Multi-source coupled error follow-up compensation method for robot dynamic grasping
By constructing an end-to-end error compensation model, and combining a differentiable Gaussian mixture model and a sparse variational Gaussian process model, the problem of multi-source error coupling in robot dynamic grasping was solved, achieving high-precision error compensation and stability improvement.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- DALIAN UNIV OF TECH
- Filing Date
- 2026-05-19
- Publication Date
- 2026-08-04
AI Technical Summary
In existing robotic dynamic grasping, multi-source error coupling is difficult to compensate effectively, resulting in insufficient positioning accuracy and stability. Traditional methods cannot balance global modeling accuracy and computational complexity, and are prone to getting trapped in local optima.
An end-to-end error compensation method is constructed that integrates a differentiable Gaussian mixture model and a sparse variational Gaussian process model. By collecting multi-source heterogeneous data, performing feature selection and partition modeling, it can achieve accurate modeling and real-time prediction of multi-source errors and perform dynamic compensation during visual guidance.
It significantly improves the positioning accuracy and stability of the robot during dynamic grasping, enhances the end-positioning accuracy and grasping success rate of vision-guided grasping tasks, simplifies the model implementation, and has good engineering applicability.
Smart Images
Figure CN122231906B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of robot error compensation technology, and in particular to a multi-source coupling error follow-up compensation method for dynamic grasping of robots. Background Technology
[0002] In the aerospace field, composite robots are now widely used for the inspection and repair of large components. Precise target localization and guidance are crucial for these robots to complete tasks such as inspection, processing, and scheduling. However, AGVs (Automated Guided Vehicles), as the mobile devices of composite robots, have low repeatability (20mm), making traditional teaching-based localization methods unsuitable. To address this issue, composite robots employ binocular cameras for target localization and guidance; however, the coupling of multi-source errors between the vision system and the robot system during the process leads to visual guidance failure. Therefore, to ensure the accuracy of visual guidance, compensation for multi-source errors in the process is necessary.
[0003] Gaussian Process Regression (GPR) has natural advantages in nonlinear function approximation, few-shot learning, and uncertainty quantification, but its training and inference costs are high in high-dimensional, large-sample scenarios. Therefore, it is necessary to combine strategies such as feature selection, dimensionality reduction, and local modeling to ensure online performance.
[0004] The patent "A Waterborne Adhesive Coating Control Method and System Based on Artificial Intelligence Optimization" (patent number CN120630740A) proposes to model coating history data and coating quality using Gaussian process regression, which can achieve real-time and efficient prediction of coating quality and generate optimal control actions to improve overall quality. However, this method does not take into account the execution error of the robot system and does not perform error modeling for the entire system.
[0005] Tao Chao, Junxi Tian, Ming Yang. A Gaussian Process Regression-Based Multitarget Orientation Compensation Algorithm for Distance Measurement[J].IEEE Transactions on Instrumentation and Measurement, 2024, 739516111.1-9516111.11. This paper proposes a Gaussian Process Regression-Based Multitarget Orientation Compensation Algorithm (GPRMOCA), which can calculate distance compensation values and uncertainties to accurately compensate for measurement deviations in multitarget localization. However, it only studies the ranging error caused by the target orientation, and other error sources (such as sensor noise, mechanical jitter, and temperature drift) are not included in the unified modeling and compensation framework, which limits the overall accuracy improvement potential.
