Data labeling method suitable for six-degree-of-freedom grabbing network
Through the diversified sampling and grab stability detection algorithm combined with Monte Carlo importance sampling, the problem of inaccurate labeling and inefficiency in the production of six-degree-of-freedom robot crawling data sets is solved, and efficient and real data set generation is achieved, which is suitable for a variety of robot crawling scenarios.
Patent Information
- Application Number
- CN202510930084.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-07
- Publication Date
- 2025-08-15
AI Technical Summary
The existing robot data set production method is mainly based on four degrees of freedom, which cannot meet the flexibility and adaptability needs of six degrees of freedom robots. The existing methods have problems such as inaccurate labeling, inefficient and untrue data, resulting in weak generalization capabilities.
A diversified sampling algorithm is used to obtain the model contact points, a two-finger and multi-finger grasp stability detection algorithm is designed, and a grab position is generated by combining Monte Carlo importance sampling, and collision detection between the fixture and the model is carried out to achieve fully automatic labeling.
It improves the production efficiency of the six-degree-of-freedom crawling data set and the authenticity of the data set, enhances the generalization ability in practical applications, and is suitable for a variety of robot crawling scenarios.
Smart Images
Figure CN120480925A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of robotics technology, and in particular to a data labeling method suitable for a six-degree-of-freedom crawling network. Background Art
[0002] In robotics, the generation of grasping datasets is crucial for optimizing robotic grasping capabilities. Existing methods for generating grasping datasets primarily rely on four-degree-of-freedom grasping, generating grasping poses from two-dimensional images. This approach meets the basic requirements of robotic grasping to a certain extent, but in practice, with the advancement of robotics technology, six-degree-of-freedom grasping is more suitable for practical applications. Therefore, improving existing grasping dataset generation methods and enhancing dataset quality have become urgent challenges in robotics.
[0003] However, the prior art has the following problems and shortcomings:
[0004] 1. Limitations of grasping pose generation methods: Existing grasping pose generation methods are mainly based on four-degree-of-freedom grasping, which cannot meet the grasping requirements of six-degree-of-freedom robots. In practical applications, six-degree-of-freedom grasping has greater flexibility and adaptability, and is more consistent with actual human operations.
[0005] 2. Insufficient dataset preparation methods: Existing methods for preparing six-DOF grasping datasets have the following problems:
[0006] (1) Manual labeling: The data scenario is simple, the labels are imprecise, and the subjectivity is high, resulting in low quality of the dataset.
[0007] (2) Machine collection: low efficiency, non-repeatability, and difficulty meeting the needs of large-scale data sets.
[0008] (3) Simulation collection: There are certain discrepancies with the real scene, resulting in the weak generalization ability of the dataset in practical applications.
[0009] Therefore, a new technical solution is urgently needed to solve the above technical problems. Summary of the Invention
[0010] The purpose of the present invention is to overcome the problems of the above-mentioned prior art and provide a data labeling method suitable for six-degree-of-freedom grasping networks, so as to solve the technical problems existing in the existing six-degree-of-freedom robot grasping data set production method, such as inaccurate labels, low production efficiency and untrue data, which lead to weak generalization ability in applications.
[0011] The above objectives are achieved through the following technical solutions:
[0012] A data annotation method for a six-degree-of-freedom grasping network, comprising:
[0013] Step (1) obtaining 3D model data of the object to be grasped;
[0014] Step (2) performing a diversified sampling algorithm on the acquired 3D model data to obtain the contact points of the model, and calculating the center of the contact points as the grasping points;
[0015] Step (3) determines whether the input grasping point is a reliable grasping point through a grasping stability detection algorithm. If so, output it until a preset total number is reached; otherwise, repeat step (2); the grasping stability detection algorithm is a detection algorithm constructed based on the force closure principle, including multi-finger and dual-finger detection algorithms;
[0016] Step (4) Generate a corresponding set of grasping poses for each stable grasping point based on the Monte Carlo importance sampling algorithm;
[0017] Step (5) performs collision detection between the fixture and the model for all grasping postures generated in step (4);
[0018] Step (6) outputs and saves the collision-free grasping pose as the final grasping pose of the model.
[0019] As an optimization of this method, the diversified sampling algorithms in step S2 include uniform sampling, Gaussian sampling and counter sampling.
[0020] As an optimization of this method, in step (2), the center of each pair of contact points is calculated as the grasping point after filtering the contact points, and the filtering method is to set the distance threshold according to the fixture size.
