Circular suction cup gripping method and system

Abstract

The application discloses a circular suction cup grabbing method and system, belongs to the technical field of robot vision grabbing and three-dimensional point cloud deep learning, and solves the technical defects of existing circular suction cup grabbing algorithms, such as posture redundancy, poor rotation robustness, large normal prediction error, only single-point judgment of adsorption, and score susceptible to noise interference. Relying on the axisymmetric geometric characteristics of the circular suction cup, the grabbing pose is simplified to the adsorption center and the directed approach axis, and the yaw angle around the axis without physical action is discarded; a multi-order spherical harmonic isovar encoder is used to extract point cloud l0 / l1 / l2 hierarchical features, and the network directly predicts the suction cup approach axis; a cylindrical coordinate system is established based on the predicted axis, and sealing evaluation indexes such as height variance and circular coverage are calculated based on the real sealing ring size partition of the suction cup; the isovar features are contracted into global rotation invariant features along the suction cup axis, and a stable adsorption score is solved by inputting a conditional diffusion model; and finally, the optimal grabbing pose is screened through joint non-maximum suppression, collision and inverse kinematics.

Description

Technical Field

[0001] This invention relates to the fields of robot intelligent grasping, three-dimensional machine vision, point cloud deep learning, and automated sorting control of vacuum suction cups, and particularly to a circular suction cup grasping method and system. Background Technology

[0002] Existing circular suction cup visual grasping algorithms have five major technical shortcomings: First, there are redundant degrees of freedom in the attitude space: the rotation of the circular suction cup around its own axis does not change the lip sealing contact state, but the existing network forces the regression of the complete six degrees of freedom attitude, introduces a yaw angle with no physical meaning, the same grasping effect corresponds to multiple sets of labels, the training supervision conflict is large, and the model convergence speed is slow. Secondly, the model has poor robustness to rotation: conventional PointNet-like networks do not have rotational equivariance properties. After the scene rotates, the prediction results of the adsorption direction and grasping score drift, requiring massive rotation data for enhancement. It also has weak generalization ability for unknown orientations and new types of workpieces. Thirdly, the normal prediction error propagation: existing solutions first calculate the surface normal separately through plane fitting and general geometry library, and then feed it into the grasping network. The normal in sparse point cloud, reflective, and occluded areas jitters violently, and the error propagates downward, causing the grasping axis prediction to fail. Fourth, judging the feasibility of adsorption based on a single point: scoring based on the point cloud features of a single candidate point without verifying the flatness of the surrounding curved surface according to the actual sealing ring radius of the suction cup, failing to identify circumferential gaps and steps, resulting in a large number of false high-scoring candidates that seem feasible but actually leak air. Fifth, the adsorption scores lack contextual constraints: the scores regress independently point by point, ignoring the spatial correlation between adjacent candidate points, point cloud missing, and local similar regions are prone to generating isolated false high scores, and there is no smoothing and denoising mechanism, resulting in low capture and screening accuracy.

[0003] Existing technologies cannot integrate the four core concepts of suction cup axisymmetric geometric constraints, rotational isovariant deep learning, global geometric verification of sealing rings, and context diffusion scoring. This results in low success rates for grasping in industrial settings and high computational costs for model deployment. Therefore, this invention proposes a complete technical solution. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing circular suction cup grasping algorithms, such as attitude redundancy, inconsistent scene rotation prediction, large normal estimation error, single-point judgment of seal failure, and strong noise interference of adsorption fraction. It provides a circular suction cup grasping method and system based on axisymmetric isovariant geometric representation and contact ring analysis. By relying on the axisymmetric characteristics of the suction cup to eliminate useless yaw degrees of freedom, and employing a spherical harmonic tensor product isovariant network to ensure feature consistency under any scene rotation, the network directly outputs the grasping proximity axis. Based on the actual sealing dimensions of the suction cup, the flatness of the sealing ring is verified across the entire domain. Through rotation-invariant features combined with a conditional diffusion model, a stable adsorption fraction is smoothly output, significantly improving the robustness and success rate of suction cup grasping in cluttered industrial scenes.