[0006] Therefore, this invention proposes a multi-source coupling error follow-up compensation method for dynamic grasping of robots. Through feature selection, partitioned modeling, and optimization with ARD kernel functions, error analysis of the visual guidance process is achieved, and the GPR error prediction model is trained. Error compensation is then performed based on the calculation results of the visual guidance, effectively compensating for coupling errors caused by perception, coordinate transformation, and execution. Summary of the Invention
[0007] The technical challenge addressed by this invention is the difficulty in effectively compensating for multi-source coupling errors in existing robotic dynamic grasping methods. Furthermore, it highlights the difficulty in balancing global modeling accuracy and computational complexity, and the tendency of traditional two-stage strategies (partitioning followed by modeling) to get trapped in local optima. This results in insufficient positioning accuracy and stability for composite robots in grasping and assembly operations. The invention proposes a dynamic compensation method for multi-source coupling errors in robotic dynamic grasping. This method collects intermediate variables such as camera measurement data, coordinate transformation information, and robot operating status. It constructs an end-to-end model fusing a differentiable Gaussian mixture model and a sparse variational Gaussian process model to achieve accurate modeling and real-time prediction of coupling errors. Dynamic compensation for visual guidance positions is then performed during robot operation, significantly improving the positioning accuracy and stability of the operation.
[0008] The technical solution of this invention: A multi-source coupling error follow-up compensation method for dynamic grasping of robots, the specific steps of which are as follows: Based on a composite robot vision guidance system, intermediate variables during task execution are simultaneously acquired by the robot and a vision camera to construct a multi-source heterogeneous dataset. This dataset includes the predicted end-effector pose information output by the vision camera and the robot's motion execution state information. The actual end-effector pose information when the gripper is fully engaged with the target object is obtained through teaching. The difference between the actual end-effector pose information and the predicted end-effector pose information output by the vision camera is calculated to obtain the corresponding multi-dimensional error, thus constructing a multi-source error dataset. The multi-source error dataset is then preprocessed to obtain a standardized training dataset. The standardized training dataset is used as input features. Based on the model's partitioning of the network, the input features are calculated. x Weight in workspace sub-regions Prediction results of variational sparse Gaussian process regression models for each workspace sub-region Weighted fusion is performed; the model partitioning network is a differentiable Gaussian mixture model, which is composed of... Composed of Gaussian distribution components This represents the total number of sub-regions in the workspace. Each Gaussian distribution component corresponds to one sub-region in the workspace, resulting in the final prediction result. : Construct an end-to-end error compensation model; combine workspace sub-region weights And the variational distribution of each variational sparse Gaussian process regression model, constructing a weighted variational lower bound as the overall optimization objective function: Where N represents the total number of samples in the standardized training dataset. To indicate the first n Input features x n Belongs to the k The weight of each workspace sub-region For the first k The expected log-likelihood of the variational sparse Gaussian process regression model; KL(·) is the Kullback-Leibler penalty term used to constrain the . k Variational posterior distribution of a variational sparse Gaussian process regression model Approximating its prior distribution ; The Adam adaptive optimizer is invoked to uniformly update the end-to-end error compensation model; in each iteration, the model partition network parameters are simultaneously optimized. μ k , L k , πk}, the kernel function hyperparameters of the variational sparse Gaussian process regression model { σ f , l 1 ,…, l D}, σ f The standard deviation of the kernel function is used to characterize the output amplitude of the function. l 1 ,…, l D The length scale parameter corresponds to each dimension of the input feature; D is the number of dimensions of the input feature, and the location of the induced point. Z k Variational parameters { m kq , S kq}, m kq Let be the mean vector of the posterior distribution of the induced variables in the k-th variational sparse Gaussian process regression model; S kq Let be the covariance matrix of the posterior distribution of the induced variables in the regression model of the k-th variational sparse Gaussian process; π k For the first k The weights of the Gaussian distribution components satisfy the following condition: , For the first k The mean vector of each Gaussian distribution component is used to characterize the center location of the sub-region of the workspace. It is a lower triangular matrix; the above parameters are calculated with respect to the overall optimization objective function through backpropagation. The gradient is calculated, and the update is performed based on the gradient, so that the end-to-end error compensation model gradually converges to the optimal solution; During the process of visual guidance of the robot, the robot's end-effector position and the predicted end-effector pose calculated by the vision camera are recorded in real time. B P calculated , as input features x Input end-to-end error compensation model, output multidimensional error prediction value v ={Δ x ,Δ y ,Δ z The robot control system fuses the multidimensional error prediction values with the original guidance coordinates output by the vision camera to obtain the error-compensated target guidance coordinates. The calculation method is as follows: in, B Pcompensated These are the target guidance coordinates after error compensation; The robot control system generates motion control commands based on the error-compensated target guidance coordinates, driving the end effector of the robotic arm to complete the target grasping and positioning operations; during the continuous operation of the robot, the error prediction and error compensation processes are triggered and executed in real time with the visual guidance, correcting the dynamic errors in the robot's guidance path online.
