A robot stable grasping judgment control method for pipe feeding

Abstract

The application provides a kind of robot stable grasping determination control method for pipe feeding, belongs to grasping determination control technical field, under the condition of complex background and material frame mixed loading, background area, bearing area and candidate pipeline area can be separated first, the influence of interference point on identification result is significantly reduced, the accuracy of target extraction is improved;For local occlusion, point cloud missing and pipeline contact pressure stacking, through local geometric consistency judgment, only the target with sufficient visible features and stable grasping conditions is retained, false identification and false grasping are avoided;Through the fusion decision of geometric features and artificial intelligence semantic results, the identification reliability and grasping confidence of similar pipelines and random attitude pipelines are improved;Through grasping point scoring, six-degree-of-freedom pose output and low-score alarm mechanism, the robot preferentially performs grasping action that can reach, clamp and stably carry, reduces the risk of dropping, deviation and collision.

Description

Technical Field

[0001] This invention belongs to the field of gripping and control technology, specifically relating to a stable gripping and control method for pipe fittings, which is a step in the nuclear power pipeline processing control process. Background Technology

[0002] While current technologies have achieved a certain level of automation in material handling, they generally rely on deep learning segmentation, point cloud registration, or grasping point scoring to determine the object to be grasped. These methods are typically applicable to general industrial parts or flexible materials, but for workpieces such as nuclear power pipelines, which have extremely high requirements for grasping safety, they often lack fine-grained judgment on local occlusion, missing point clouds, and structural consistency. Especially when multiple pipelines are stacked and in contact with each other, the system is prone to misjudging some exposed areas as complete targets, leading to unstable grasping actions and even causing secondary collisions and scratches on the workpiece surface.

[0003] The prior art solution with patent publication number CN121849649A discloses a method and system for loading irregular products based on robot vision. It mainly acquires point clouds of stacked scenes through 3D vision, and generates grasping actions by combining instance segmentation, pose estimation, grasping interference evaluation, and a dual-mode grasping strategy. This solution is more geared towards planning the loading of irregular products. Although it considers the grasping order and information completeness, it cannot further determine whether the pipeline has a stable gripping area for scenarios such as nuclear power pipelines, where long axes, partial occlusion, and random postures coexist. Therefore, it cannot achieve a precise determination of whether the pipeline can be stably grasped. The prior art solution with patent publication number CN121698087A discloses an automatic loading and precise positioning control system for flexible materials based on machine vision. It achieves automated loading through image acquisition, flexible material recognition, deformation parameter quantification, grasping strategy selection, and visual servo positioning. This solution emphasizes the deformation adaptation and precise positioning of flexible materials and is suitable for loading deformable materials. While material control is possible, it cannot jointly determine the local geometric integrity, grasping stability, and risky grasping areas of rigid nuclear power pipelines. Therefore, the risk of misgrabbing and collision still exists in nuclear power pipeline scenarios. The prior art solution with patent publication number CN121820202A discloses a machine sorting method, system, device, and medium based on vector paper semantic parsing. It mainly obtains geometric and non-structural information through vector paper, and then determines the attribute relationships of parts through attribute association maps and graph neural networks, and generates a sorting task list. This solution can achieve part recognition and task assignment at the drawing level, but it cannot directly address the issues of point cloud missingness, occlusion interference, and grasping stability in the real working environment. Therefore, it cannot make on-site judgments on the actual grasping feasibility of nuclear power pipelines. The prior art solution with patent publication number CN121733539A discloses a technical route for robot material loading through 3D vision, point cloud analysis, and grasping strategy generation. It usually includes scene parsing, sequence planning, and grasping point determination. The solution can complete general material loading tasks, but it lacks in-depth joint judgment on the stability and graspability confidence of the target. It cannot fully screen the grasping results in high-risk scenarios such as nuclear power pipelines, and therefore still cannot meet the requirements of high safety and high reliability for nuclear power pipeline material loading. Summary of the Invention

[0004] The purpose of this invention is to provide a robot-based stable grasping and control method for pipeline loading. Nuclear power pipelines are critical components in nuclear power equipment manufacturing, requiring high quality and structural integrity. If a robot is allowed to grasp a pipeline prematurely without sufficient target identification and clear grasping conditions, it is prone to collisions, dropped parts, or clamping misalignment due to incorrect posture judgment, partial obstruction, or workpiece stacking interference. This could damage the robot and fixtures, but most importantly, it could cause irreversible damage to the pipeline, affecting subsequent processing and assembly quality. Furthermore, while existing technologies can identify and grasp pipelines, they lack effective judgment on the stability and feasibility of the grasping process. This invention aims to establish a joint judgment mechanism for grasping execution. By performing layered preprocessing of the scene point cloud, candidate target screening, local geometric consistency estimation, and geometric feature and artificial intelligence semantic fusion analysis, the graspability, stability, and reliability of the pipeline target are comprehensively judged, thereby improving the automation level, safety, and grasping success rate of nuclear power pipeline loading operations.

[0005] The present invention employs the following technical solution.

[0006] A robot-based stable gripping and control method for pipe fitting feeding includes:

[0007] Step 1: Collect and preprocess scene data for the robot's grasping operation for loading nuclear power plant pipe parts;

[0008] Step 2: Within the candidate nuclear power pipeline area, extract candidate targets based on the effective point cloud set;

[0009] Step 3: After obtaining the set of candidate pipeline targets, perform local geometric consistency estimation to obtain candidate targets that meet the set conditions;

[0010] Step 4: After obtaining candidate targets that meet the set conditions, perform a geometric feature and artificial intelligence semantic fusion judgment;

[0011] Step 5: Perform geometric feature and artificial intelligence semantic fusion judgment, and execute the crawling evaluation output.

[0012] Furthermore, step 1 specifically includes:

[0013] First, a 3D camera was used to collect point cloud data of the work area where the robot was loading materials onto the nuclear power plant pipeline, resulting in the original point cloud set: ,in , represents the 3D coordinates of the i-th point in the original point cloud set in the camera coordinate system. This represents the grayscale value or reflection intensity value of the i-th point in the original point cloud set, and N represents the total number of points in the original point cloud.

[0014] Furthermore, step 1 specifically includes: firstly, performing height constraint filtering on each point in the original point cloud set, retaining points that satisfy the following formula: ,in, and These represent the lower and upper limits of the working layer's height, respectively; further filtering based on grayscale or reflection intensity yields: ,in, and These are the lower and upper limits of the set intensity filtering threshold, respectively. This represents the grayscale value or reflection intensity value of the i-th point in the original point cloud set; further, outlier removal is performed by combining the point neighborhood density, for any point... The number of its neighborhood points is defined as: ,in For point and The three-dimensional Euclidean distance between them, when At that time, the point These are considered discrete noise points and discarded. The threshold for the number of points in the smallest neighborhood. This is the set neighborhood radius threshold.