[0005] In a first aspect, this embodiment provides a circular suction cup grasping method, including the following steps: S1. Point Cloud Acquisition and Candidate Point Preprocessing: Acquire scene point clouds using 3D imaging equipment. The original point cloud is processed by working space pruning, outlier removal, voxel downsampling, and farthest point sampling to obtain a set of spatial candidate points; the radius r of the circular suction cup sealing ring is pre-calibrated. s lip width w s Tool height and perimeter safety distance are used as fixed constants of the system; S2. Multi-scale rotational isotropic geometric coding: The local space of candidate points is divided into multi-scale spherical neighborhoods. A spherical harmonic tensor product isotropic encoder is used to extract hierarchical features of the relative coordinates of the point cloud, outputting l0 scalar features, l1 vector features, and l2 second-order tensor features. When the scene undergoes arbitrary spatial rotation, the l0 features remain unchanged, the l1 features rotate synchronously, and the l2 tensor satisfies Q'=RQR. T ; S3. Suction Cup Proximity Axis Prediction: Based on the features of each candidate point l0, l1, and l2, gating weights are constructed. The multi-channel l1 vectors are weighted, synthesized, and normalized to obtain the directed proximity axis a of the suction cup. i Synchronous output axis confidence q i Introducing the camera's line-of-sight vector eliminates the ambiguity of the normal's positive or negative sign, outputting only the adsorption center p. i With near axis a i It does not return to the yaw angle around the axis; S4, Geometric analysis of axisymmetric contact ring: with (p i ,a i Establish a local cylindrical coordinate system and calculate the axial height z of the neighboring points. ij Radial distance ρ ij Based on the suction cup calibration parameters, the central area, sealing ring area, and outer safety area are divided. Gaussian soft weighting is used to calculate the weighted mean height, height variance, and circumferential harmonic coverage amplitude of the sealing ring to characterize the continuity of the suction cup seal. S5. Rotation-Invariant Feature Compression: The l1 vector and the l2 second-order tensor are subjected to axial projection, tangential norm, and tensor anisotropy contraction operations along the prediction suction cup axis. Combined with the sealing ring statistical index, a globally rotation-invariant conditional feature F is formed. i F under arbitrary scene rotation i The value remains constant; S6, Scalar Conditional Diffusion-Adsorption Score: Based on rotationally invariant characteristic F i Given the condition, a diffusion denoising model is used to iteratively solve for the scalar score s of adsorption feasibility at each candidate point. i Only neighborhood context diffusion is performed on the adsorption fraction; diffusion updates are not performed on the suction cup proximity axis or yaw angle. S7. Candidate Grasping and Screening and Robot Execution: Based on Adsorption Score s i , axis confidence q i The sealing ring integrity and peripheral collision risk are comprehensively ranked, and a position-direction joint NMS is used to screen the Top-K candidates. For each candidate, within the yaw equivalence interval without changing the sealing contact, the optimal yaw angle that minimizes the collision cost and joint cost is solved. i ,a i The optimal yaw angle is converted into the robot's end-effector's six-degree-of-freedom grasping pose and executed.

[0006] In one embodiment, in step S2, the multi-scale spherical neighborhood is selected at a scale of 0.5r. s 1.0r s 2.0r s 4.0r s The equivariant encoder is based on the PointNet++ hierarchical architecture, and uses the spherical harmonic function Ylr and tensor product operation to realize equivariant message aggregation. Each layer outputs three types of hierarchical features: l0, l1, and l2.

[0007] In one embodiment, in step S3, the formula for calculating the gating weight is as follows: Weighted composite vector Normalization yields suction cups close to the axis .

[0008] In one embodiment, in step S4, the radial interval of the sealing ring satisfies Soft radial weight ; through the m-th order circular harmonic amplitude A i,m Characterizes circumferential gaps and local discontinuities in the sealing ring.

[0009] In one embodiment, rotation-invariant compression in step S5 includes: axial projection of vector l1. Tangential residual norm ; l2 tensor axial response Anisotropic tensor norm of the tangent plane The above indicators are concatenated with the sealing ring variance and coverage amplitude to obtain the fixed-dimensional feature F. i .

[0010] In one embodiment, in step S6, the loss function during the diffusion model training phase includes axis loss, confidence loss, sealing supervision loss, equivariance consistency loss, and diffusion denoising loss; during the inference phase, DDIM is used to generate adsorption scores through few-step sampling, which can be replaced by a direct scalar regression network in scenarios with limited computing power.

