A progressive error compensation method for robotic vision point cloud registration

Abstract

The application discloses a progressive error compensation method for robot vision point cloud registration, relates to the technical field of industrial robot vision and vision point cloud registration, and solves the problems of the limitations of a single linear hand-eye correction model, etc.The original calibration data set is collected in an offline mode, a hand-eye transformation matrix is solved, and an MLP network is constructed, error compensation is realized through an online scanning mode, the physical cause-and-effect relationship that the rotation error is a function of the spatial translation position is utilized, a cascade sequential decoupling architecture is constructed, translation linear independent solving, translation results are taken as prior conditions to input the MLP, and the MLP maps the residual rotation error, the modeling logic of error compensation is changed, and the model structure is matched with the physical law of the error by changing from mixed-dimension synchronous solving to cause-and-effect cascade conditional modeling.

Description

Technical Field

[0001] This invention relates to the field of industrial robot vision and visual point cloud registration technology, specifically to a progressive error compensation method for robot visual point cloud registration. Background Technology

[0002] For a robot 3D scanning system with a vision sensor mounted on the end effector of a robotic arm, the general processing flow is as follows: First, acquire multiple sets of calibration pose data to complete linear hand-eye calibration, and solve for the rigid body linear transformation matrix of the vision sensor relative to the end effector of the robotic arm; then, based on this linear transformation matrix, uniformly transform the original point cloud in the vision sensor coordinate system to the robot's base coordinate system to achieve point cloud alignment. This entire linear processing flow assumes that all system errors follow a linear variation law, which has significant limitations in practical applications.

[0003] 1. Existing multi-source error coupling leads to misalignment of point cloud layers;

[0004] Multi-source error coupling causes misalignment of point cloud layers. The core errors originate from robot positioning deviation, visual sensor lens imaging distortion, parallax measurement errors, calibration residuals, etc. These various errors are coupled with each other to form a nonlinear composite deviation. After fitting with a linear calibration model, the errors cannot be effectively compensated, resulting in a large residual layering error.

[0005] 2. Ignoring the essential differences between translation and rotation, and using mixed-dimensional synchronous solution leads to model defects;

[0006] Current techniques for handling multi-source errors treat translation and rotation errors as six independent variables solved simultaneously. However, translation and rotation errors have different generation mechanisms and variation characteristics; the distribution of rotation error is influenced by the spatial translation position. Directly solving these errors simultaneously in mixed dimensions easily leads to rotational nonlinear components interfering with the translation fitting results. Current techniques, by solving translation and rotation simultaneously in the same model, fail to recognize the physical causal relationship that rotation error is a function of spatial translation position. This results in translation results being contaminated by rotational nonlinear errors, or translation information being merely input into the neural network as ordinary input features without exerting its constraint on rotation errors. Consequently, the model's fitting ability is limited.

[0007] Meanwhile, under the existing technology's simultaneous translation and rotation solution paradigm, even with the introduction of neural networks to fit the six-dimensional error as a whole, there are still problems such as a large number of model parameters, high training sample requirements, and long inference time, making it difficult to meet the requirements of real-time scanning in industrial settings.

[0008] 3. Hardware upgrades are costly;

[0009] Relying on high-precision robots and calibration equipment requires significant hardware investment and cannot fundamentally eliminate the complex deviations caused by the coupling of positioning, imaging, parallax, and calibration.

[0010] 4. Poor point cloud registration after correction;

[0011] Even after linear model correction, the point cloud still has a large initial layering deviation, which exceeds the effective convergence range of the Iterative Closest Point (ICP) registration algorithm. Directly performing global registration is prone to getting trapped in local optima, making it difficult to achieve complete point cloud alignment. The final reconstructed 3D model cannot meet the needs of subsequent measurement, reverse modeling, and other applications.

[0012] In summary, existing correction schemes that rely solely on a single linear model cannot adapt to the nonlinear error conditions caused by multi-source coupling in the system. The residual deviation after correction is large and the point cloud registration stability is poor. There is an urgent need for a data-driven progressive error compensation method for multi-source coupling errors. Summary of the Invention

[0013] To address the limitations of existing single linear hand-eye calibration models, namely: multi-source error coupling leading to misalignment of point cloud layers; neglecting the essential differences between translation and rotation, resulting in model defects due to mixed-dimensional synchronous solution; poor point cloud registration effect after calibration; and high hardware upgrade costs, this invention provides a progressive error compensation method for robot vision point cloud registration.

