A robot grasping posture estimation method, device, equipment and storage medium

Abstract

The present invention discloses a robot grasping posture estimation method, device, equipment and storage medium. By acquiring the first point cloud data of the target object, the first point cloud data is preprocessed to obtain the second point cloud data of the target object; according to the grasping posture sampling data, a grasping quality assessment network is constructed, and the second point cloud data is input into the grasping quality assessment network, so that the grasping quality assessment network outputs the optimal grasping posture. Compared with the prior art, the robot grasping posture estimation can be performed directly based on the three-dimensional point cloud, the position of the grasped target object in the three-dimensional space can be accurately obtained, and by directly processing the first point cloud data, the original point cloud information is maximized, and the grasping posture is estimated more effectively and accurately.

Description

Technical Field

[0001] The present invention relates to the field of robotics technology, and in particular to a robot grasping posture estimation method, device, equipment and storage medium. Background Art

[0002] At present, most deep learning methods for grasping pose estimation still rely on two-dimensional image input. However, due to the lack of depth information, two-dimensional images cannot accurately obtain the position of the grasped object in three-dimensional space. On the other hand, for scenes with complex backgrounds or severe occlusion, two-dimensional images are far less sensitive to object boundaries than depth images. Some grasping models require the setting of complex manual features to process data, and few models consider the point cloud information of three-dimensional objects, which limits their scope of application and also leads to problems such as low accuracy of grasping operations and unstable algorithms. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a method, device, equipment and storage medium for estimating the grasping posture of a robot, so as to improve the accuracy of the grasping posture of the robot.

[0004] In order to solve the above technical problems, the present invention provides a robot grasping posture estimation method, comprising:

[0005] Acquire first point cloud data of a target object, and preprocess the first point cloud data to obtain second point cloud data of the target object;

[0006] A grasping quality assessment network is constructed according to the grasping posture sampling data, and the second point cloud data is input into the grasping quality assessment network so that the grasping quality assessment network outputs an optimal grasping posture.

[0007] In a possible implementation, constructing a grasping quality assessment network according to grasping posture sampling data specifically includes:

[0008] A grasping pose estimation dataset is obtained, grasping pose sampling is performed on each model in the grasping pose estimation dataset, and each grasping pose is scored according to a preset quality assessment standard, and a grasping quality assessment network is constructed based on PointNet, wherein the grasping quality assessment network includes a multi-layer perceptron, a maximum pooling layer, and a fully connected layer.

[0009] In a possible implementation, preprocessing is performed on the first point cloud data, wherein the preprocessing includes point cloud direct filtering processing;

[0010] Coordinate axis thresholds in various directions of the spatial coordinate axis are set, the first point cloud data is segmented according to the coordinate thresholds in various directions to obtain a region of interest of the first point cloud data, and target point cloud recognition is performed on the region of interest to generate through-filtered point cloud data.

[0011] In a possible implementation, the first point cloud data is preprocessed, wherein the preprocessing further includes stage plane segmentation processing;

[0012] The background point cloud approximate plane in the through-filtered point cloud data is estimated based on a random sampling consistency algorithm, and based on the background point cloud approximate plane, the through-filtered point cloud data is segmented into a target object point cloud and a background point cloud to obtain a target object point cloud.

[0013] In a possible implementation manner, preprocessing is performed on the first point cloud data, wherein the preprocessing further includes point cloud downsampling processing;

[0014] Creating a three-dimensional voxel grid according to the target object point cloud, wherein the size of the three-dimensional voxel grid is determined by the resolution and sampling density of the target object point cloud;

[0015] Placing the target object point cloud into the three-dimensional voxel grid to generate a plurality of small cubes;

[0016] According to the center of gravity points of the plurality of small cubes, the distance between each target object point cloud in each small cube and the center of gravity point is calculated, and the target object point cloud with the smallest distance is retained.