[0011] In one embodiment, in step S7, the location-direction joint NMS distance calculation formula is as follows: The yaw angle optimization objective is .

[0012] Secondly, this embodiment also provides a circular suction cup grasping system for performing the circular suction cup grasping method described in the first aspect, including: The point cloud acquisition module is used to acquire 3D point clouds of the scene and perform preprocessing such as cropping and downsampling. A multi-scale rotational equivariant geometry encoding module is used to extract l0 scalar, l1 vector, and l2 second-order tensor hierarchical equivariant features from point clouds; The suction cup axis prediction module is used to calculate the directional proximity axis and axis confidence of the suction cup based on hierarchical isovariant features. The axisymmetric contact ring analysis module is used to construct a cylindrical coordinate system based on the predicted suction cup axis and solve for sealing evaluation parameters such as sealing ring height and circumferential coverage. The rotation invariant compression module is used to shrink the isovariant feature along the chuck axis to generate rotationally invariant conditional features; The scalar conditional diffusion scoring module is used to output an adsorption feasibility score for each candidate point through a diffusion model, based on invariant characteristics. The candidate selection and robot control module is used to jointly solve the optimal six-degree-of-freedom grasping pose by combining nonmaximum suppression, collision detection, and inverse kinematics, and then send the results to the robot to execute the grasping action. Each module is stored in the memory of the edge computing device / robot control cabinet and is scheduled and coordinated by the processor.

[0013] In one embodiment, the point cloud acquisition module is any one of a depth camera, a structured light camera, or a lidar.

[0014] In one embodiment, the circular suction cup gripping system is further configured with a computer-readable storage medium storing a computer program that, when executed by a processor, implements all the steps of the circular suction cup gripping method described in the first aspect.

[0015] The circular suction cup gripping method and system provided by this invention have the following significant advantages compared with the prior art: 1. By eliminating redundant yaw degrees of freedom, training and supervision conflicts are significantly reduced. The grasping pose space is simplified from the complete SE(3) to R 3 ×S 2 There is no issue of multiple sets of labels corresponding to the same crawling effect, and the model converges faster.

[0016] 2. Naturally rotational and equivariant, with strong cross-scenario generalization ability. Based on the spherical harmonic tensor product, a hierarchical isovariant feature is constructed. After arbitrary scene rotation, the adsorption axis and adsorption fraction have theoretical consistency. No large amount of rotation data augmentation is required, and the success rate of grasping unknown workpieces in messy stacks in the warehouse is significantly improved.

[0017] 3. Integrated network predicts and grasps axes, eliminating normal error propagation. By abandoning the external geometric normal estimation process, all directional features are uniformly generated within the isovariant network. Normal jitter will not occur in sparse, reflective, or occluded areas of the point cloud, and the stability of directional prediction is greatly improved.

[0018] 4. Match the suction cup and real sealing ring for full-area verification to accurately identify air leakage defects. The sealing ring area is divided by the suction cup calibration size. The sealing capability of the curved surface is quantified by the height variance and the circumferential coverage amplitude. False candidates with high single-point feature scores but circumferential gaps are effectively filtered out, and the probability of negative pressure grasping failure is greatly reduced.

[0019] 5. Combining rotation-invariant features with diffusion denoising suppresses point cloud noise interference. Axial contraction yields globally constant condition features, while the diffusion model introduces neighborhood context to smooth single-point scores, eliminating point cloud missingness and isolated false high scores caused by local similar regions, thus improving the reliability of candidate selection.

[0020] 6. Post-processing optimization of yaw, balancing algorithm computing power and robot reachability. The network only outputs the core grasping center and approach axis. The yaw angle is optimized in the post-processing collision and inverse kinematics solution stages without increasing the computational power consumption of the network inference, while improving the reachability of the robot's grasping.