[0014] A progressive error compensation method for robot visual point cloud registration is proposed. This method involves offline acquisition of the original calibration dataset, solving the hand-eye transformation matrix, and constructing an MLP network, followed by online scanning to achieve error compensation. The method comprises the following steps:

[0015] Step 1: Collect the original calibration dataset;

[0016] Step 2: Solve the hand-eye transformation matrix by performing hand-eye calibration iterative calculations on the calibration dataset described in Step 1, and generate a paired dataset with error labels;

[0017] Step 3: Construct a translational linear model using the paired dataset described in Step 2; output the translational compensation amount through the translational linear model;

[0018] Step 4: The translation compensation amount is spliced ​​with the six-dimensional pose of the robot's end effector as the input of the MLP network, and the rotation compensation amount is output through the MLP network.

[0019] Step 5: Using the hand-eye transformation matrix described in Step 2, the translational linear model in Step 3, and the MLP network in Step 4, real-time correction is achieved online to complete error compensation.

[0020] The beneficial effects of this invention: The error compensation method described in this invention utilizes the physical causal relationship that rotation error is a function of spatial translation position, constructing a cascaded sequential decoupled architecture where translation is solved linearly and independently, the translation result is used as a priori condition input to the MLP, and the MLP maps the rotation residual error. Unlike existing technologies that treat translation and rotation as parallel variables and solve them simultaneously in the same model, this invention transforms the translation component from the output target to be fitted into conditional prior input features; and reduces the dimensionality of the MLP network from six-dimensional synchronous fitting to three-dimensional conditional mapping. This invention changes the modeling logic of error compensation, shifting from mixed-dimensional synchronous solution to causal cascaded conditional modeling, making the model structure match the physical laws of the error. Specifically, it possesses the following technical advantages:

[0021] 1. This invention employs Singular Value Decomposition (SVD) for rigid body transformation and linear model solution for translation, combined with a progressive error compensation architecture using an MLP network to compensate for residual rotation errors. This effectively suppresses point cloud layer misalignment and can control the initial layer deviation of the point cloud within the convergence radius of the ICP registration algorithm. In preferred cases, the translation deviation can be ≤1mm and the rotation deviation ≤2°. Compared with the pure linear hand-eye calibration scheme, the point cloud registration success rate is expected to be improved by more than 60%.

[0022] 2. The error compensation method described in this invention uses the translation compensation amount obtained from the linear model as a priori input to participate in network training. The MLP network only needs to fit the three-dimensional rotation compensation amount. The number of network parameters is small and the inference operation speed is fast. It can achieve millisecond-level processing speed on ordinary industrial control computers, which meets the usage cycle of high-speed industrial scanning.

[0023] 3. The error compensation method described in this invention has good sample utilization efficiency and can still achieve good compensation effect under the condition of relatively sparse training samples, thereby reducing the workload of data collection for on-site calibration.

[0024] 4. The error compensation method described in this invention reuses the standard checkerboard calibration board to automatically generate error labels, eliminating the need for manual labeling and additional precision acquisition equipment.

[0025] 5. The error compensation method described in this invention uses a modular pre-correction output standard point cloud, which is directly compatible with the ICP registration algorithm and does not require modification to the existing eye-on-hand robot 3D vision scanning process.

[0026] 6. The error compensation method described in this invention relies entirely on algorithms to achieve error compensation optimization, without the need to replace the high-precision robotic arm and calibration equipment, resulting in low hardware deployment and modification costs. Attached Figure Description

[0027] Figure 1This is a schematic diagram of the hardware layout of a typical robot 3D scanning system in a progressive error compensation method for robot visual point cloud registration as described in this invention.

[0028] Figure 2 A flowchart of a progressive error compensation method for robot vision point cloud registration according to the present invention;

[0029] Figure 3 This is a schematic diagram of the point cloud of the corner point of the calibration board after hand-eye calibration conversion when the compensation method of the present invention is not used; wherein, (a) is a schematic diagram of the top view and (b) is a schematic diagram of the side view.