[0017] The present invention also provides a robot grasping posture estimation device, comprising: a point cloud data processing module and an optimal grasping posture output module;

[0018] The point cloud data processing module is used to obtain first point cloud data of the target object, pre-process the first point cloud data, and obtain second point cloud data of the target object;

[0019] The optimal grasping posture output module is used to construct a quality assessment network according to the grasping posture sampling data, and input the second point cloud data into the grasping quality assessment network so that the grasping quality assessment network outputs the optimal grasping posture.

[0020] In a possible implementation, the optimal grasping posture output module is used to construct a grasping quality evaluation network according to the grasping posture sampling data, specifically including:

[0021] A grasping pose estimation dataset is obtained, grasping pose sampling is performed on each model in the grasping pose estimation dataset, and each grasping pose is scored according to a preset quality assessment standard, and a grasping quality assessment network is constructed based on PointNet, wherein the grasping quality assessment network includes a multi-layer perception machine, a maximum pooling layer and a fully connected layer.

[0022] In a possible implementation, the point cloud data processing module is used to pre-process the first point cloud data, wherein the pre-processing includes point cloud direct filtering processing;

[0023] Setting coordinate axis thresholds in various directions of the spatial coordinate axis, segmenting the first point cloud data according to the coordinate thresholds in various directions to obtain a region of interest of the first point cloud data, and performing target point cloud recognition on the region of interest to generate through-filtered point cloud data;

[0024] In a possible implementation, the point cloud data processing module is used to pre-process the first point cloud data, wherein the pre-processing further includes stage plane segmentation processing;

[0025] The background point cloud approximate plane in the through-filtered point cloud data is estimated based on a random sampling consistency algorithm, and based on the background point cloud approximate plane, the through-filtered point cloud data is segmented into a target object point cloud and a background point cloud to obtain a target object point cloud.

[0026] In a possible implementation, the point cloud data processing module is used to pre-process the first point cloud data, wherein the pre-processing further includes point cloud downsampling processing;

[0027] Creating a three-dimensional voxel grid according to the target object point cloud, wherein the size of the three-dimensional voxel grid is determined by the resolution and sampling density of the target object point cloud;

[0028] Placing the target object point cloud into the three-dimensional voxel grid to generate a plurality of small cubes;

[0029] According to the center of gravity points of the plurality of small cubes, the distance between each target object point cloud in each small cube and the center of gravity point is calculated, and the target object point cloud with the smallest distance is retained.

[0030] The present invention also provides a terminal device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the robot grasping posture estimation method as described in any one of the above items is implemented.

[0031] The present invention also provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the robot grasping pose estimation method as described in any one of the above.

[0032] The robot grasping posture estimation method, device, equipment and storage medium of the embodiment of the present invention have the following beneficial effects compared with the prior art:

[0033] By acquiring the first point cloud data of the target object and preprocessing the first point cloud data, the second point cloud data of the target object is obtained; according to the grasping posture sampling data, a grasping quality assessment network is constructed, and the second point cloud data is input into the grasping quality assessment network so that the grasping quality assessment network outputs the optimal grasping posture. Compared with the existing technology, the robot grasping posture estimation can be performed directly based on the three-dimensional point cloud, the position of the grasped target object in the three-dimensional space can be accurately obtained, and by directly processing the first point cloud data, the original point cloud information can be maximized, and the grasping posture can be estimated more effectively and accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of an embodiment of a robot grasping posture estimation method provided by the present invention;

[0035] Figure 2 It is a structural schematic diagram of an embodiment of a robot grasping posture estimation device provided by the present invention;

[0036] Figure 3 is a schematic diagram of first point cloud data of an embodiment provided by the present invention;

[0037] Figure 4 is a schematic diagram of the result of point cloud through filtering according to an embodiment of the present invention;

[0038] Figure 5 is a schematic diagram of the voxel filtering principle of an embodiment provided by the present invention;

[0039] Figure 6 It is a schematic diagram of the result of point cloud downsampling according to an embodiment of the present invention.