[0021] 7. The entire algorithm is deployed in a modular fashion, adaptable to low-computing-power devices at the industrial edge. The encoding, axis prediction, geometric analysis, scoring, and screening modules are independent of each other, and the computing power consumption can be adjusted as needed. Both ordinary industrial robot control cabinets and embedded edge machines can perform real-time inference, making them suitable for large-scale automated sorting production lines. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the overall data flow of the circular suction cup grasping method according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the principle of multi-order rotational equivariant local geometric coding in an embodiment of the present invention. Figure 3 This is a simplified schematic diagram of the axisymmetric grasping posture of the circular suction cup according to an embodiment of the present invention; Figure 4 is an axisymmetric geometric analysis diagram of the circular suction cup contact ring according to an embodiment of the present invention; Figure 5 is a flowchart of the axial rotation-invariant compression process of an embodiment of the present invention; Figure 6 is a flowchart of the algorithm training and inference process in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] It should be noted that when a component is said to be "mounted on" another component, it can be directly mounted on the other component or may be interspersed with a component. When a component is said to be "set on" another component, it can be directly set on the other component or may be interspersed with a component. When a component is said to be "fixed to" another component, it can be directly fixed to the other component or may be interspersed with a component.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0026] Please see Figures 1 to 6 Firstly, this embodiment provides a complete implementation of a circular suction cup grasping method and system.

[0027] This embodiment provides a general methodological concept. For a circular suction cup of fixed specifications, this invention represents the contact pose of the suction cup as the adsorption center and a directed approach axis, without predicting the yaw angle around the approach axis. Let the adsorption center be p. i ∈R 3 The suction cup is close to the axis a i ∈S 2 The contact pose of the circular suction cup belongs to R. 3 *S 2 This is equivalent to eliminating the SO(2) symmetric degrees of freedom of the circular suction cup about its own axis from the pose space of the complete rigid body.

[0028] According to the formula X can be obtained circle .

[0029] It should be noted that, with this setup, this embodiment sequentially performs multi-scale rotational isovariant geometry encoding, suction cup axis prediction, axisymmetric contact ring analysis, invariant compression along the predicted axis, scalar conditional diffusion scoring, and robot executability screening. The sealing ring radius r of the fixed suction cup is... s lip width w s Tool height h s and safe distance r safe Obtained through calibration and used as a system constant.

[0030] Based on the overall methodological concept, the implementation is extended through the following steps: Step S1: Point cloud acquisition and candidate point establishment Acquire scene point clouds using depth cameras, structured light cameras, LiDAR, or multi-view reconstruction devices. Among them, p i For three-dimensional coordinates, c i This can be an optional color, depth confidence, reflection intensity, or semantic scalar. The network convolution only uses relative coordinates p. j -p i This ensures that the features remain equivariant to global translation.

[0031] Perform workspace pruning, outlier removal, voxel downsampling, and farthest point sampling on the point cloud to obtain a candidate point set. Preferably, the neighborhood scale is related to the fixed suction cup radius r. s Associated settings, for example, using 0.5r s 1.0r s 2.0r s and 4.0r s A multi-scale spherical neighborhood is used to simultaneously describe the central micro-undulations, the lip-mouth contact range, and the peripheral barriers.

[0032] Step S2, Multi-scale Rotational Isotropic Geometric Coding A point cloud isotropic encoder with irreducible representation features is employed. The encoder retains the PointNet++-style farthest point sampling, neighborhood query, encoder-decoder, and feature propagation structure, but replaces the ordinary local multilayer perceptron aggregation with spherical harmonic functions and tensor product convolution. For candidate point p... i and neighboring point p j ,definition , , .

[0033] ; Among them, Y （lr） For the l r The order of the spherical harmonic function, W is the learnable radial weight determined by distance, the symbol ⊗ represents the tensor product, and the square bracket subscript indicates the projection to the specified output order.

[0034] Preferably, the output has three types of features: l=0, l=1, and l=2. The l=0 scalar is used to represent flatness, density, and semantics; the l=1 vector is used to represent normal, edge direction, and free space direction; and the l=2 second-order feature is used to represent curvature, groove, and anisotropy.

[0035] ; When the point cloud is rotated by R, the features of each order satisfy the following: scalar h （0） The vector h remains unchanged. （l） Synchronous rotation, second-order tensor Q satisfies Q'=RQR T .

[0036] Therefore, the network does not need to memorize the adsorption direction for each spatial angle.