[0021] As an optimization of this method, the method of determining whether the input grasping point is a reliable grasping point through the grasping stability detection algorithm described in step (3) includes: manually selecting the clamp type for different application requirements; and applying different grasping stability detection algorithms for different clamp types.
[0022] As an optimization of this method, the grasping stability detection algorithm for the two-finger gripper specifically includes:
[0023] Step (3-101) traverses the two contact points and calculates the relative position vector D between the contact points. The calculation formula is as follows:
[0024] D=c1-c2
[0025] Where D is the relative position vector, c1 is the first contact point object, which contains the position and normal vector of the contact point; c2 is the second contact point object, which contains the position and normal vector of the contact point;
[0026] Step (3-102) calculates the corresponding normal projection N. If the projection direction is negative, the grab point is deleted;
[0027] Step (3-103) calculates the friction angle α of the retained grasping point, and the calculation formula is as follows:
[0028]
[0029] Among them, α is the friction angle, N is the normal projection, and D is the relative position vector;
[0030] In step (3-104), the friction coefficient μ is selected to calculate the friction cone angle β. The calculation formula is as follows:
[0031] β=arctan(μ)
[0032] If the friction angle a of the gripping point is not less than β, the gripping point is output.
[0033] As an optimization of this method, the grasping stability detection algorithm for multi-finger grippers specifically includes:
[0034] Step (3-201) For each contact point, calculate the force F and torque τ; where F is the normal force F n and tangential force F t1 、F t2 The calculation formula is as follows:
[0035] F n =N
[0036] F t1 =μ*t1
[0037] F t2 =μ*t2
[0038] τ=P*F
[0039] Where μ is the friction coefficient, t1 is the first tangent vector generated by the cross product of the normal vector N, t2 is another orthogonal tangent vector generated by the cross product of N and t1; P is the position vector of the contact point.
[0040] Step (3-202) combines the wrench vectors composed of the force and torque of each contact point into the wrench matrix W, and the calculation formula is as follows:
[0041]
[0042] Step (3-203) calculates the rank of the spliced wrench matrix. If the rank is 6, the grasping point is output.
[0043] As an optimization of this method, the friction coefficient μ is set to 0.8.
[0044] As an optimization of this method, step (4) specifically includes:
[0045] Step (401) uses the KNN algorithm to establish a correspondence between the grasping points output in step (3) and the original point cloud;
[0046] Step (402) calculates the normal vector of each grasp point in the original point cloud and evaluates the importance of each potential grasping perspective using the importance function f. The calculation formula of function f is as follows:
[0047]
[0048] in, is the spherical coordinate, θ is the azimuth, is the polar angle, N is the normal vector, and ω is the angle between the viewing direction vector and the normal vector;
[0049] Step (403) uses the Monte Carlo method to perform importance sampling, generating a total of N p The generated grasping perspective is rotated at different angles on the z-axis for each grasping perspective, and the grasping posture is generated and output.
[0050] As an optimization of this method, the N p The total number of preset capture angles.
[0051] As an optimization of this method, the total number of gripping points preset in step (3) is 100.
[0052] The present invention provides a data labeling method suitable for a six-degree-of-freedom grasping network. It uses a diversified sampling algorithm to obtain model grasping points and designs two sets of grasping stability detection algorithms to perform grasping point stability tests. It is applicable to two-finger parallel grippers and multi-finger grippers. Monte Carlo importance sampling is designed based on the grasping characteristics to generate grasping postures. Finally, collision detection is performed on the gripper and the model to achieve fully automatic labeling of grasping postures and improve the efficiency of data set production. This method is applicable to various robot grasping scenarios such as two-finger parallel grippers and multi-finger grippers. It can meet the grasping label requirements of the six-degree-of-freedom grasping algorithm and has strong versatility. The specific advantages are as follows:
[0053] 1. Strong versatility: Two sets of grasping stability detection algorithms are designed, which are applicable to two-finger parallel grippers and multi-finger grippers to meet the needs of various robot grasping scenarios.
[0054] 2. Increased number of effective grasps: The Monte Carlo-based importance sampling algorithm considers the geometric features of the object surface, optimizes the grasping perspective selection, and significantly increases the number of effective grasps.