[0009] The differentiable Gaussian mixture model is designed for input features. x The overall probability of the entire end-effector workspace Represented as: in, (·) represents a Gaussian distribution. The covariance matrix is used to describe the shape and scale of the data distribution within the sub-regions of the workspace; The covariance matrix is decomposed using the Cholesky method. in, It is the transpose of the lower triangular matrix; In the process of model partitioning network computation, the input features are calculated first. x In the k Logarithmic response value under Gaussian distribution components : Normalization is performed using the Softmax function to map the logarithmic response values of each Gaussian distribution component to a weight distribution that satisfies probability constraints, thereby obtaining the input features. x Belongs to the k Weight of each workspace sub-region : in, Representing input features x In the j Logarithmic response value under Gaussian distribution components.
[0010] A Gaussian mixture model based on the expectation-maximization algorithm is introduced to pre-partition the workspace of the robotic arm's end effector, and the resulting mean vector and covariance matrix are used as the initialization parameters of the model's partitioning network.
[0011] A variational sparse Gaussian process regression model is constructed in each sub-region of the workspace; For the k Each workspace sub-region is designed for input features. x Another input feature in the standardized training dataset Construct a Gaussian regression process model f k ( x Its prior distribution is defined as: in, (·) represents a Gaussian process. This is the mean of the sub-region of the workspace, which is 0 in the absence of prior knowledge; This is the kernel function for this sub-region of the workspace, used to characterize the correlation between the input feature and another input feature in the standardized training dataset; For Gaussian regression process model f k ( x ), introduce induced variables to perform a sparse approximation; by introducing a set of induced variables ,in , for the first k The set of guiding points within each sub-region of the workspace. Indicates the first k The first workspace sub-region M One induction point, a set of induction variables prior distribution Represented as: in, The kernel matrix between the induced points; Introducing variational distribution Approximation: in, For variational mean, For variational covariance, Parameterization using Cholesky decomposition ensures positive definiteness; based on variational distribution. For each input feature x ,calculate f k ( x Approximate posterior distribution of the value at this input feature : mean With variance They are respectively: in, The kernel matrix between the input features and the induced points. for The transpose of .
[0012] The kernel function is a Matern-3 / 2 kernel function with automatic correlation resolution, and its form is as follows: in, For signal variance, l For length scale, This represents the Euclidean distance between an input feature and another input feature in the standardized training dataset.
[0013] The composite robot vision guidance system includes an AGV mobile platform, a robotic arm, a vision camera, and a gripper; the vision camera is integrated on the gripper; the gripper is fixed to the end of the robotic arm; the robotic arm is mechanically connected to the AGV mobile platform.
[0014] The data preprocessing includes outlier removal, data alignment, normalization, and principal component analysis.
[0015] The beneficial effects of this invention are as follows: This invention proposes a multi-source coupling error follow-up compensation method for robot dynamic grasping. By constructing a multi-source error modeling and prediction mechanism covering the entire process of visual perception, coordinate transformation, and mechanical execution, it achieves unified representation and collaborative compensation for multiple types of errors. This method closely integrates data-driven modeling with the robot's motion process. It eliminates the need for establishing complex analytical error models or relying on precise system mechanism identification, enabling high-precision prediction and correction of comprehensive errors in the robot's dynamic grasping process. This effectively reduces the impact of incomplete models and parameter uncertainties on grasping accuracy in traditional methods.