[0037] Step S3: Predicting the directional suction cup axis A circular suction cup approximation axis is directly generated from the l=1 vector features, without relying on external normal estimation. First, the scalar gating weight g for each vector channel is calculated using scalar features, vector norm, and second-order feature norm. i,c Then, perform a weighted summation of the vector channels: ; ; Since the gating weights are rotation-invariant... Since it is an isovariant vector, the prediction axis satisfies To eliminate the ambiguity of the positive and negative directions of the local surface normal, the observation direction from the camera's optical center to the candidate point, the visible free space direction, or the signed occupancy gradient are used as additional l=1 vector inputs, enabling the network to select the side from which the robot can approach in free space.

[0038] Simultaneously, the axis confidence q is predicted based on scalar features, vector synthesis magnitude, and line-of-sight consistency. i For candidate points with geometric degradation, severe occlusion, or uncertain orientation, q i The lower score reduces its priority in subsequent ratings.

[0039] Step S4: Geometric Analysis of the Axisymmetric Contact Ring With candidate point p i Origin, prediction axis a i Establish local cylindrical coordinates for the axis that do not require in-plane reference directions. For a neighboring point p i ,calculate: ; ; Among them, z ij ρ is the height along the suction cup axis. ij This is the radial distance to the suction cup axis. When the scene rotates as a whole, p... j -p i and a i Synchronous rotation, therefore z ij and ρ ij The parameters remain unchanged. Based on the fixed suction cup parameters, the neighborhood is divided into a central region, a sealing ring region, and a peripheral safety region. The sealing ring region preferably satisfies... .

[0040] To maintain differentiability, we use a method around r. s Soft radial weights: ; Calculate the weighted average height, weighted height variance, and robust maximum residual of the sealing ring based on this weight: ; .

[0041] The sealing ring height variance describes the overall smoothness of the lip coverage area, while the maximum residual is used to detect localized steps, gaps, or protrusions. The central area is used to detect severe protrusions and holes inside the suction cup, while the outer safety area is used to estimate the collision risk to the tool housing and approach path.

[0042] To avoid misjudgments due to a large number of points covering only half a circle, the projection direction of the sealing ring can be represented as an angle. And calculate the amplitudes of several order circular harmonic moments: , A i,mThe amplitude is independent of the choice of the tangent plane reference axis and can be used to describe the degree of circumferential missing sectors and coverage unevenness. In another embodiment, soft-angle binning, point-to-point distance statistics, or closed-loop integrity based on neighborhood graphs can also be used instead of the circumferential harmonic moment.

[0043] Step S5: Compression of rotational invariant along the suction cup axis To provide stability conditions for the scalar scoring network, the equivariant vector and the second-order tensor are placed along the prediction axis a. i Perform contraction. For each vector channel. Calculate the axial projection and tangential energy: ; .

[0044] For second-order feature Q i,c Calculate the axial response, Frobenius norm, and anisotropy of the tangent plane: ; .

[0045] The above shrinkage results are concatenated with the scalar features of l=0, the mean and variance of the sealing ring, the maximum residual, the circumferential coverage features, the degree of central convexity, the peripheral collision features, and the axis confidence, and then compressed into a fixed-dimensional conditional feature F using a scalar multilayer perceptron. i .

[0046] ; Since the dot product, norm, tensor contraction, and cylindrical coordinate statistics mentioned above are all rotation-invariant, therefore F i (RP) = F i (P). This enables subsequent scalar scoring networks to receive a consistent conditional distribution across different scene orientations.

[0047] Step S6: Scalar Conditional Diffusion Scoring This invention preferably generates diffusion only for point-level adsorption feasibility fractions, without diffusing the suction cup axis or physically unnecessary yaw angles. Let the actual adsorbed mass yi∈[0,1], mapped as... The forward noise addition process is as follows: ; The conditional denoising network receives the current noise score x. i,t Time step embedding E(t), invariant condition F i And the candidate point neighborhood context, predict noise Or directly predict the clean score The denoising block includes input projection, temporal modulation, local scalar message aggregation, channel attention, and residual update. Neighborhood messages only use inter-point distance and F...i And scalar hidden features, instead of directly feeding the unprocessed 3D orientation into a regular multilayer perceptron.