[0055] 3. Efficient and strong generalization ability: Fully automatic labeling is achieved based on the geometric features and contact information of objects, which is efficient and repeatable. The generated dataset is highly consistent with the real scene and has strong generalization ability in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] Figure 1 This is a flowchart of the steps of a data annotation method applicable to a six-degree-of-freedom crawling network according to the present invention;
[0057] Figure 2 This is a flowchart of a data annotation method for a six-degree-of-freedom crawling network according to the present invention;
[0058] Figure 3 This is a schematic diagram of grabbing label generation for a data annotation method applicable to a six-degree-of-freedom grabbing network according to the present invention. DETAILED DESCRIPTION
[0059] The present invention will be further described in detail below with reference to the accompanying drawings and examples. The described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort shall fall within the scope of protection of the present invention.
[0060] like Figure 1 As shown, this solution provides a data annotation method suitable for six-degree-of-freedom grasping networks, including:
[0061] Step (1) obtaining 3D model data of the object to be grasped, wherein the 3D model data is 3D model data obtained after being scanned by a device such as a 3D camera or a laser radar;
[0062] Step (2) performs a diversified sampling algorithm on the acquired 3D model data to obtain the contact points of the model, and calculates the center of the contact point as the grasping point; the diversified sampling algorithm includes but is not limited to uniform sampling, Gaussian sampling and counter sampling, etc.
[0063] Step (3) determines whether the input grasping point is a reliable grasping point through a grasping stability detection algorithm. If so, output it until a preset total number is reached; otherwise, repeat step (2). The grasping stability detection algorithm is a detection algorithm constructed based on the force closure principle, including multi-finger and dual-finger detection algorithms;
[0064] Step (4) generates a corresponding set of grasping postures for each stable grasping point based on the Monte Carlo importance sampling algorithm, wherein the Monte Carlo importance sampling algorithm is a grasping perspective generation method based on normal vector importance sampling;
[0065] Step (5) performs collision detection between the fixture and the model for all grasping postures generated in step (4);
[0066] Step (6) outputs and saves the collision-free grasping pose as the final grasping pose of the model.
[0067] In step (2) of this embodiment, the contact points are filtered and the center of each pair of contact points is calculated as the grasping point. The filtering method is to set a distance threshold according to the fixture size.
[0068] The method described in step (3) of this embodiment for determining whether the input grasping point is a reliable grasping point through a grasping stability detection algorithm includes: manually selecting the clamp type for different application requirements; and applying different grasping stability detection algorithms for different clamp types (multi-finger and dual-finger).
[0069] Among them, for the grasping stability detection algorithm of the two-finger gripper, the specific steps include:
[0070] Step (3-101) traverses the two contact points and calculates the relative position vector D between the contact points. The calculation formula is as follows:
[0071] D=c1-c2
[0072] Where D is the relative position vector, c1 is the first contact point object, which contains the position and normal vector of the contact point; c2 is the second contact point object, which contains the position and normal vector of the contact point;
[0073] Step (3-102) calculates the corresponding normal projection N, which represents the component of the relative position between the two contact points in the normal direction; if the projection direction is negative, it means that the contact point is in the opposite direction of the normal, indicating that the force closure condition is not met, and the grasping point is deleted;
[0074] Step (3-103) calculates the friction angle α of the retained grasping point using the inverse cosine function, which is calculated by the projection and the length of the direction vector. This angle represents the angle between the normal and the line connecting the two contact points. The calculation formula is as follows:
[0075]
[0076] Among them, ɑ is the friction angle, N is the normal projection, and D is the relative position vector;
[0077] In step (3-104), a smaller friction coefficient μ is selected to calculate the friction cone angle β. The calculation formula is as follows:
[0078] β=arctan(μ)
[0079] Compare the friction angle a of the grasping point with the friction cone angle β. If the friction angle a of the grasping point is less than β, it is insufficient to form a force seal. Conversely, if the friction angle a of the grasping point is not less than β, it means that the force seal condition is met and the grasping point is output.
[0080] For the grasping stability detection algorithm of multi-finger grippers, the specific steps include:
[0081] Step (3-201) For each contact point, calculate the force F and torque τ; where F is the normal force F n and tangential force F t1 、F t2 The calculation formula is as follows:
[0082] F n =N
[0083] F t1 =μ*t1
[0084] F t2 =μ*t2
[0085] τ=P*F
[0086] Where μ is the friction coefficient, t1 is the first tangent vector generated by the cross product of the normal vector N (or other orthogonal operation), t2 is another orthogonal tangent vector generated by the cross product of N and t1; P is the position vector of the contact point.
[0087] Step (3-202) combines the wrench vectors composed of the force and torque of each contact point into the wrench matrix W, and the calculation formula is as follows:
[0088]
[0089] Step (3-203) calculates the rank of the spliced wrench matrix. If the rank is 6, the force closure condition is met and the grasping point is output.