[0016] By learning from multi-source error data under limited sample conditions, an error prediction model is established. During robot operation, real-time acquired visual measurement information, pose calculation results, and execution state parameters are input, and the corresponding error compensation amount can be output, achieving online follow-up correction during the grasping process. This method can work stably under dynamic targets, complex environments, and multi-coordinate system coupling conditions, significantly improving the end-effector positioning accuracy and grasping success rate of vision-guided grasping tasks. Furthermore, the method has a clear implementation process, is simple to implement, and is easily integrated into existing robot vision guidance systems, demonstrating good engineering applicability and promotional value. Attached Figure Description
[0017] Figure 1 This is a flowchart of a multi-source coupling error follow-up compensation method for dynamic grasping of robots.
[0018] Figure 2(a) is a diagram of multi-source error analysis; (b) is a diagram of visual measurement error; (c) is a diagram of coordinate transformation error; (d) is a diagram of robot execution error. Figure 3 The diagram shows the experimental verification platform; (a) is a composite robot vision guidance system, and (b) is a laser tracker.
[0019] In the diagram: 1-Laser tracker, 2-Target ball, 3-Gripper, 4-Vision camera, 5-Robotic arm, 6-AGV mobile platform. Detailed Implementation
[0020] The specific embodiments of the present invention are described in detail below with reference to the technical solutions and accompanying drawings.
[0021] The first step is to construct a composite robot vision guidance system; like Figure 3 As shown, a composite robot vision guidance system is constructed, including an AGV mobile platform 6, a robotic arm 5, a vision camera 4, and a gripper 3; the vision camera 4 is integrated on the gripper 3; the gripper 3 is fixed to the end of the robotic arm; the robotic arm 5 is mechanically connected to the AGV mobile platform 6.
[0022] The second step is error analysis of the robot vision guidance process. like Figure 2 As shown, the errors in the robot vision guidance process are analyzed and mainly categorized into three parts: visual measurement error, coordinate transformation error, and robot execution error. Among these, visual measurement error is as follows: Figure 2 (a) mainly occurs when the binocular camera identifies and locates the guided target, including binocular camera calibration error, binocular stereo matching error, etc., as shown in the figure. O l , O r Indicates the center point of the left and right images; coordinate transformation error is as follows: Figure 2 (b) mainly arises during the calculation of coordinate transformations between the camera coordinate system {C}, the robot end effector coordinate system {E}, and the robot base coordinate system {B}. This includes errors in camera and robot hand-eye calibration, robot DH parameter modeling errors, and visual camera and end effector offset errors caused by vibration after calibration. Robot execution errors include... Figure 2 (c) Errors mainly occur when the robot receives coordinates and performs guided grasping, including errors caused by factors such as robot inverse kinematics calculation errors, robot joint motor transmission errors, and end effector load / thermal deformation. These various errors are coupled with each other, resulting in insufficient overall positioning accuracy of the composite robot in tasks such as grasping and assembly.
[0023] The third step is to construct a multi-source error dataset and perform data preprocessing. Based on the constructed composite robot vision guidance system, the robot and the vision camera simultaneously collect intermediate variables during the task execution process to construct a multi-source heterogeneous dataset. The multi-source heterogeneous dataset includes the predicted end-effector pose information output by the vision camera and the robot's motion execution state information. The actual end-effector pose information when the gripper is fully engaged with the target object is obtained through teaching. The difference between the actual end-effector pose information and the predicted end-effector pose information output by the vision camera is calculated to obtain the corresponding multi-dimensional error, thereby constructing a multi-source error dataset.
[0024] The multi-source error dataset is preprocessed, including outlier removal, data alignment, normalization, and principal component analysis, to form a training dataset containing input features and multi-dimensional errors for subsequent error modeling.