[0048] ; During the inference phase, starting from Gaussian noise or a preset initial score, iterative noise reduction is performed using DDPM, DDIM, or an equivalent deterministic sampler to obtain the adsorption score s for each candidate point. i .

[0049] Preferably, 10 to 20 sampling steps are used; in embodiments with high speed requirements, a few-step distillation, a one-step consistency model, or a direct scalar regression head can be used to replace diffusion scoring, while retaining the aforementioned isotropic axis prediction and contact ring analysis.

[0050] Step S7: Candidate Screening and Robot Execution Based on the adsorption fraction s i , axis confidence q i The sealing ring quality and collision risk are used to form a comprehensive ranking, and Top-K candidates are selected. Position-direction joint non-maximum suppression is employed, and the distance between candidates i and j can be defined as: .

[0051] Collision detection was then performed on the suction cup housing, flange, and approach path, followed by robot inverse kinematics screening. The contact pose output by the neural network is (p i ,a i If the complete tool structure is also axisymmetric about the suction cup, the yaw angle can be fixed; if the air tube, flange, or robot wrist disrupts the axisymmetry, then the representative with the lowest collision and joint costs is selected from the set of yaw equivalences that do not change the contact state. .

[0052] Based on the calibration transformation from the suction cup contact surface to the robot flange, the contact center p i Approaching axis a i The selected yaw angle is converted into the six-DOF pose of the end effector and sent to the robot controller.

[0053] This embodiment also provides a system, apparatus, and storage medium to implement the above method, including a point cloud acquisition module, a rotationally isotropic geometry encoding module, a suction cup axis prediction module, an axisymmetric contact ring analysis module, a rotationally invariant compression module, a scalar conditional denoising module, a candidate screening module, and a robot control module. These modules can be implemented by one or more processors executing programs in memory, or they can be deployed in an industrial computer, edge computing device, robot control cabinet, or cloud server.

[0054] This embodiment also covers a computer-readable storage medium storing program instructions, which, when run on a processor, causes the processor to perform the above steps.

[0055] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0056] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for grasping objects with a circular suction cup, characterized in that, Includes the following steps: S1. Point Cloud Acquisition and Candidate Point Preprocessing: Acquire scene point clouds using 3D imaging equipment. The original point cloud is processed by working space pruning, outlier removal, voxel downsampling, and farthest point sampling to obtain a set of spatial candidate points. Pre-calibrate the radius r of the circular suction cup sealing ring s lip width w s Tool height and perimeter safety distance are used as fixed constants of the system; S2. Multi-scale rotational isotropic geometric coding: The local space of candidate points is divided into multi-scale spherical neighborhoods. A spherical harmonic tensor product isotropic encoder is used to extract hierarchical features of the relative coordinates of the point cloud, outputting l0 scalar features, l1 vector features, and l2 second-order tensor features. When the scene undergoes arbitrary spatial rotation, the l0 features remain unchanged, the l1 features rotate synchronously, and the l2 tensor satisfies Q'=RQR. T ; S3. Suction Cup Proximity Axis Prediction: Based on the features of each candidate point l0, l1, and l2, gating weights are constructed. The multi-channel l1 vectors are weighted, synthesized, and normalized to obtain the directed proximity axis a of the suction cup. i Synchronous output axis confidence q i Introducing the camera's line-of-sight vector eliminates the ambiguity of the normal's positive or negative sign, outputting only the adsorption center p. i With near axis a i It does not return to the yaw angle around the axis; S4, Geometric analysis of axisymmetric contact ring: with (p i ,a i Establish a local cylindrical coordinate system and calculate the axial height z of the neighboring points. ij Radial distance ρ ij Based on the suction cup calibration parameters, the central area, sealing ring area, and outer safety area are divided. Gaussian soft weighting is used to calculate the weighted mean height, height variance, and circumferential harmonic coverage amplitude of the sealing ring to characterize the continuity of the suction cup seal. S5. Rotation-Invariant Feature Compression: The l1 vector and the l2 second-order tensor are subjected to axial projection, tangential norm, and tensor anisotropy contraction operations along the prediction suction cup axis. Combined with the sealing ring statistical index, a globally rotation-invariant conditional feature F is formed. i F under arbitrary scene rotation i The value remains constant; S6, Scalar Conditional Diffusion-Adsorption Score: Based on rotationally invariant characteristic F i Given the condition, a diffusion denoising model is used to iteratively solve for the scalar score s of adsorption feasibility at each candidate point. i Only neighborhood context diffusion is performed on the adsorption fraction; diffusion updates are not performed on the suction cup proximity axis or yaw angle. S7. Candidate Grasping and Screening and Robot Execution: Based on Adsorption Score s i , axis confidence q i The sealing ring integrity and peripheral collision risk are comprehensively ranked, and a position-direction joint NMS is used to screen the Top-K candidates. For each candidate, within the yaw equivalence interval without changing the sealing contact, the optimal yaw angle that minimizes the collision cost and joint cost is solved. i ,a i The optimal yaw angle is converted into the robot's end-effector's six-degree-of-freedom grasping pose and executed.