[0090] The specific steps of step (4) in this embodiment include:
[0091] Step (401) uses the KNN algorithm to establish a correspondence between the grasping points output in step (3) and the original point cloud;
[0092] Step (402) calculates the normal vector of each grasp point in the original point cloud and evaluates the importance of each potential grasping perspective using the importance function f. The calculation formula of function f is as follows:
[0093]
[0094] in, is the spherical coordinate, θ is the azimuth, is the polar angle, N is the normal vector, and ω is the angle between the viewing direction vector and the normal vector;
[0095] Step (403) uses the Monte Carlo method to perform importance sampling, generating a total of N pGrasping viewpoints, and then generating each grasping viewpoint to rotate at different angles on the z axis, generating and outputting the grasping posture. p The total number of preset capture angles.
[0096] This embodiment adopts an importance sampling algorithm based on Monte Carlo, takes into account the geometric features of the object surface, optimizes the selection of grasping perspective, greatly increases the number of effective grasps, and the generated grasping perspective is more in line with actual grasping needs.
[0097] like Figure 2 As shown, as a specific embodiment of this solution, a data annotation method suitable for a six-degree-of-freedom crawling network is provided, which includes the following steps:
[0098] Step (1) Use a 3D camera to scan and photograph the object being grasped to obtain 3D model data of the object. Repeat this operation to obtain data for different objects one by one.
[0099] After obtaining the data collected in step (1), step (2) first uniformly samples the acquired 3D model data to obtain the contact points of the model, sets the distance threshold according to the fixture size, filters the contact points, and finally calculates the center of each pair of contact points as the grasping point.
[0100] Step (3) determines whether the input grasping point is a reliable grasping point through a two-finger grasping stability detection algorithm. If so, output it until a preset total number is reached, otherwise repeat step (2); specifically:
[0101] First, traverse the two contact points and calculate the relative position vector D between the contact points. The calculation formula D is as follows:
[0102] D=c1-c2
[0103] Where D is the relative position vector, c1 is the first contact point object, which contains the position and normal vector of the contact point, and c2 is the second contact point object, which contains the position and normal vector of the contact point.
[0104] Then, the corresponding normal projection N is calculated. This projection represents the normal component of the relative position between the two contact points. If the projection direction is negative, it means that the contact point is in the opposite direction of the normal, indicating that the force closure condition is not met and the grip point is deleted.
[0105] Next, calculate the friction angle ɑ of the remaining grip point using the inverse cosine function using the projection and the length of the direction vector. This angle represents the angle between the normal and the line connecting the two contact points. The calculation formula for ɑ is as follows:
[0106]
[0107] Among them, ɑ is the friction angle, N is the normal projection, and D is the relative position vector.
[0108] Finally, the friction coefficient μ = 0.8 is selected to calculate the friction cone angle β. The specific β calculation formula is as follows:
[0109] β=arctan(μ)
[0110] Compare the friction angle α of the grasping point with the friction cone angle β. If the friction angle α of the grasping point is less than β, it is insufficient to form a force seal. Otherwise, it means that the force seal condition is met and the grasping point is output. Calculate each grasping point in turn until the preset 100 grasping points are met. Otherwise, repeat step (2) for sampling.
[0111] Step (4) generates a corresponding set of grasping postures for each stable grasping point based on the Monte Carlo importance sampling algorithm. Specifically:
[0112] First, the grasping points output in step (3) are used to establish the correspondence between the grasping points and the original point cloud through the KNN algorithm.
[0113] Next, for each grasped point in the original point cloud, its normal vector is calculated. The importance of each potential grasping perspective is evaluated using the importance function f. Specifically, the calculation formula of function f is as follows:
[0114]
[0115] in, is the spherical coordinate, θ is the azimuth, is the polar angle, N is the normal vector, and ω is the angle between the viewing direction vector and the normal vector.
[0116] Then the Monte Carlo method is used for importance sampling to generate a total number N p = 50. Then, each grasping view is rotated at different angles on the z-axis to generate and output the grasping pose.
[0117] Step (5) performs collision detection between the fixture and the model for all grasping postures generated in step (4).
[0118] Step (6), output and save the collision-free grasping posture as the final grasping posture of the model. Figure 3 As shown in Figure 2, effective multi-view grasping poses are generated for a single model and a single grasping point.