[0025] The fourth step is to construct an end-to-end error compensation model; like Figure 1 As shown, after completing the training dataset, an end-to-end error compensation model that integrates model partitioning and local modeling is proposed to address the complex multi-source errors during the robot's dynamic grasping process. Specifically: First, to address the accuracy limitations of traditional single global models, a differentiable Gaussian Mixture Model (GMM) is constructed as the model partitioning network to divide the robotic arm's end effector workspace into partitions. Subsequently, local modeling is performed within each partition's workspace. The expectation-maximization algorithm is used to pre-partition the workspace of the differentiable Gaussian Mixture Model, and the resulting mean and covariance parameters are used to initialize the model partitioning network. The workspace is divided into k sub-regions, with each sub-region having a weight of [value missing]. .
[0026] Secondly, to address the issue of high computational cost due to the high dimensionality and complex coupling of input features within the workspace, a Sparse Variational Gaussian Process (SVGP) model is constructed in each sub-region of the workspace. This enables high-precision modeling of the input features and improves computational efficiency, resulting in the predicted values of the SVGP model within each sub-region of the workspace. .
[0027] Finally, after completing the construction of the model partitioning network and the local modeling network, this paper proposes an end-to-end error compensation model that integrates the differentiable Gaussian mixture model (GMM) and the sparse variational Gaussian process (SVGP) for collaborative optimization. This model achieves joint learning of the model partitioning network and the local modeling network by unifying the optimization objective, thereby effectively improving the global error fitting accuracy and the model generalization ability.
[0028] The predicted value of the input feature across the entire workspace is And it can be represented as a forward propagation network as: Construct the following weighted variational lower bound (Mixture ELBO) as the overall optimization objective function: in, For the first k The expected log-likelihood of the SVGP model is used to measure the reconstruction and fitting ability of local error data; KL(·) is the Kullback-Leibler penalty term used to constrain the SVGP model. k Variational posterior distribution of an SVGP model Approximating its prior distribution This helps prevent overfitting of local models.
[0029] Then, the Adam adaptive optimizer is invoked to uniformly update the model parameters. In each iteration, the parameters of the differentiable GMM network are simultaneously optimized. μ k , L k , π k}, SVGP kernel function hyperparameters { σ f , l 1 ,…, l D}, location of the induction point Z k Variational parameters { m kq , S kq The above parameters are calculated with respect to the overall loss function via backpropagation. The gradient is calculated, and updates are performed based on the gradient, so that the model gradually converges to the optimal solution.
[0030] The fifth step is real-time error compensation and precise robot vision guidance. After completing the construction of the end-to-end error compensation model, it is used for real-time prediction and online compensation of visual guidance errors. Specifically, during the robot's visual guidance process, the robot's end-effector working position and the predicted end-effector pose calculated by the visual camera are recorded in real time. B P calculated , as input features x Input the pre-built end-to-end error compensation model, and the model outputs multidimensional error prediction values. v ={Δ x ,Δ y ,Δ z The robot control system fuses the predicted error value with the original guidance coordinates output by the vision camera to obtain the error-compensated target guidance coordinates. The calculation method is as follows: in, B P compensated These are the final target guidance coordinates after error compensation; The robot control system generates motion control commands based on the compensated target guidance coordinates, driving the end effector of the robotic arm to complete the target grasping and positioning operations. During the continuous operation of the robot, the above error prediction and compensation process is triggered and executed in real time with the visual guidance task, realizing online correction of dynamic errors in the robot guidance path, thereby improving the positioning accuracy of the robot end effector and the consistency of visual guidance, and completing the robot's precise visual guidance.
[0031] Step 6: Verification of the guidance error compensation effect; Figure 3 This is a schematic diagram for verifying the accuracy of the guidance error. The verification platform consists of a laser tracker 1 (LT, Leica ATS600), the composite robot vision guidance system constructed in the first step, a target ball mount (SMN), a target ball 2 (SMR) positioned by the LT, and a target ball 2 (SBR) positioned by vision. Six SMNs are arranged on the experimental support in a non-collinear, non-coplanar manner. First, the position of the SMR measured by the laser tracker is used as the true value. Then, the SBR is visually identified and guided to the calculated position. Finally, the end effector center is replaced with the SMR, and the position measured by the laser tracker is used as the calculated value. This allows the calculation of the error before and after GPR model compensation.