[0015] Furthermore, step 2 specifically includes:

[0016] Within the candidate nuclear power pipeline area, the effective point cloud set is... Perform Euclidean distance clustering to obtain an effective set of point clouds. Any two points and Euclidean distance between Defined as: When satisfied When the two points are grouped into the same connected point cluster, that is, the candidate point cluster, then... This represents the set clustering distance threshold, from which the candidate target set is obtained. ,in Indicates the first There are 3 candidate point clusters, k = 1, 2, ..., n, where n is the number of candidate point clusters. For each candidate point cluster... Furthermore, its geometric and statistical characteristics are calculated.

[0017] Furthermore, step 2 specifically includes: a method for further calculating its geometric statistical characteristics, specifically including:

[0018] Let its center of mass be for: And construct the covariance matrix The three eigenvalues ​​are obtained by solving the covariance matrix. , , , Further define the linearity index and firmness index When the candidate point cluster satisfies the preset constraints on length range, number of points, linearity, and compactness, that is: , , , These are retained as candidate pipeline targets, and all candidate pipeline targets constitute the candidate pipeline target set. This represents the estimated length of the point cluster along the main direction. The minimum effective length threshold for nuclear power plant pipelines is set. The maximum effective length threshold for nuclear power plant pipelines is set. For the first The total number of points in each candidate point cluster. and These are the set linearity threshold and compactness threshold, respectively. This is the minimum point cloud quantity threshold required to determine whether a nuclear power pipeline is valid.

[0019] Furthermore, step 3 specifically includes:

[0020] Let the theoretical length of the standard pipe model be... For the first A cluster of candidate points as candidate targets Let its effective visible length along the principal axis be... The visibility coverage rate of the candidate target is then defined as follows. ,in Used to characterize the effective visibility of candidate targets along the principal axis; when a candidate target satisfies the visibility coverage condition. When a candidate target is deemed to have the basis for further local geometric consistency estimation, for candidate targets that meet the visibility coverage condition, one or more local sections are selected along its principal axis, and the points on the local sections are projected onto a local section coordinate system perpendicular to the principal axis. The circle fitting relationship is expressed as follows: ,in The coordinates of the center of the circle for local cross-section fitting. The fitted radius of the local cross section is . and Let the set of points on the cross section be the two-dimensional abscissa and ordinate of a single point on the cross section in the local cross section coordinate system. Further define the fitting residual as ,in M is used to characterize the deviation between the local cross-sectional geometry of the candidate target and the ideal circular cross-section, where M is the total number of points within the current local cross-section participating in circle fitting. and These are the two-dimensional abscissa and ordinate of the m-th point within the current local section in the local section coordinate system, respectively.

[0021] Furthermore, step 3 specifically includes:

[0022] Define the boundary continuity index as ,in This indicates the number of points in the current local section that satisfy the continuous boundary conditions. This represents the total number of points in the current local section. Furthermore, a local geometric consistency score is constructed. ,in When satisfied , , , Only under these four conditions will the local geometric consistency score of the candidate target be used as a valid input for subsequent judgment. This represents the maximum allowable fitting residual threshold for the set local cross-section. This indicates the set boundary continuity threshold. This represents the set local geometric consistency score threshold. For candidate targets that pass the local geometric consistency estimation, i.e., meet the four conditions, they are retained as valid inputs for subsequent geometric feature modeling, semantic fusion determination, and crawling point evaluation. For candidate targets that fail the local geometric consistency estimation, they are directly removed from the subsequent crawling candidate set.

[0023] Furthermore, step 4 specifically includes:

[0024] After obtaining candidate targets that meet the four conditions, the candidate targets are judged from both the geometric and semantic sides, and a comprehensive credibility result for crawling and execution is formed through a fusion mechanism.

[0025] Furthermore, in step 4, after obtaining candidate targets that satisfy the four conditions, the candidate targets are judged from both the geometric and semantic perspectives, and a comprehensive credibility result for crawling execution is formed through a fusion mechanism. The specific implementation method includes:

[0026] For the candidate target that satisfies the above four conditions, the first one candidate targets Extract its geometric feature vector ,in Indicates the target length. Indicates the target diameter. Indicates the spindle direction parameter. Indicates the roundness characteristics of the cross section. Indicates the surface normal distribution characteristics, Indicates visible coverage. Indicates boundary continuity, The local fitting residual is denoted as the normalized residual after normalizing each geometric feature. One geometric feature is Geometric decision scores are constructed based on normalized geometric features. ,in The weights corresponding to the m-th geometric feature satisfy the following conditions: , ;

[0027] Input the point cloud clusters, depth maps, or projection views corresponding to the candidate targets into the artificial intelligence semantic recognition model to obtain semantic output vectors. ,in Indicates the target category output. Indicates the confidence level for category identification. This indicates the popularity of the candidate capture area. The result of the occlusion estimation is represented by the normalization of the semantic features, and the normalized result is denoted as the i-th. The semantic features are Constructing semantic judgment scores based on normalized semantic features ,in For the corresponding weights, the following should be satisfied: , ;

[0028] The fusion weights are dynamically adjusted based on the geometric completeness and semantic credibility of the candidate targets; that is, a semantic credibility term is first constructed. ,in The semantic occlusion correction coefficient is used, and then the geometric reliability term is constructed. ,in , , Let the weighting coefficients satisfy: =1, defining dynamic fusion weights based on semantic reliability and geometric reliability terms. in, For balance coefficient, To prevent extremely small positive numbers with a denominator of zero, and ;

[0029] Construct a basic fusion score based on geometric and semantic decision scores. Furthermore, to enhance the ability to suppress geometric decision-making and semantic conflict scenarios, a consistency compensation term is introduced. And construct the corrected fusion result: ,in This is the conflict penalty coefficient. This is the local geometric consistency compensation coefficient. The obtained local geometric consistency score;

[0030] Set a comprehensive judgment threshold When the candidate target satisfies: When, retain it as a valid candidate for crawling, when At the same time, a second judgment is performed by combining the degree of occlusion, the integrity of the boundary and the continuity of the candidate grabbing area. If the candidate target still has a stable local grabbing area, its grabbing priority is reduced and it is retained; otherwise, it is eliminated.