[0030] Figure 4 This is a schematic diagram of the point cloud obtained after aligning the corner points of the calibration board using the progressive error compensation method described in this invention.

[0031] In the diagram: 1. Industrial robotic arm, 2. Industrial control computer, 3. Vision sensor, 4. Load-bearing turntable. Detailed Implementation

[0032] Specific Implementation Method 1: Combination Figures 1 to 4 This embodiment describes a progressive error compensation method for robot visual point cloud registration. This method is highly versatile and can be adapted to any hand-held robot 3D visual scanning system. The method includes an offline modeling module and an online module. The offline modeling module sequentially completes four steps: calibration data acquisition, hand-eye calibration iterative solution, translational linear model construction, and training of a multilayer perceptron (MLP) network. The online module performs the real-time error compensation step.

[0033] In this embodiment, Figure 1 The hardware described herein is merely a general-purpose supporting carrier required for the operation of this method, and represents one possible hardware layout example for implementing this method. It does not limit the applicable equipment models or whether a turntable is included. The industrial robotic arm 1 is a multi-degree-of-freedom motion carrier used to mount the vision sensor 3. The vision sensor 3 serves as an image acquisition unit, capturing images of the checkerboard calibration board and the object under test, and generating the original 3D point cloud based on parallax calculation. The carrying turntable 4 (optional supporting equipment): only equipped in some scanning systems, used to place the object under test, expand the scanning perspective, and enrich the spatial distribution of calibration samples. When there is no carrying turntable, the workpiece pose can be manually adjusted instead. The checkerboard calibration board (an essential reference object for this method): provides the true values ​​of corner points with known geometric coordinates, serving as the core reference for automatically generating model training pose labels. The industrial control computing host 2 serves as the method's operating carrier, completing point cloud preprocessing, translational linear solution, offline training of the rotational MLP model based on translational priors, real-time inference, and ICP point cloud registration calculation.

[0034] like Figure 1As shown, the industrial robotic arm 1 and the vision sensor 3 are mechanically fixed together. The vision sensor 3 is mounted on the end flange of the industrial robotic arm 1 and moves synchronously with the robotic arm. The industrial robotic arm 1 and the industrial control computing host 2 are connected by bidirectional electrical communication. The industrial control computing host 2 issues motion control commands, and the industrial robotic arm 1 uploads the six-dimensional pose data of the robot's end effector to the industrial control computing host 2. The vision sensor 3 transmits data unidirectionally to the industrial control computing host 2 and outputs the acquired image and the original three-dimensional point cloud. The vision sensor 3 faces the placement station above the support turntable 4 and is used to collect images and point clouds of objects on the station. In this embodiment, the support turntable 4 is optional hardware and is used to provide the placement station. During offline calibration, a checkerboard calibration plate is placed at the station, and during scanning operations, the station is replaced with the workpiece to be measured. When the system does not have the support turntable 4, the placement station can be achieved by the ground or a fixed bracket.

[0035] The error compensation method described in this embodiment is specifically implemented by the following steps:

[0036] Step S1: Collect the original calibration dataset and use it as the training dataset;

[0037] In this embodiment, the industrial robotic arm 1 equipped with vision sensor 3 is controlled to move along a preset trajectory. Using the chessboard calibration board as a reference, the industrial control computing host 2 synchronously collects multiple sets of six-dimensional poses of the robot end effector, the original point cloud output by the vision sensor, and the chessboard corner point images. The coordinates of the chessboard corner points are extracted to obtain the measured coordinates of the chessboard corner points in the vision sensor coordinate system under each set of sampled poses (measured coordinates of the lower corner points of the vision sensor). All six-dimensional poses of the robot end effector, measured coordinates of the lower corner points of the vision sensor, and coordinates of the chessboard corner points are saved to form the original calibration dataset.

[0038] Step S2: Solve for the hand-eye transformation matrix to generate a paired dataset with error labels;

[0039] Step S21: For the coordinates of multiple pairs of checkerboard corner points and the corresponding six-dimensional poses of the robot end effector in the original calibration dataset obtained in step S1, solve the hand-eye transformation matrix of the vision sensor 2 relative to the end flange of the robotic arm through hand-eye calibration iterative calculation.