[0032] The results show that after GPR error compensation, the average visual guidance errors in each direction are 0.813 mm, 0.783 mm, and 0.422 mm, respectively. Compared with before error compensation, the guidance accuracy in each direction is improved by 94.27%, 88.80%, and 89.99%, respectively. Therefore, the real-time follow-up compensation method for multi-source errors in the robot's dynamic grasping process is feasible.
Claims
1. A multi-source coupled error follow-up compensation method for robot dynamic grasping, characterized in that, The specific steps are as follows: Based on a composite robot vision guidance system, intermediate variables during task execution are simultaneously collected by the robot and a vision camera to construct a multi-source heterogeneous dataset. The multi-source heterogeneous dataset includes the predicted end-effector pose information output by the vision camera and the robot's motion execution state information. The actual end-effector pose information when the gripper is fully engaged with the target object is obtained through teaching. The difference between the actual end-effector pose information and the predicted end-effector pose information output by the vision camera is calculated to obtain the corresponding multi-dimensional error and construct a multi-source error dataset. The multi-source error dataset is preprocessed to obtain a standardized training dataset; The standardized training data set is taken as an input feature, and a model partition network is used to calculate the input feature x The weights of the workspace sub-regions The prediction results of the variational sparse Gaussian process regression model of each workspace sub-region The weighted fusion is performed; the model partition network is a differentiable Gaussian mixture model, and the differentiable Gaussian mixture model is composed of Gaussian distribution components, The total number of workspace sub-regions is represented, each Gaussian distribution component corresponds to a workspace sub-region, and a final prediction result is obtained : Constructing an end-to-end error compensation model; combining workspace sub-region weights And the variational distribution of each variational sparse Gaussian process regression model, construct a weighted variational lower bound as the overall optimization objective function: Where N represents the total number of samples in the standardized training dataset. To indicate the first n Input features x n Belongs to the k The weight of each workspace sub-region For the first k The expected log-likelihood of the variational sparse Gaussian process regression model; KL(·) is the Kullback-Leibler penalty term used to constrain the . k Variational posterior distribution of a variational sparse Gaussian process regression model Approximating its prior distribution ; The Adam adaptive optimizer is invoked to uniformly update the end-to-end error compensation model; in each iteration, the model partition network parameters are simultaneously optimized. μ k , L k , π k }, the kernel function hyperparameters of the variational sparse Gaussian process regression model { σ f , l 1 ,…, l D }, σ f The standard deviation of the kernel function is used to characterize the output amplitude of the function. l 1 ,…, l D The length scale parameter corresponds to each dimension of the input feature; D is the number of dimensions of the input feature, and the location of the induced point. Z k Variational parameters { m kq , S kq }, m kq Let be the mean vector of the posterior distribution of the induced variables in the k-th variational sparse Gaussian process regression model; S kq Let be the covariance matrix of the posterior distribution of the induced variables in the regression model of the k-th variational sparse Gaussian process; π k For the first k The weights of the Gaussian distribution components satisfy the following condition: , For the first k The mean vector of each Gaussian distribution component is used to characterize the center location of the sub-region of the workspace. It is a lower triangular matrix; the above parameters are calculated with respect to the overall optimization objective function through backpropagation. The gradient is calculated, and the update is performed based on the gradient, so that the end-to-end error compensation model gradually converges to the optimal solution; During the process of visual guidance of the robot, the robot's end-effector position and the predicted end-effector pose calculated by the vision camera are recorded in real time. B P calculated , as input features x Input end-to-end error compensation model, output multidimensional error prediction value v ={Δ x ,Δ y ,Δ z The robot control system fuses the multidimensional error prediction values with the original guidance coordinates output by the vision camera to obtain the error-compensated target guidance coordinates. The calculation method is as follows: in, B P compensated These are the target guidance coordinates after error compensation; The robot control system generates motion control commands based on the error-compensated target guidance coordinates, driving the end effector of the robotic arm to complete the target grasping and positioning operations; during the continuous operation of the robot, the error prediction and error compensation processes are triggered and executed in real time with the visual guidance, correcting the dynamic errors in the robot's guidance path online.