2. The circular suction cup grasping method according to claim 1, characterized in that, In step S2, the multi-scale spherical neighborhood is selected at a scale of 0.5r. s 1.0r s 2.0r s 4.0r s The equivariant encoder is based on the PointNet++ hierarchical architecture, and uses the spherical harmonic function Ylr and tensor product operation to realize equivariant message aggregation. Each layer outputs three types of hierarchical features: l0, l1, and l2.

3. The circular suction cup grasping method according to claim 1, characterized in that, In step S3, the formula for calculating the gating weight is as follows: Weighted composite vector Normalization yields suction cups close to the axis .

4. The circular suction cup grasping method according to claim 1, characterized in that, In step S4, the radial range of the sealing ring satisfies Soft radial weight ; through the m-th order circular harmonic amplitude A i,m Characterizes circumferential gaps and local discontinuities in the sealing ring.

5. The circular suction cup grasping method according to claim 1, characterized in that, The rotation-invariant compression in step S5 includes: axial projection of vector l1. Tangential residual norm ; l2 tensor axial response Anisotropic tensor norm of the tangent plane The above indicators are concatenated with the sealing ring variance and coverage amplitude to obtain the fixed-dimensional feature F. i .

6. The circular suction cup grasping method according to claim 1, characterized in that, In step S6, the loss function during the diffusion model training phase includes axis loss, confidence loss, sealing supervision loss, equivariance consistency loss, and diffusion denoising loss; during the inference phase, DDIM is used to generate adsorption scores through few-step sampling, which can be replaced by a direct scalar regression network in scenarios with limited computing power.

7. The circular suction cup grasping method according to claim 1, characterized in that, In step S7, the formula for calculating the joint position-direction NMS distance is as follows: The yaw angle optimization objective is .

8. A circular suction cup gripping system, characterized in that, The method for performing the circular suction cup grasping method according to any one of claims 1-7 includes: The point cloud acquisition module is used to acquire 3D point clouds of the scene and perform preprocessing such as cropping and downsampling. A multi-scale rotational equivariant geometry encoding module is used to extract l0 scalar, l1 vector, and l2 second-order tensor hierarchical equivariant features from point clouds; The suction cup axis prediction module is used to calculate the directional proximity axis and axis confidence of the suction cup based on hierarchical isovariant features. The axisymmetric contact ring analysis module is used to construct a cylindrical coordinate system based on the predicted suction cup axis and solve for sealing evaluation parameters such as sealing ring height and circumferential coverage. The rotation invariant compression module is used to shrink the isovariant feature along the chuck axis to generate rotationally invariant conditional features; The scalar conditional diffusion scoring module is used to output an adsorption feasibility score for each candidate point through a diffusion model, based on invariant characteristics. The candidate selection and robot control module is used to jointly solve the optimal six-degree-of-freedom grasping pose by combining nonmaximum suppression, collision detection, and inverse kinematics, and then send the results to the robot to execute the grasping action. Each module is stored in the memory of the edge computing device / robot control cabinet and is scheduled and coordinated by the processor.

9. The circular suction cup gripping system according to claim 8, characterized in that, The point cloud acquisition module can be any one of a depth camera, a structured light camera, or a lidar.

10. The circular suction cup gripping system according to claim 8, characterized in that, The system is also configured with a computer-readable storage medium storing a computer program that, when executed by a processor, implements all the steps of the grasping method according to any one of claims 1-7.

Images