[0031] Furthermore, step 5 specifically includes:

[0032] For the valid candidate targets after screening, the system further evaluates their candidate capture points, that is, let the first... The first candidate target The crawl points are A crawling evaluation function is established for it, which includes:

[0033] First, define the exposure index. It is used to characterize the degree of exposure of the gripping area and the accessibility of the grippers, and defines a stability index. It is used to characterize the risk of the target rolling, swaying, or falling off during the grasping process, and defines the interference risk. It is used to characterize the probability of the grasping path colliding with surrounding workpieces or material frame structures, and defines the reachability index. The grasp point score is used to characterize the pose realizability of the robot's end effector at the current position. ,in The obtained target comprehensive judgment score Introducing crawl point-level decision-making, we obtain the final crawl point score: ,in, The set fusion judgment compensation coefficient; the system applies the following to all candidate capture points: Sort the results and select the highest-rated result as the current preferred crawl point: For the selected preferred grasping point, output its spatial coordinates and attitude parameters to form the six-degree-of-freedom grasping pose of the robot's end effector: in, Indicates the coordinates of the grab point location. The corresponding attitude angle parameters are represented, and the pose information is sent to the robot control system to guide the end effector to complete a stable grasping action. When the maximum grasping score is lower than a preset threshold... When this happens, the system outputs a signal indicating a failure to capture or requiring manual review, in order to prevent the robot from performing erroneous capture operations in low-confidence scenarios.

[0034] The beneficial effects of the present invention are as follows, compared with the prior art:

[0035] This invention achieves stable grasping in nuclear power pipeline loading operations through process optimization and system integration, including scene-layered preprocessing, candidate target extraction, local geometric consistency estimation, geometric and AI semantic fusion judgment, and grasping point evaluation output. Under complex backgrounds and mixed material loading conditions, it can first separate the background area, the load-bearing area, and the candidate pipeline area, significantly reducing the impact of interference points on the recognition results and improving target extraction accuracy. For situations involving partial occlusion, missing point clouds, and pipeline contact and overlapping, local geometric consistency judgment retains only targets with sufficient visible features and stable grasping conditions, avoiding misidentification and misgrabbing. The fusion decision-making based on geometric features and AI semantic results improves the reliability of identifying similar pipelines and pipelines with random postures, as well as the grasping confidence. Through grasping point scoring and ranking, six-degree-of-freedom pose output, and a low-score alarm mechanism, the robot prioritizes grasping actions that are reachable, grippable, and stably transportable, reducing the risk of dropped parts, swaying, and collisions. Therefore, this invention can effectively improve the automation level, grasping success rate, operational safety, and flexible adaptability of nuclear power pipeline loading, and has significant engineering application value. Attached Figure Description

[0036] Figure 1 This is a partial flowchart illustrating the principle of the robot's stable grasping and control method for pipe fitting feeding in this invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, other embodiments obtained by those skilled in the art without creative effort are all within the protection scope of this invention.

[0038] like Figure 1 As shown, this invention proposes a robot stable gripping and control method for pipe fitting feeding, comprising the following steps:

[0039] This invention provides a robot grasping and recognition method for automated material handling in nuclear power plant pipelines. This method combines 3D point cloud processing, local geometric consistency analysis, and a geometric feature and AI semantic fusion judgment mechanism to achieve stable recognition and grasping pose output for pipeline targets under conditions of complex stacking, random postures, partial occlusion, and missing point clouds. Specifically, it includes the following steps:

[0040] Step 1: Collect and preprocess scene data for the robot's grasping operation for loading nuclear power plant pipe parts;

[0041] In a preferred but non-limiting embodiment of the present invention, step 1 specifically includes:

[0042] First, a 3D camera was used to collect point cloud data of the work area where the robot was loading materials onto the nuclear power plant pipeline, resulting in the original point cloud set: ,in , represents the 3D coordinates of the i-th point in the original point cloud set in the camera coordinate system. This represents the grayscale value or reflection intensity value of the i-th point in the original point cloud set, and N represents the total number of points in the original point cloud.

[0043] In a preferred but non-limiting embodiment of the present invention, step 1 further includes: to remove the ground, the area outside the material frame, and discrete noise points, firstly, performing height constraint filtering on each point in the original point cloud set, retaining points that satisfy the following formula: ,in, and These represent the lower and upper limits of the work layer's height, respectively. When there is strong reflection, abnormal background reflection, or significant material differences in the work scene, further filtering is performed based on grayscale or reflection intensity to obtain: ,in, and These are the lower and upper limits of the set intensity filtering threshold, respectively. This represents the grayscale value or reflection intensity value of the i-th point in the original point cloud set; furthermore, outlier removal can be performed by combining the point neighborhood density, for any point... The number of its neighborhood points is defined as: ,in For point and The three-dimensional Euclidean distance between them, when At that time, the point These are considered discrete noise points and discarded. This is the threshold for the number of points in the smallest neighborhood. After preprocessing, the effective point cloud set is obtained. Based on height distribution, spatial location, and reflection characteristics, the scene is divided into background area, carrying area, and candidate pipeline area, thus providing a stable data foundation for subsequent candidate target extraction. For the set neighborhood radius threshold, It can be designed according to specific requirements.

[0044] Step 2: Within the candidate nuclear power pipeline area, extract candidate targets based on the effective point cloud set;

[0045] In a preferred but non-limiting embodiment of the present invention, step 2 specifically includes:

[0046] Within the candidate nuclear power pipeline area, the effective point cloud set is... Perform Euclidean distance clustering to obtain an effective set of point clouds. Any two points and Euclidean distance between Defined as: When satisfied When the two points are grouped into the same connected point cluster, that is, the candidate point cluster, then... This represents the set clustering distance threshold, from which the candidate target set is obtained. ,in Indicates the first There are 3 candidate point clusters, k = 1, 2, ..., n, where n is the number of candidate point clusters. For each candidate point cluster... Furthermore, its geometric and statistical characteristics are calculated.

[0047] In a preferred but non-limiting embodiment of the present invention, step 2 further includes: a method for further calculating its geometric statistical characteristics, specifically including:

[0048] Let its center of mass be for: And construct the covariance matrix The three eigenvalues ​​are obtained by solving the covariance matrix. , , , Then we can obtain the corresponding Main direction vector Based on this, the principal axis direction and length distribution of the point cluster are estimated, and a linearity index is further defined. and firmness index When the candidate point cluster satisfies the preset constraints on length range, number of points, linearity, and compactness, that is: , , , These are retained as candidate pipeline targets, and all candidate pipeline targets constitute the candidate pipeline target set. This represents the estimated length of the point cluster along the main direction. The minimum effective length threshold for nuclear power plant pipelines is set. The maximum effective length threshold for nuclear power plant pipelines is set. For the first The total number of points in each candidate point cluster. and These are the set linearity threshold and compactness threshold, respectively. This is the minimum point cloud quantity threshold required to determine a valid nuclear power pipeline. For candidate point clusters that are in contact, stacked, or locally adhered, the local normal change rate or cross-sectional spacing change can be calculated along the principal direction. When the abrupt change in local direction exceeds the threshold or the boundary gap is greater than the set threshold, the corresponding point cluster is split. When the principal direction difference between adjacent point clusters is less than the threshold, the end spacing is less than the threshold, and the contour continuity meets the condition, a merging process is performed to improve the accuracy of candidate target extraction.