[0040] Figure 7 It is a schematic diagram of the structure of each layer of the space transformer network STN in PointNet of an embodiment provided by the present invention;

[0041] Figure 8 1 is a schematic diagram of a PointNet-based grasping quality assessment network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the accompanying drawings in the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Example 1

[0044] See also Figure 1 , Figure 1 FIG. 1 is a flow chart of an embodiment of a robot grasping posture estimation method provided by the present invention. Figure 1 As shown, the method includes steps 101 to 102, which are specifically as follows:

[0045] Step 101: Acquire first point cloud data of a target object, and preprocess the first point cloud data to obtain second point cloud data of the target object.

[0046] In one embodiment, a depth camera is used to acquire first point cloud data of an object, wherein the first point cloud data is a set of vectors in a three-dimensional coordinate system, which is a representation of a three-dimensional model, and each point cloud is usually represented by three-dimensional coordinates (x, y, z).

[0047] Compared with two-dimensional images and grid images, point cloud data contains information such as RGB values ​​and grayscale values, which can intuitively depict the real world. The first schematic diagram of point cloud data is as follows: Figure 3 shown.

[0048] In one embodiment, due to reasons such as ambient light and camera errors, the first point cloud data collected by the depth camera will contain data noise, and in addition to the point cloud information of the target object, there will also be redundant background information, stage plane information, etc. These noises will affect the recognition and calculation of the target. Therefore, in this embodiment, in order to avoid unnecessary calculations and improve subsequent recognition efficiency and accuracy, it is necessary to preprocess the acquired first point cloud data.

[0049] In one embodiment, the first point cloud data is preprocessed, wherein the preprocessing includes: point cloud through-filtering processing, stage plane segmentation processing and point cloud downsampling processing.

[0050] In one embodiment, the first point cloud data is subjected to point cloud through-filtering processing, and the coordinate axis thresholds in each direction of the spatial coordinate axis are set, and the first point cloud data is segmented according to the coordinate thresholds in each direction to obtain the region of interest of the first point cloud data, and the target point cloud is identified in the region of interest to generate through-filtered point cloud data. The effect of the point cloud through-filtering processing on the first point cloud data is as follows: Figure 4 As shown, Figure 4 Schematic diagram of the result of point cloud direct filtering.

[0051] Specifically: For the first point cloud data collected by the depth camera, in order to find the area of ​​interest where the target object is located more quickly, the point cloud direct filtering algorithm is used to remove data noise and outliers. The point cloud direct filtering is cropped within the specified coordinate range, and you can choose to keep points within the range or points outside the range. Therefore, the number of outliers and noise point clouds outside the target range can be greatly reduced.

[0052] Since the distance between the target object point cloud to be identified and the depth camera remains unchanged, the coordinate axis thresholds in each direction of the spatial coordinate axis can be set to intercept the point cloud data within a certain range in the spatial coordinate axis direction as the region of interest, and then identify the target point cloud. The distance threshold settings in each direction of the spatial coordinate axis are as follows:

[0053]

[0054] In the formula, X min , X max , Y min , Y max , Z min , Z max The threshold can be set reasonably according to the actual situation to achieve a balance between reducing the number of point clouds and retaining the area of ​​interest.

[0055] In one embodiment, the first point cloud data is subjected to stage plane segmentation processing, the background point cloud approximate plane in the through-filtered point cloud data is estimated based on a random sampling consistency algorithm, and based on the background point cloud approximate plane, the through-filtered point cloud data is segmented into a target object point cloud and a background point cloud to obtain a target object point cloud.

[0056] Specifically: The essence of the random sampling consensus algorithm is to find the best patch in the three-dimensional motor, estimate and optimize the mathematical model in the point cloud data containing external points by random selection iteration, and set the mathematical model as a plane model. Separate the stage plane from the target object.

