoencoder model is adopted, and the actual grasping posture and the corresponding conditional feature information are input into the encoder. The conditional feature information includes enhanced local geometric features and shape prior features. S32: The encoder maps the actual grasping posture and conditional feature information, outputting the probability distribution parameters of the latent variables, including the mean vector μ and the variance vector σ, and constructing an approximate posterior distribution of the latent variables: ； in, As a latent variable, To accurately capture posture, For enhanced local geometric features, These are shape prior features; S33: Based on the mean vector μ and variance vector σ, a reparameterization mechanism is used to sample the latent variables and output the latent variables. : ； in, Let be a random variable that follows a standard normal distribution; S34: Decoder for latent variables The conditional feature information is processed to output the predicted grasping posture. : ； in, This indicates the decoder.

9. The grasping pose generation method based on enhanced single-view point cloud features according to claim 8, characterized in that, In step S31, the loss function of the conditional variational autoencoder model The design is as follows: ； in, These are weight hyperparameters; For reconstruction loss; The loss is a posterior distribution constraint. By minimizing the loss function The parameters of the encoder and decoder are optimized to enable the conditional variational autoencoder model to model the ambiguity and uncertainty of the grasping posture given enhanced local geometric features and shape prior features.

Images