[0049] Step 3: After obtaining the candidate pipeline target set in Step 2, perform local geometric consistency estimation and judgment to obtain candidate targets that meet the set conditions;

[0050] In a preferred but non-limiting embodiment of the present invention, step 3 specifically includes:

[0051] After obtaining the candidate pipeline target set in step 2, if a candidate pipeline target has partial occlusion, incomplete boundaries, or missing point clouds, this invention does not directly treat it as a complete workpiece for forced restoration. Instead, it first determines whether the candidate pipeline target still retains sufficient visible geometric features to support subsequent grasping and analysis. Specifically, the theoretical length of the standard pipeline model is set to... The theoretical length The parameters are obtained from the specification library, production task sheet, process database, or manually preset parameters of the nuclear power plant pipeline to be captured. For the first... A cluster of candidate points as candidate targets Let its effective visible length along the principal axis be... The visibility coverage rate of the candidate target is then defined as follows. ,in Used to characterize the effective visibility of candidate targets along the principal axis; when a candidate target satisfies the visibility coverage condition. When the candidate target is determined to have a basis for further local geometric consistency estimation, preferably, It can be set to 0.7. For candidate targets that meet the visibility coverage condition, one or more local sections are further selected along their principal axis, and the points on the local sections are projected into a local section coordinate system perpendicular to the principal axis. Since the pipe is a cylindrical structure, a circular model can be used to fit the set of points on the local section in the local section coordinate system. The circular fitting relationship can be expressed as: ,in The coordinates of the center of the circle for local cross-section fitting. The fitted radius of the local cross section is . and Let the set of points on the cross section be the two-dimensional abscissa and ordinate of a single point on the cross section in the local cross section coordinate system. The fitting residual can be further defined as ,in M is used to characterize the deviation between the local cross-sectional geometry of the candidate target and the ideal circular cross-section, where M is the total number of points within the current local cross-section participating in circle fitting. and These are the two-dimensional abscissa and ordinate of the m-th point within the current local section in the local section coordinate system, respectively. The smaller this value is, the more the local geometry conforms to the cylindrical structure characteristics of the pipeline.

[0052] In a preferred but non-limiting embodiment of the present invention, step 3 further includes:

[0053] Furthermore, to assess the boundary continuity of candidate targets, a boundary continuity index is defined as follows: ,in This indicates the number of points in the current local section that satisfy the continuous boundary conditions. This represents the total number of points in the current local section, when the boundary continuity index... The larger the value, the more complete and continuous the local boundary of the candidate target. To comprehensively evaluate the local geometric completeness and local shape consistency of candidate targets, a local geometric consistency score is further constructed. ,in When satisfied , , , Only under these four conditions will the local geometric consistency score of the candidate target be used as a valid input for subsequent judgment. This represents the maximum allowable fitting residual threshold for the set local cross-section. This indicates the set boundary continuity threshold. This represents the set local geometric consistency scoring threshold. Candidate targets that pass the local geometric consistency estimation (i.e., meet the four conditions) are retained as valid input for subsequent geometric feature modeling, semantic fusion determination, and grab point evaluation. Candidate targets that fail the local geometric consistency estimation have their grab priority reduced, or are directly removed from the subsequent grab candidate set. Candidate targets that meet the four set conditions are considered to satisfy the criteria. , , , Candidate targets for these four conditions.

[0054] Step 4: After obtaining candidate targets that meet the above four conditions, perform a geometric feature and artificial intelligence semantic fusion judgment;

[0055] In a preferred but non-limiting embodiment of the present invention, step 4 specifically includes:

[0056] This step is a key step in the present invention. After obtaining candidate targets that meet the four conditions, the candidate targets are judged from both geometric and semantic perspectives, and a comprehensive credibility result for crawling and execution is formed through a fusion mechanism.

[0057] In a preferred but non-limiting embodiment of the present invention, in step 4, after obtaining candidate targets that satisfy the four conditions, the candidate targets are judged from both the geometric and semantic perspectives, and a comprehensive credibility result for crawling execution is formed through a fusion mechanism. The specific implementation method includes:

[0058] For the candidate target that satisfies the above four conditions, the first one candidate targets Extract its geometric feature vector ,in Indicates the target length. Indicates the target diameter. Indicates the spindle direction parameter. Indicates the roundness characteristics of the cross section. Indicates the surface normal distribution characteristics, Indicates visible coverage. Indicates boundary continuity, The local fitting residual is denoted as the normalized residual after normalizing each geometric feature. One geometric feature is Geometric decision scores are constructed based on normalized geometric features. ,in The weights corresponding to the m-th geometric feature satisfy the following conditions: , ;

[0059] Input the point cloud clusters, depth maps, or projection views corresponding to the candidate targets into an AI semantic recognition model (the AI ​​semantic recognition model outputs target category confidence, candidate capture region heat, and occlusion estimation results; the model type can be a point cloud segmentation network, a depth map segmentation network, or a multi-view recognition network) to obtain a semantic output vector. ,in Indicates the target category output. Indicates the confidence level for category identification. This indicates the popularity of the candidate capture area. The result of the occlusion estimation is represented by the normalization of the semantic features, and the normalized result is denoted as the i-th. The semantic features are Constructing semantic judgment scores based on normalized semantic features ,in For the corresponding weights, the following should be satisfied: , ;

[0060] To improve the stability of judgment under occlusion, missing parts, and complex backgrounds, this invention does not use a fixed fusion ratio, but dynamically adjusts the fusion weights based on the geometric completeness and semantic credibility of the candidate targets. That is, a semantic credibility term is first constructed. ,in The semantic occlusion correction coefficient is used, and then the geometric reliability term is constructed. ,in , , Let the weighting coefficients satisfy: =1, defining dynamic fusion weights based on semantic reliability and geometric reliability terms. in, For balance coefficient, To prevent extremely small positive numbers with a denominator of zero, and When the semantic recognition confidence of the candidate target is high, the crawling popularity is concentrated, and the occlusion degree is low, Larger Increased semantic judgment ratio leads to higher semantic judgment ratios. This is especially true when the candidate target has high local geometric integrity, good boundary continuity, and a small fitting residual. Larger The proportion of geometric determinations is relatively reduced, while the proportion of geometric determinations is increased.