[0057] In the robot grasping scenario, the target object placement platform can be regarded as a plane, but due to the noise in the depth camera sampling process, the background point cloud is not completely on the same plane, but distributed near a certain plane. Based on this, the background point cloud approximate plane is estimated by using a random sampling consistency algorithm, and then the target object point cloud and the background point cloud are segmented with reference to the background point cloud approximate plane to obtain an independent target object point cloud.

[0058] In one embodiment, the first point cloud data is subjected to point cloud downsampling processing, and a three-dimensional voxel grid is created based on the target object point cloud, wherein the size of the three-dimensional voxel grid is determined by the resolution and sampling density of the target object point cloud; the target object point cloud is placed into the three-dimensional voxel grid accordingly to generate a plurality of small cubes; based on the center of gravity points of the plurality of small cubes, the distance between each target object point cloud in each small cube and the center of gravity point is calculated, and the target object point cloud with the smallest distance is retained.

[0059] Specifically: Since the first point cloud data acquired by the depth camera is relatively dense, in order to improve the speed of subsequent point cloud processing, a voxel filtering algorithm is used to downsample the target object point cloud while retaining the object feature information.

[0060] The voxel filtering algorithm creates a three-dimensional voxel grid based on the point cloud data, and then determines the size of the cube grid based on the point cloud resolution, sampling density, etc. For example, for large-scale point cloud data with low point cloud resolution and low sampling density, the cube side length can be set to 1cm to 1m. For small-scale point cloud data with high point cloud resolution and high sampling density, the cube side length can be set to 0.1mm to 10mm. The specific side length can be set manually. Then align the target object point cloud data into the three-dimensional voxel grid to form multiple small cubes. Use the center of gravity of the small cube to replace the points in the cube to sparse the point cloud data. Figure 5 As shown, Figure 5 Schematic diagram of voxel filtering principle. Figure 5 The point P in the small cube shown in FIG. 1 replaces all the points in this small cube.

[0061] The center of gravity of each small cube (x c ,y c ,z c ) is calculated as:

[0062]

[0063] Where n is the number of point cloud data points in the small cube. The distance between each point in the small cube and the centroid is calculated, and the point cloud closest to the centroid is retained. The remaining data points are deleted, and the point cloud closest to the centroid is used as the representative of the data points in the area where the small cube is located, thereby achieving point cloud downsampling. The result of point cloud downsampling is shown in Figure 6 shown.

[0064] In this embodiment, the original point cloud information is retained to the maximum extent by directly processing the acquired first point cloud data, so that the grasping posture can be estimated more effectively and accurately later.

[0065] Step 102: construct a grasping quality assessment network according to the grasping posture sampling data, and input the second point cloud data into the grasping quality assessment network so that the grasping quality assessment network outputs the optimal grasping posture.

[0066] In one embodiment, a grasping pose estimation dataset YCB is obtained, wherein the grasping pose estimation dataset YCB includes 77 common object models, 46,200 RGB images and corresponding depth images. The grasping pose estimation dataset YCB is used for subsequent network model training, and grasping pose sampling data is obtained by sampling the grasping pose of each model in the grasping pose estimation dataset YCB, and each grasping pose obtained by sampling is scored according to a preset quality evaluation standard.

[0067] In one embodiment, when sampling the grasping pose of each model in the grasping pose estimation data set YCB, first randomly find two surface points of the object model as the contact points of the grasper, and then specify a grasping angle of 0 to 90 degrees, and then construct a grasping pose according to the specified contact point and grasping angle; repeatedly execute the above grasping pose construction process to obtain a large number of candidate grasping poses, and screen the obtained large number of candidate grasping poses, and judge whether the grasper will collide with the grid after executing the grasping action. If so, remove the grasping pose, if not, save the appropriate candidate grasping pose.

[0068] In one embodiment, in order to evaluate the quality of the grasping posture, each grasping posture obtained by sampling needs to be scored according to a preset quality evaluation standard.