[0119] The above description is only for explaining the embodiments of the present invention and is not intended to limit the present invention. For those skilled in the art, any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A data annotation method for a six-degree-of-freedom crawling network, characterized in that: include: Step (1) obtaining 3D model data of the object to be grasped; Step (2) performing a diversified sampling algorithm on the acquired 3D model data to obtain the contact points of the model, and calculating the center of the contact points as the grasping points; Step (3) determines whether the input grasping point is a reliable grasping point through a grasping stability detection algorithm. If so, output it until a preset total number is reached; otherwise, repeat step (2); the grasping stability detection algorithm is a detection algorithm constructed based on the force closure principle, including multi-finger and dual-finger detection algorithms; Step (4) Generate a corresponding set of grasping poses for each stable grasping point based on the Monte Carlo importance sampling algorithm; Step (5) performs collision detection between the fixture and the model for all grasping postures generated in step (4); Step (6) outputs and saves the collision-free grasping pose as the final grasping pose of the model.
2. A data annotation method for a six-degree-of-freedom crawling network according to claim 1, characterized in that: The diversified sampling algorithms in step S2 include uniform sampling, Gaussian sampling and counter sampling.
3. A data annotation method for a six-degree-of-freedom crawling network according to claim 1 or 2, characterized in that: In step (2), the contact points are filtered and the center of each pair of contact points is calculated as the grasping point. The filtering method is to set the distance threshold according to the fixture size.
4. A data annotation method for a six-degree-of-freedom crawling network according to claim 1, characterized in that: The method of determining whether the input grasping point is a reliable grasping point by grasping stability detection algorithm described in step (3) includes: manually selecting the clamp type for different application requirements; and applying different grasping stability detection algorithms for different clamp types.
5. A data annotation method for a six-degree-of-freedom crawling network according to claim 4, characterized in that: The grasping stability detection algorithm for the two-finger gripper specifically includes: Step (3-101) traverses the two contact points and calculates the relative position vector D between the contact points. The calculation formula is as follows: D=c1-c2 Where D is the relative position vector, c1 is the first contact point object, which contains the position and normal vector of the contact point; c2 is the second contact point object, which contains the position and normal vector of the contact point; Step (3-102) calculates the corresponding normal projection N. If the projection direction is negative, the grab point is deleted; Step (3-103) calculates the friction angle α of the retained grasping point, and the calculation formula is as follows: Among them, α is the friction angle, N is the normal projection, and D is the relative position vector; In step (3-104), the friction coefficient μ is selected to calculate the friction cone angle β. The calculation formula is as follows: β=arctan(μ) If the friction angle a of the gripping point is not less than β, the gripping point is output.
6. A data annotation method for a six-degree-of-freedom crawling network according to claim 4 or 5, characterized in that: The grasping stability detection algorithm for multi-finger grippers specifically includes: Step (3-201) For each contact point, calculate the force F and torque τ; where F is the normal force F n and tangential force F t1 、F t2 The calculation formula is as follows: F n =N F t1 =μ*t1 F t2 =μ*t2 τ=P*F Where μ is the friction coefficient, t1 is the first tangent vector generated by the cross product of the normal vector N, t2 is another orthogonal tangent vector generated by the cross product of N and t1; P is the position vector of the contact point. Step (3-202) combines the wrench vectors composed of the force and torque of each contact point into the wrench matrix W, and the calculation formula is as follows: Step (3-203) calculates the rank of the spliced wrench matrix. If the rank is 6, the grasping point is output.
7. A data annotation method for a six-degree-of-freedom crawling network according to claim 6, characterized in that: The friction coefficient μ is set to 0.
8.
8. The data annotation method for a six-degree-of-freedom crawling network according to claim 1, characterized in that: The step (4) specifically includes: Step (401) uses the KNN algorithm to establish a correspondence between the grasping points output in step (3) and the original point cloud; Step (402) calculates the normal vector of each grasp point in the original point cloud and evaluates the importance of each potential grasping perspective using the importance function f. The calculation formula of function f is as follows: in, is the spherical coordinate, θ is the azimuth, is the polar angle, N is the normal vector, and ω is the angle between the viewing direction vector and the normal vector; Step (403) uses the Monte Carlo method to perform importance sampling, generating a total of N p The generated grasping perspective is rotated at different angles on the z-axis for each grasping perspective, and the grasping posture is generated and output.
9. A data annotation method for a six-degree-of-freedom crawling network according to claim 8, characterized in that: The N p The total number of preset capture angles.
10. The data annotation method for a six-degree-of-freedom crawling network according to claim 1, characterized in that: The preset total number of gripping points in step (3) is 100.