2. The multi-source coupling error follow-up compensation method for robot dynamic grasping according to claim 1, characterized in that, The differentiable Gaussian mixture model is designed for input features. x The overall probability of the entire end-effector workspace Represented as: in, (·) represents a Gaussian distribution. The covariance matrix is used to describe the shape and scale of the data distribution within the sub-regions of the workspace; The covariance matrix is decomposed using the Cholesky method. in, It is the transpose of the lower triangular matrix; In the process of model partitioning network computation, the input features are calculated first. x In the k Log response value of Gaussian distribution components : Normalization is performed using the Softmax function to map the logarithmic response values of each Gaussian distribution component to a weight distribution that satisfies probability constraints, thereby obtaining the input features. x Belongs to the k Weight of each workspace sub-region : in, Representing input features x In the j Logarithmic response value under Gaussian distribution components.
3. The multi-source coupling error follow-up compensation method for robot dynamic grasping according to claim 2, characterized in that, A Gaussian mixture model based on the expectation-maximization algorithm is introduced to pre-partition the workspace of the robotic arm's end effector, and the resulting mean vector and covariance matrix are used as the initialization parameters of the model's partitioning network.
4. The multi-source coupling error follow-up compensation method for robot dynamic grasping according to claim 1, characterized in that, A variational sparse Gaussian process regression model is constructed in each sub-region of the workspace; For the k Each workspace sub-region is designed for input features. x Another input feature in the standardized training dataset Construct a Gaussian regression process model f k ( x Its prior distribution is defined as: in, (·) represents a Gaussian process. This is the mean of the sub-region of the workspace, which is 0 in the absence of prior knowledge; This is the kernel function for this sub-region of the workspace, used to characterize the correlation between the input feature and another input feature in the standardized training dataset; For Gaussian regression process model f k ( x ), introduce induced variables to perform a sparse approximation; by introducing a set of induced variables ,in , for the first k The set of guiding points within each sub-region of the workspace. Indicates the first k The first workspace sub-region M One induction point, a set of induction variables prior distribution Represented as: in, The kernel matrix between the induced points; Introducing variational distribution Approximation: in, For variational mean, For variational covariance, Parameterization using Cholesky decomposition ensures positive definiteness; based on variational distribution. For each input feature x ,calculate f k ( x Approximate posterior distribution of the value at this input feature : mean With variance They are respectively: in, The kernel matrix between the input features and the induced points. for The transpose of .
5. The multi-source coupling error follow-up compensation method for robot dynamic grasping according to claim 4, characterized in that, The kernel function is a Matern-3 / 2 kernel function with automatic correlation determination, and the kernel function form is as follows: in, For signal variance, l For length scale, This represents the Euclidean distance between an input feature and another input feature in the standardized training dataset.
6. The multi-source coupling error follow-up compensation method for robot dynamic grasping according to claim 1, characterized in that, The composite robot vision guidance system includes an AGV mobile platform (6), a robotic arm (5), a vision camera (4), and a gripper (3); the vision camera (4) is integrated on the gripper (3); the gripper (3) is fixed to the end of the robotic arm (5); the robotic arm (5) is mechanically connected to the AGV mobile platform (6).
7. The multi-source coupling error follow-up compensation method for robot dynamic grasping according to claim 1, characterized in that, The data preprocessing includes outlier removal, data alignment, normalization, and principal component analysis.