[0069] Q based on force closure fc and Q based on the grasping spinor space gws There are two mainstream grasping quality indicators. Among them, the force closure method takes into account the friction between the grasper and the object, while the analysis method based on the grasping spinor space can be used in the case of frictionlessness. fc The value of depends on the friction coefficient. The lower the friction coefficient, the higher the grasping quality and the more robust the grasping. So Qfc The specific value of Q is calculated by the inverse of the friction coefficient. gws The specific value of is the GWS radius. Based on the above, this embodiment uses the combined force to close the Q fc and Q of the grasping spinor space gws The preset quality evaluation standard ensures the scoring quality of the grasping posture, thereby improving the accuracy and credibility of the posture estimation. The preset quality evaluation ratio standard is as follows:

[0070] Q=αQ fc +βQ gws ;

[0071] Among them, due to Q gws Q fc is much larger, so the coefficients (α, β) are set to (1.0, 0.01).

[0072] In one embodiment, a grasping quality assessment network is constructed based on PointNet, wherein the grasping quality assessment network is composed of shared 64×64 and 64×128×1024 multi-layer perceptrons, a maximum pooling layer, and a fully connected layer with an output size of 512×256×C (C is the number of categories). The schematic diagram of the grasping quality assessment network based on PointNet is shown in the figure. Figure 8 shown.

[0073] Specifically: Input an N×3 matrix to the spatial transformation network STN in PointNet and perform an input matrix transformation, where N is the number of point clouds, such as Figure 7 As shown, Figure 7 This is a schematic diagram of the structure of each layer of the spatial transformation network STN in PointNet. The spatial transformation network STN outputs a 3×3 matrix, and then projects each point into a 64-dimensional high-dimensional space through a multi-layer perceptron, and then performs a high-dimensional space transformation to form a more normalized 64-dimensional matrix. The multi-layer perceptron is used to map the 64-dimensional to 1024-dimensional. Symmetric operations can be performed in 1024 dimensions to obtain global features with a dimension of 1024. Symmetric operations ensure the symmetry of the point cloud data by taking the maximum value. Finally, the point cloud is classified through a cascaded fully connected network to output the optimal grasping posture. Among them, the point cloud classification categories include two categories, namely good grasping and bad grasping.

[0074] In one embodiment, the second point cloud data is input into the grasping quality assessment network so that the grasping quality assessment network outputs an optimal grasping posture.

[0075] In this embodiment, since the second point cloud data is obtained after preprocessing the first point cloud data, it is different from the prior art that uses the entire object point cloud data as input. In this embodiment, the number of input point clouds can be reduced, thereby greatly reducing the amount of calculation, and in the network training and prediction stages, the efficiency of network learning and inference can be effectively improved. In addition, the grasping posture is directly output based on the second point cloud data, and there is no need to project the 3D point cloud into multiple 2D photos, nor is it necessary to rasterize it into dense 3D volume data. It can maximize the geometric information of the original point cloud, more effectively infer the grasping quality, and improve the grasping efficiency.

[0076] Example 2

[0077] See also Figure 2 , Figure 2 FIG. 1 is a flow chart of an embodiment of a robot grasping posture estimation method provided by the present invention. Figure 2 As shown, the device includes a point cloud data processing module 201 and an optimal grasping posture output module 202, which are specifically as follows:

[0078] The point cloud data processing module 201 is used to obtain first point cloud data of the target object, and pre-process the first point cloud data to obtain second point cloud data of the target object.

[0079] The optimal grasping posture output module 202 is used to construct a quality assessment network according to the grasping posture sampling data, and input the second point cloud data into the grasping quality assessment network so that the grasping quality assessment network outputs the optimal grasping posture.