[0061] Construct a basic fusion score based on geometric and semantic decision scores. Furthermore, to enhance the ability to suppress geometric decision-making and semantic conflict scenarios, a consistency compensation term is introduced. And construct the corrected fusion result: ,in This is the conflict penalty coefficient. This is the local geometric consistency compensation coefficient. The local geometric consistency score obtained in step 3;

[0062] Set a comprehensive judgment threshold When the candidate target satisfies: When, retain it as a valid candidate for crawling, when At the same time, a second judgment is performed by combining the degree of occlusion, the integrity of the boundary and the continuity of the candidate grabbing area. If the candidate target still has a stable local grabbing area, its grabbing priority is reduced and it is retained; otherwise, it is eliminated.

[0063] Through the above method, the present invention combines traditional geometric constraints with artificial intelligence semantic recognition to make joint decisions, not only determining whether the target is a nuclear power pipeline, but also focusing on whether it is suitable for stable grasping, thereby improving the reliability of grasping and recognition under complex working conditions.

[0064] Step 5: Perform geometric feature and artificial intelligence semantic fusion judgment, and execute the crawling evaluation output.

[0065] In a preferred but non-limiting embodiment of the present invention, step 5 specifically includes:

[0066] For the valid candidate targets after screening, the system further evaluates their candidate capture points, that is, let the first... The first candidate target The crawl points are A crawling evaluation function is established for it, which includes:

[0067] First, define the exposure index. It is used to characterize the degree of exposure of the gripping area and the accessibility of the grippers, and defines a stability index. It is used to characterize the risk of the target rolling, swaying, or falling off during the grasping process, and defines the interference risk. It is used to characterize the probability of the grasping path colliding with surrounding workpieces or material frame structures, and defines the reachability index. The grasp point score, used to characterize the pose realizability of the robot's end effector at its current position, can be expressed as: ,in The target comprehensive judgment score obtained in step 4 is then used. Introducing crawl point-level decision-making, we obtain the final crawl point score: ,in, The set fusion judgment compensation coefficient; the system applies the following to all candidate capture points: Sort the results and select the highest-rated result as the current preferred crawl point: For the selected preferred grasping point, output its spatial coordinates and attitude parameters to form the six-degree-of-freedom grasping pose of the robot's end effector: in, Indicates the coordinates of the grab point location. The corresponding attitude angle parameters are represented, and the pose information is sent to the robot control system to guide the end effector to complete a stable grasping action. When the maximum grasping score is lower than a preset threshold... When this happens, the system outputs a signal indicating a failure to capture or requiring manual review, in order to prevent the robot from performing erroneous capture operations in low-confidence scenarios.

[0068] This invention addresses the challenges of automated pipeline loading scenarios, including background interference, disordered workpiece stacking, partial occlusion, random posture, missing point clouds, and difficulty in stably determining gripping points. It proposes a strategy and technical solution for robot gripping and recognition based on 3D vision, local geometric consistency judgment, and the fusion of geometry and AI semantics. This invention focuses on constructing the gripping process based on a gripping-capable rather than a recognizable strategy. Most 3D vision gripping solutions or disordered gripping patents prioritize workpiece recognition, emphasizing how to grip without addressing the stability of workpiece gripping. This invention, however, addresses the joint judgment of target recognizability, gripping capability, and gripping area stability in complex scenarios for automated robot gripping tasks. Background interference is reduced through scene layering and candidate target extraction mechanisms. Instead of directly processing the entire point cloud of the work area, this invention first layers the work scene based on the height and grayscale / reflection intensity information of the 3D point cloud, dividing the scene into a background area, a load-bearing area, and a candidate pipeline area. This method effectively extracts key regions by identifying the background, the bounding box, and the candidate pipelines. It employs a local geometric consistency judgment mechanism to adapt to occlusion and point cloud missing fields. When candidate targets have partial occlusion, incomplete boundaries, or missing point clouds, traditional methods sometimes forcibly match them with the model to obtain pipeline information. This invention, however, performs local geometric consistency judgment on the visible area. In other words, it compares the point cloud information with the pipeline model information to determine the current point cloud pipeline coverage. If the coverage is too low, the candidate pipeline is eliminated. A fusion decision-making mechanism combining geometric features and AI semantic recognition is used. Traditional geometric feature judgment is combined with AI semantic segmentation, and the weights of the two results are used to determine the confidence level of the crawling. Finally, a crawling point evaluation and failure handling mechanism is implemented for crawling execution. Candidate grasping points are scored and sorted, and the one with the highest score is selected as the current preferred grasping point. The robot's six-degree-of-freedom grasping pose parameters are output. If the current score is too low and cannot meet the grasping requirements, an alarm mechanism is triggered to remind humans to investigate the cause and prevent the robot from performing grasping operations in untrusted scenarios.

[0069] Specific embodiments of the present invention are shown below:

[0070] According to an embodiment of the present invention, the workflow of the pipeline feeding industrial robot for grasping and determining the material is divided into the following steps:

[0071] 1. Acquire image data of the work area

[0072] After the robot workstation is started, the vision acquisition device scans the loading area of ​​the material frame, conveyor line, or processing equipment to acquire 3D point cloud data including the position, orientation, and stacking status of the pipes, providing basic data for subsequent target recognition. In this embodiment, the vision acquisition device is installed about 2500mm above the material frame, and the coverage area of ​​a single scan is 6500mm×1500mm×1200mm, acquiring approximately 1.28 million points of raw point cloud data. The material frame contains 8 circular pipes to be grasped, including 4 three-inch circular pipes with an outer diameter of 73.0mm, a wall thickness of 2.11mm, a length of 6000mm, and a single piece weight of 10.01kg; 2 two-inch circular pipes with an outer diameter of 60.3mm, a wall thickness of 1.65mm, a length of 6000mm, and a single piece weight of 7.21kg; and 2 three-inch circular pipes with an outer diameter of 101.6mm, a wall thickness of 2.11mm, a length of 6000mm, and a single piece weight of 15.60kg. The pipes are randomly stacked in the material frame, and there is contact, obstruction and partial overlap between some pipes.

[0073] 2. Perform point cloud data processing

[0074] The control system filters, denoises, and trims the acquired point cloud data, removing irrelevant point clouds such as those from the material frame edges, worktable surfaces, and background objects, retaining only valid point cloud areas that may belong to the pipeline. In this embodiment, firstly, height-constrained filtering is performed on the original point cloud, retaining point clouds with heights ranging from 80mm to 780mm from the bottom of the material frame; then, the point cloud is trimmed according to the valid area of ​​the material frame, retaining points within the range of -3200mm to 3200mm in the X direction and -550mm to 550mm in the Y direction; subsequently, filtering is performed based on reflection intensity, retaining points with intensity values ​​between 35 and 220; finally, a neighborhood outlier removal method is used, treating points with fewer than 6 neighboring points within a 25mm radius as noise points and deleting them. After processing, the number of point clouds is reduced from approximately 1.28 million points to approximately 456,000 points, of which approximately 382,000 points are valid point clouds belonging to the candidate pipeline area.