[0080] In one embodiment, the optimal grasping posture output module 202 is used to construct a grasping quality assessment network based on the grasping posture sampling data, specifically including: obtaining a grasping posture estimation data set, performing grasping posture sampling on each model in the grasping posture estimation data set, and scoring each grasping posture according to a preset quality assessment standard, and constructing a grasping quality assessment network based on PointNet, wherein the grasping quality assessment network includes a multi-layer perception machine, a maximum pooling layer and a fully connected layer.

[0081] In one embodiment, the point cloud data processing module 201 is used to preprocess the first point cloud data, wherein the preprocessing includes point cloud through filtering processing; by setting coordinate axis thresholds in various directions of the spatial coordinate axis, the first point cloud data is segmented according to the coordinate thresholds in various directions to obtain the area of ​​interest of the first point cloud data, and the target point cloud is identified in the area of ​​interest to generate through filtered point cloud data.

[0082] In one embodiment, the point cloud data processing module 201 is used to preprocess the first point cloud data, wherein the preprocessing also includes stage plane segmentation processing; based on a random sampling consistency algorithm, an approximate plane of a background point cloud in the through-filtered point cloud data is estimated, and based on the approximate plane of the background point cloud, the through-filtered point cloud data is segmented into a target object point cloud and a background point cloud to obtain a target object point cloud.

[0083] In one embodiment, the point cloud data processing module 201 is used to preprocess the first point cloud data, wherein the preprocessing also includes point cloud downsampling processing; create a three-dimensional voxel grid based on the target object point cloud, wherein the size of the three-dimensional voxel grid is determined by the resolution and sampling density of the target object point cloud; place the target object point cloud into the three-dimensional voxel grid accordingly to generate multiple small cubes; calculate the distance between each target object point cloud in each small cube and the center of gravity point based on the center of gravity points of the multiple small cubes, and retain the target object point cloud with the smallest distance.

[0084] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0085] It should be noted that the above-mentioned embodiment of the robot grasping posture estimation device is only illustrative, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical units, that is, they may be located in one place, or may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment.

[0086] Based on the above-mentioned embodiment of the robot grasping pose estimation method, another embodiment of the present invention provides a robot grasping pose estimation terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the robot grasping pose estimation method of any embodiment of the present invention is implemented.

[0087] Exemplarily, in this embodiment, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program in the robot grasping posture estimation terminal device.

[0088] The robot grasping posture estimation terminal device may be a computing device such as a desktop computer, a notebook, a palm computer, a cloud server, etc. The robot grasping posture estimation terminal device may include, but is not limited to, a processor and a memory.

[0089] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the robot grasping posture estimation terminal device, and uses various interfaces and lines to connect various parts of the entire robot grasping posture estimation terminal device.

[0090] The memory can be used to store the computer program and / or module, and the processor realizes various functions of the robot grasping posture estimation terminal device by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function, etc.; the data storage area can store data created according to the use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0091] Based on the above-mentioned embodiment of the robot grasping pose estimation method, another embodiment of the present invention provides a storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the storage medium is located is controlled to execute the robot grasping pose estimation method of any embodiment of the present invention.

[0092] In this embodiment, the storage medium is a computer-readable storage medium, and the computer program includes computer program code, which may be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0093] In summary, the present invention provides a robot grasping posture estimation method, device, equipment and storage medium, which obtains the first point cloud data of the target object, pre-processes the first point cloud data, and obtains the second point cloud data of the target object; constructs a grasping quality assessment network according to the grasping posture sampling data, and inputs the second point cloud data into the grasping quality assessment network, so that the grasping quality assessment network outputs the optimal grasping posture. Compared with the prior art, the robot grasping posture estimation can be performed directly based on the three-dimensional point cloud, and the position of the grasped target object in the three-dimensional space can be accurately obtained. By directly processing the first point cloud data, the original point cloud information is kept to the maximum extent, and the grasping posture is estimated more effectively and accurately.

[0094] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention. These improvements and substitutions should also be regarded as the scope of protection of the present invention.

Claims

1. A robot grasping posture estimation method, characterized in that: include: Acquire first point cloud data of a target object, and preprocess the first point cloud data to obtain second point cloud data of the target object; According to the grasping posture sampling data, a grasping quality assessment network is constructed, and the second point cloud data is input into the grasping quality assessment network so that the grasping quality assessment network outputs an optimal grasping posture; The method of constructing a grasping quality evaluation network based on grasping posture sampling data specifically includes: obtaining a grasping posture estimation data set, performing grasping posture sampling on each model in the grasping posture estimation data set, and scoring each grasping posture according to a preset quality evaluation standard, and constructing a grasping quality evaluation network based on PointNet, wherein the grasping quality evaluation network includes a multi-layer perceptron, a maximum pooling layer, and a fully connected layer; wherein the first point cloud data is preprocessed, wherein the preprocessing includes point cloud through filtering processing; setting coordinate axis thresholds in various directions of the spatial coordinate axis, segmenting the first point cloud data according to the coordinate axis thresholds in various directions, obtaining a region of interest of the first point cloud data, and performing target point cloud recognition on the region of interest to generate through-filtered point cloud data; Among them, the first point cloud data is preprocessed, wherein the preprocessing also includes stage plane segmentation processing; based on a random sampling consistency algorithm, the background point cloud approximate plane in the through-filtered point cloud data is estimated, and based on the background point cloud approximate plane, the through-filtered point cloud data is segmented into a target object point cloud and a background point cloud to obtain a target object point cloud.

2. A robot grasping posture estimation method as claimed in claim 1, characterized in that: Preprocessing the first point cloud data, wherein the preprocessing further includes point cloud downsampling processing; Creating a three-dimensional voxel grid according to the target object point cloud, wherein the size of the three-dimensional voxel grid is determined by the resolution and sampling density of the target object point cloud; Placing the target object point cloud into the three-dimensional voxel grid to generate a plurality of small cubes; According to the center of gravity points of the plurality of small cubes, the distance between each target object point cloud in each small cube and the center of gravity point is calculated, and the target object point cloud with the smallest distance is retained.

3. A robot grasping posture estimation device, characterized in that: include: Point cloud data processing module and optimal grasping posture output module; The point cloud data processing module is used to obtain first point cloud data of the target object, pre-process the first point cloud data, and obtain second point cloud data of the target object; The optimal grasping posture output module is used to construct a grasping quality evaluation network according to the grasping posture sampling data, and input the second point cloud data into the grasping quality evaluation network so that the grasping quality evaluation network outputs the optimal grasping posture; The method of constructing a grasping quality evaluation network based on grasping posture sampling data specifically includes: obtaining a grasping posture estimation data set, performing grasping posture sampling on each model in the grasping posture estimation data set, and scoring each grasping posture according to a preset quality evaluation standard, and constructing a grasping quality evaluation network based on PointNet, wherein the grasping quality evaluation network includes a multi-layer perceptron, a maximum pooling layer, and a fully connected layer; wherein the first point cloud data is preprocessed, wherein the preprocessing includes point cloud through filtering processing; setting coordinate axis thresholds in various directions of the spatial coordinate axis, segmenting the first point cloud data according to the coordinate axis thresholds in various directions, obtaining a region of interest of the first point cloud data, and performing target point cloud recognition on the region of interest to generate through-filtered point cloud data; Among them, the first point cloud data is preprocessed, wherein the preprocessing also includes stage plane segmentation processing; based on a random sampling consistency algorithm, the background point cloud approximate plane in the through-filtered point cloud data is estimated, and based on the background point cloud approximate plane, the through-filtered point cloud data is segmented into a target object point cloud and a background point cloud to obtain a target object point cloud.

4. A terminal device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the robot grasping pose estimation method according to any one of claims 1 to 2 is implemented.

5. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the robot grasping pose estimation method according to any one of claims 1 to 2.

Images