[0075] 3. Identify candidate pipeline targets

[0076] The system segments and clusters the effective point cloud based on the pipe's length, diameter, cylindrical profile, and spatial distribution characteristics to obtain one or more candidate pipe targets, and calculates the center position and axial direction of each candidate pipe. In this embodiment, Euclidean distance clustering is used for the candidate region point cloud, with a clustering distance threshold set to 35mm and a minimum number of cluster points set to 2500, initially obtaining 11 candidate point clusters. Then, the centroid, covariance matrix, and principal direction are calculated for each point cluster, and its length, linearity, and compactness are estimated. Point clusters with estimated lengths less than 800mm, linearities less than 0.78, or compactness less than 0.12 are discarded, leaving 9 candidate pipe targets. For one of the sticky point clusters formed due to contact between adjacent pipes, the system calculates the local normal change rate every 100mm along the principal direction. When the direction change angle of a certain area reaches 15.8° and the abrupt change in cross-sectional spacing reaches 48mm, the point cluster is split into two independent candidate targets. Ten candidate pipe targets were ultimately obtained, eight of which corresponded to actual pipes, while the other two were incomplete candidate targets formed by partial occlusion. Taking one of the 3-inch circular pipes as an example, its estimated center position coordinates are X=1265mm, Y=-182mm, Z=458mm, the angle between the main axis direction and the length direction of the material frame is approximately 4.3°, and the estimated visible length is approximately 4720mm.

[0077] 4. Assess the feasibility of pipe grabbing

[0078] For each candidate pipe, the system analyzes its occlusion level, exposed length, surrounding gaps, and clamp interference to determine whether the pipe meets the conditions for stable grasping. If the pipe is severely occluded or the clamp cannot enter, it is temporarily not considered as the current grasping target. In this embodiment, the system calculates the visible coverage of candidate targets based on the theoretical length of 6000mm in the pipe specification library. For the aforementioned 3-inch circular pipe, its effective visible length along the main axis is 4720mm, and the visible coverage is 0.787; the system sets the visible coverage threshold to 0.70, therefore the target meets the further judgment conditions. Subsequently, five local sections are selected along the main axis of the pipe, each with a width of 50mm and a spacing of 800mm, and circular fitting is performed in the local section coordinate system. The fitting radii of the five sections are 36.4mm, 36.8mm, 36.7mm, 36.5mm, and 36.9mm, respectively, with a corresponding average fitting residual of 1.46mm and an average boundary continuity of 0.86. In this embodiment, the local cross-section fitting residual threshold is set to 3.0 mm, the boundary continuity threshold is set to 0.75, and the local geometric consistency score threshold is set to 0.70. Since the average fitting residual of this target is less than 3.0 mm and the average boundary continuity is greater than 0.75, the calculated local geometric consistency score is 0.83, thus it is determined to meet the stable grasping conditions. Conversely, for another candidate target, its effective visible length is only 3180 mm, the visible coverage is 0.53, and the boundary continuity is only 0.62, so it is determined to be severely occluded and is not considered as the current grasping target.

[0079] 5. Determine the preferred capture points

[0080] For pipes that meet the gripping conditions, the system determines candidate gripping points based on the pipe's axial direction, exposed area, and clamp structure. These candidate gripping points are then scored, and the position with the highest score is selected as the preferred gripping point for this gripping operation. In this embodiment, a double-jaw clamping clamp is used, with a 600mm distance between the two support points. The clamp is suitable for outer diameters ranging from 60mm to 120mm. The system generates five candidate gripping points within the visible area of ​​the aforementioned 3-inch circular pipe. Each gripping point must be at least 450mm from the visible edge of the end, avoiding areas with partial obstruction and areas where the distance between adjacent pipe sections is less than 25mm. The five candidate gripping points are located along the main axis at distances of 1120mm, 1680mm, 2350mm, 3010mm, and 3560mm from the left end, respectively. The system calculated the exposure, stability, interference risk, and reachability of each grab point, and combined these with the overall target assessment score to obtain final scores of 0.76, 0.84, 0.89, 0.81, and 0.73 for the five grab points. Since the third candidate grab point had the highest score of 0.89, and there were no abrupt changes in obstruction or significant interference areas within 300mm on either side of it, it was selected as the preferred grab point for this grab.

[0081] 6. Generate robot grasping posture

[0082] The control system generates the gripping pose of the robot's end effector based on the spatial coordinates of the preferred gripping point and the pipe's orientation. The gripping pose includes the gripping position, approach direction, gripper angle, and lifting direction. In this embodiment, the preferred spatial coordinates of the gripping point in the robot's base coordinate system are X=1240mm, Y=-185mm, Z=465mm, and the angle between the pipe's main axis and the X-axis of the robot's base coordinate system is 4.3°. Based on the gripper opening direction and the workpiece's orientation, the control system generates the end effector's gripping posture angle parameters as Rx=178.5°, Ry=2.0°, and Rz=4.3°, thus forming a six-degree-of-freedom gripping pose (1240mm, -185mm, 465mm, 178.5°, 2.0°, 4.3°). Meanwhile, to prevent the end clamp from colliding with adjacent pipes during approach, the system also generates a pre-grabbing pose, positioned 120mm directly above the gripping point, i.e. (1240mm, -185mm, 585mm, 178.5°, 2.0°, 4.3°). The clamp's opening width for this 3-inch round pipe is set to 76mm, and the single-sided clamping force is set to 120N, which can meet the stable clamping requirements of a 10.01kg pipe.

[0083] 7. Perform the pipe grabbing action.

[0084] The industrial robot moves to the target pipe according to the generated gripping posture. The end effector approaches the pipe along a preset approach direction and performs clamping, suction, or lifting actions upon reaching the gripping position, achieving stable gripping of the pipe. In this embodiment, the robot first moves from the initial safe position to the pre-gripping posture, then descends 120mm along the negative Z-axis at an approach speed of 80mm / s to reach the gripping posture. After reaching the position, the gripper performs a closing action, gripping a circular pipe with an outer diameter of 73.0mm. The feedback opening width after the gripper closes is 72.8mm, and the gripping force feedback range is 118N to 129N, which is within the preset allowable range. After gripping, the robot first slowly raises 50mm at a speed of 60mm / s for preliminary lifting detection. If the gripping force fluctuation is less than 20N and the workpiece does not slip, it continues to raise to a safe height. During this process, the system detected a maximum gripping force feedback fluctuation of 16N, indicating stable gripping, and therefore the gripping was determined to be successful.

[0085] 8. Transport and place pipes

[0086] The robot lifts the grasped pipe to a safe height, avoiding the material frame, surrounding pipes, and equipment structures, then moves to the target loading position and places the pipe onto the processing equipment according to a preset placement posture. In this embodiment, the robot lifts the pipe to a safe height of Z=760mm, avoiding the edge of the material frame and surrounding ungrabbed pipes. The robot then transports the pipe to the loading station, with the target placement position coordinates being X=3860mm, Y=420mm, Z=318mm, and placement posture angles of Rx=180.0°, Ry=0.0°, and Rz=0.0°. During placement, the robot first moves to a position 100mm above the placement point, then descends to the placement position at a speed of 60mm / s. The gripper gradually reduces its clamping force to below 20N and fully opens, completing the stable placement.

[0087] 9. Execute feedback results and update data.

[0088] After the pipe is placed, the control system receives execution feedback from the robot and the gripper to determine whether the gripping was successful. If successful, the current pipe status within the material frame is updated; if unsuccessful, data is re-acquired and the gripping strategy is adjusted. In this embodiment, the system uses the following success criteria: no collision alarm on the robot trajectory, gripper opening width between 70mm and 76mm, gripping force between 100N and 150N, force feedback fluctuation during handling not exceeding 30N, and visual re-inspection confirming the disappearance of the target location point cloud. After the gripping is completed, the vision system re-scans the material frame area, acquiring approximately 1.19 million points in the original point cloud, which, after preprocessing, yields approximately 431,000 valid point clouds. The system confirms that the original target pipe has been removed from the candidate set, updates the number of remaining visible pipes within the material frame to 7, and re-performs identification, local geometric consistency determination, fusion determination, and gripping point evaluation on the remaining pipes. If the maximum grasping point score is detected to be lower than 0.78, the target comprehensive judgment score is lower than 0.75, or the clamp closure position is abnormal during a grasping process, the system will output a grasping failure or manual review signal to prevent the robot from performing erroneous grasping operations in low-confidence scenarios.

[0089] 10. Complete the loading operation in a loop.

[0090] The system repeatedly executes the above identification, judgment, grasping, and placement steps until all pipes in the material frame are loaded or all processed pipes are unloaded, thereby achieving automated operation of pipe loading. In this embodiment, when continuously loading the above 8 pipes, the system performs 8 rounds of visual recognition and grasping decisions. The first 6 rounds were successfully identified and grasped on the first attempt. In the last 2 rounds, due to significant obstruction from the lower pipes, the system re-identified and grasped the pipes after the upper pipes were removed. During the entire loading process, each visual recognition and grasping decision takes approximately 5 to 8 seconds, each robot grasping and handling takes approximately 125 seconds, and the total time for loading all 8 pipes is approximately 1064 seconds. Throughout the entire implementation, the system did not experience robot collisions, pipe drops, or significant swaying, indicating that the method described in this invention can achieve relatively stable automatic grasping and loading under conditions of random pipe stacking, partial obstruction, and missing point clouds.

[0091] It should be noted that the system in this invention can be an industrial control computer connected to a 3D camera; a stability index is defined. Define the risk of interference and define accessibility metrics It can be defined according to specific requirements.

[0092] It should be recognized that embodiments of the present invention may be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium.

[0093] The method can be implemented using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program in the computer program, wherein the storage medium is configured such that the computer operates in a specific and predefined manner.

[0094] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system; however, if required, the program can be implemented in assembly or machine language.

[0095] In any case, the language can be either compiled or interpreted.

[0096] Furthermore, for this purpose, the program can run on programmed application-specific integrated circuits.

[0097] The processes described herein (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.

[0098] Furthermore, the method can be implemented in any suitable computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices.

[0099] Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether portable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein.

[0100] Furthermore, machine-readable code, or parts thereof, can be transmitted via wired or wireless networks.

[0101] When such media includes instructions or programs that combine with a microprocessor or other data processor to implement the steps described above, the invention described herein includes these and other different types of non-transitory computer-readable storage media.

[0102] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A robot-based stable gripping and control method for pipe fitting feeding, characterized in that, include: Step 1: Collect and preprocess scene data for the robot's grasping operation for loading nuclear power plant pipe parts; Step 2: Within the candidate nuclear power pipeline area, extract candidate targets based on the effective point cloud set; Step 3: After obtaining the set of candidate pipeline targets, perform local geometric consistency estimation to obtain candidate targets that meet the four set conditions; Step 4: After obtaining candidate targets that meet the above four conditions, perform a geometric feature and artificial intelligence semantic fusion judgment; Step 5: After performing the geometric feature and artificial intelligence semantic fusion judgment, execute the crawling evaluation output; Step 3 specifically includes: Let the theoretical length of the standard pipe model be... For the first A cluster of candidate points as candidate targets Let its effective visible length along the principal axis be... The visibility coverage rate of the candidate target is then defined as follows. ,in Used to characterize the effective visibility of candidate targets along the principal axis; when a candidate target satisfies the visibility coverage condition. When a candidate target is deemed to have the basis for further local geometric consistency estimation, for candidate targets that meet the visibility coverage condition, one or more local sections are selected along its principal axis, and the points on the local sections are projected onto a local section coordinate system perpendicular to the principal axis. The circle fitting relationship is expressed as follows: ,in The coordinates of the center of the circle for local cross-section fitting. The fitted radius of the local cross section is... and Let the set of points on the cross section be the two-dimensional abscissa and ordinate of a single point on the cross section in the local cross section coordinate system. Further define the fitting residual as ,in M is used to characterize the deviation between the local cross-sectional geometry of the candidate target and the ideal circular cross-section, where M is the total number of points within the current local cross-section participating in circle fitting. and These are the two-dimensional abscissa and ordinate of the m-th point within the current local section in the local section coordinate system, respectively. Step 3 also includes: Define the boundary continuity index as ,in This indicates the number of points in the current local section that satisfy the continuous boundary conditions. This represents the total number of points in the current local section. Furthermore, a local geometric consistency score is constructed. ,in When satisfied , , , Only under these four conditions will the local geometric consistency score of the candidate target be used as a valid input for subsequent judgment. This represents the maximum allowable fitting residual threshold for the set local cross-section. This indicates the set boundary continuity threshold. This represents the set local geometric consistency score threshold. For candidate targets that pass the local geometric consistency estimation, i.e., meet the four conditions, they are retained as valid inputs for subsequent geometric feature modeling, semantic fusion determination, and crawling point evaluation. For candidate targets that fail the local geometric consistency estimation, they are directly removed from the subsequent crawling candidate set. In step 4, after obtaining candidate targets that meet the four conditions, the candidate targets are judged from both the geometric and semantic sides, and a comprehensive credibility result for crawling execution is formed through a fusion mechanism. The specific implementation method includes: For the candidate target that satisfies the above four conditions, the first one candidate targets Extract its geometric feature vector ,in Indicates the target length. Indicates the target diameter. Indicates the spindle direction parameter. Indicates the roundness characteristics of the cross section. Indicates the surface normal distribution characteristics, Indicates visible coverage. Indicates boundary continuity, The local fitting residual is denoted as the normalized residual after normalizing each geometric feature. One geometric feature is Geometric decision scores are constructed based on normalized geometric features. ,in The weights corresponding to the m-th geometric feature satisfy the following conditions: , ; Input the point cloud clusters, depth maps, or projection views corresponding to the candidate targets into the artificial intelligence semantic recognition model to obtain semantic output vectors. ,in Indicates the target category output. Indicates the confidence level for category identification. This indicates the popularity of the candidate capture area. The result of the occlusion estimation is represented by the normalization of the semantic features, and the normalized result is denoted as the i-th. The semantic features are Constructing semantic judgment scores based on normalized semantic features ,in For the corresponding weights, the following should be satisfied: ; The fusion weights are dynamically adjusted based on the geometric completeness and semantic credibility of the candidate targets; that is, a semantic credibility term is first constructed. ,in The semantic occlusion correction coefficient is used, and then the geometric reliability term is constructed. ,in , , Let the weighting coefficients satisfy: =1, defining dynamic fusion weights based on semantic reliability and geometric reliability terms. in, For balance coefficient, To prevent extremely small positive numbers with a denominator of zero, and ; Construct a basic fusion score based on geometric and semantic decision scores. Furthermore, to enhance the ability to suppress geometric decision-making and semantic conflict scenarios, a consistency compensation term is introduced. And construct the corrected fusion result: ,in This is the conflict penalty coefficient. This is the local geometric consistency compensation coefficient. The obtained local geometric consistency score; Set a comprehensive judgment threshold When the candidate target satisfies: When, retain it as a valid candidate for crawling, when At the same time, a second judgment is performed by combining the degree of occlusion, the integrity of the boundary and the continuity of the candidate grabbing area. If the candidate target still has a stable local grabbing area, its grabbing priority is reduced and it is retained; otherwise, it is eliminated.

2. The robot stable gripping and control method for pipe fitting feeding according to claim 1, characterized in that, Step 1 specifically includes: First, a 3D camera was used to collect point cloud data of the work area where the robot was loading materials onto the nuclear power plant pipeline, resulting in the original point cloud set: ,in , represents the 3D coordinates of the i-th point in the original point cloud set in the camera coordinate system. This represents the grayscale value or reflection intensity value of the i-th point in the original point cloud set, and N represents the total number of points in the original point cloud.

3. The robot stable gripping and control method for pipe fitting feeding according to claim 2, characterized in that, Step 1 specifically includes: firstly, performing height constraint filtering on each point in the original point cloud set, retaining points that satisfy the following formula: ,in, and These represent the lower and upper limits of the working layer's height, respectively; further filtering based on grayscale or reflection intensity yields: ,in, and These are the lower and upper limits of the set intensity filtering threshold, respectively. This represents the grayscale value or reflection intensity value of the i-th point in the original point cloud set; further, outlier removal is performed by combining the point neighborhood density, for any point... The number of its neighborhood points is defined as: ,in For point and The three-dimensional Euclidean distance between them, when At that time, the point These are considered discrete noise points and discarded. The threshold for the number of points in the smallest neighborhood. This is the set neighborhood radius threshold.

4. The robot stable gripping and control method for pipe fitting feeding according to claim 3, characterized in that, Step 2 specifically includes: Within the candidate nuclear power pipeline area, the effective point cloud set is... Perform Euclidean distance clustering to obtain an effective set of point clouds. Any two points and Euclidean distance between Defined as: When satisfied When the two points are grouped into the same connected point cluster, that is, the candidate point cluster, then... This represents the set clustering distance threshold, from which the candidate target set is obtained. ,in Indicates the first There are 3 candidate point clusters, k = 1, 2, ..., n, where n is the number of candidate point clusters. For each candidate point cluster... Furthermore, its geometric and statistical characteristics are calculated.

5. The robot stable gripping and control method for pipe fitting feeding according to claim 4, characterized in that, Step 2 further includes: methods for further calculating its geometric statistical characteristics, specifically including: Let its center of mass be for: And construct the covariance matrix The three eigenvalues ​​are obtained by solving the covariance matrix. , , , Further define the linearity index and firmness index When the candidate point cluster satisfies the preset constraints on length range, number of points, linearity, and compactness, that is: , , , This is retained as a candidate pipeline target, and all candidate pipeline targets constitute the candidate pipeline target set, where L k This represents the estimated length of the point cluster along the main direction. The minimum effective length threshold for nuclear power plant pipelines is set. The maximum effective length threshold for nuclear power plant pipelines is set. For the first The total number of points in each candidate point cluster. and These are the set linearity threshold and compactness threshold, respectively. This is the minimum point cloud quantity threshold required to determine whether a nuclear power pipeline is valid.

6. The robot stable gripping judgment and control method for pipe fitting feeding according to claim 5, characterized in that, Step 5 specifically includes: For the valid candidate targets after screening, the system further evaluates their candidate capture points, that is, let the first... The first candidate target The crawl points are A crawling evaluation function is established for it, which includes: First, define the exposure index. It is used to characterize the degree of exposure of the gripping area and the accessibility of the grippers, and defines a stability index. It is used to characterize the risk of the target rolling, swaying, or falling off during the grasping process, and defines the interference risk. It is used to characterize the probability of the grasping path colliding with surrounding workpieces or material frame structures, and defines the reachability index. The grasp point score is used to characterize the pose realizability of the robot's end effector at the current position. ,in The obtained target comprehensive judgment score Introducing crawl point-level decision-making, we obtain the final crawl point score: ,in, The set fusion judgment compensation coefficient; the system applies the following to all candidate capture points: Sort the results and select the highest-rated result as the current preferred crawl point: For the selected preferred grasping point, output its spatial coordinates and attitude parameters to form the six-degree-of-freedom grasping pose of the robot's end effector: in, Indicates the coordinates of the grab point location. The corresponding attitude angle parameters are represented, and the pose information is sent to the robot control system to guide the end effector to complete a stable grasping action. When the maximum grasping score is lower than a preset threshold... When this happens, the system outputs a signal indicating a failure to capture or requiring manual review, in order to prevent the robot from performing erroneous capture operations in low-confidence scenarios.

Images