[0040] Step S22: Substitute the hand-eye transformation matrix back into the original calibration dataset: Traverse all sampling records of the original calibration dataset. For each set of sampling records, use the hand-eye transformation matrix to transform and project the measured coordinates of the lower corner of the vision sensor to the robot base coordinate system to obtain the measured coordinates of the corner in the robot base coordinate system.

[0041] Step S23: After completing the coordinate transformation of all sampling records, select one set of measured corner coordinates of the visual sensor under the sampling pose as a unified reference from all sampling records of the original calibration dataset. Align the measured corner coordinates of all other sampling poses in the robot's base coordinate system with this reference. Perform SVD decomposition on the corner set after reference alignment to solve the rigid body transformation matrix. Perform analytical decomposition on the obtained rigid body transformation matrix to separate the translation error label and rotation error label, which serve as the training ground values. Combine these to form the six-dimensional error label corresponding to the sampling pose. This constructs a paired dataset with one-to-one correspondence between the robot's end effector six-dimensional pose, translation error label, and rotation error label.

[0042] In this embodiment, the translation error label serves as the fitting target for the linear regression in step S3, and the rotation error label serves as the supervised ground truth for the MLP network training in step S4. The hand-eye transformation matrix serves as an intermediate basic processing result, providing a foundation for the subsequent learning of translation compensation. Due to the inherent limitations of the linear hand-eye calibration model, the hand-eye transformation matrix cannot fully characterize the rotational nonlinear residual error present in the system, requiring further compensation processing.

[0043] Step S3: Solve the linear model based on the paired dataset from step S2 to construct a translational linear model. Output the translation compensation amount through the translational linear model to provide input information for the MLP network.

[0044] Using the robot's end-effector six-dimensional pose as input and the translation error labels in the paired dataset generated in step S2 as the fitting target, a linear mapping from the robot's end-effector pose to the translation compensation amount is established using least squares linear regression: ;

[0045] in: This is the amount of translation compensation; These represent the translation compensation components in the X, Y, and Z axes, respectively, under this translation compensation amount; Let be the linear regression coefficient matrix to be solved; This represents the six-dimensional pose vector of the robot's end effector. This is the bias term for linear regression.

[0046] Solving by closure Obtain translation compensation amount Where: Q is a sample matrix composed of multiple sets of six-dimensional pose samples of the robot's end effector; and... This serves as a priori input for subsequent MLP network modeling. The design is based on the following: translation errors follow an approximately linear, slowly varying distribution within the workspace; a linear closed-form solution can independently lock the translation reference, avoiding interference from rotational nonlinear errors; and the regression coefficient matrix is ​​obtained offline. The bias term The translational linear model is used as a parameter in the online phase.

[0047] Step S4: Train the MLP network using the robot's six-dimensional pose and translation compensation to learn the rotation error.

[0048] In this embodiment, a lightweight MLP network with nine inputs and three outputs is constructed. The preferred MLP network structure is a three-layer fully connected network, which includes two hidden layers and one output layer. The network input feature dimension is 9, the first hidden layer has 64 neurons, the second hidden layer has 32 neurons, and the output feature dimension of the output layer is 3. All layers are fully connected, the activation function is ReLU, and the loss function is the mean squared error of the rotation error.

[0049] In this embodiment, the translation compensation amount obtained in step S3 is compared with the six-dimensional pose of the robot's end effector. Concatenate to construct a 9-dimensional input feature vector ;

[0050] in: These are the six dimensional components of the robot's six-dimensional pose at the end effector; The 9-dimensional input feature vector of the MLP network is composed of a 6-dimensional pose input and a 3-dimensional translation compensation. The output rotation compensation of the MLP network is... ;in: , , These are the rotational compensation components around the X-axis, Y-axis, and Z-axis, respectively.

[0051] In this embodiment, the MLP network is trained using a stochastic gradient descent (SGD) optimizer; the initial learning rate is set to 0.001, and the number of training epochs is set to 200-5000; optionally, a learning rate decay strategy is adopted, with the learning rate decaying to 0.7 times the original value every 50 epochs; the training set and validation set are randomly partitioned into paired datasets in an 8:2 ratio, and an early stopping mechanism is adopted, whereby training is terminated early when the validation set loss no longer decreases for 20-300 consecutive epochs to prevent the MLP network from overfitting.

[0052] Step S5: Perform error compensation online in real time and output the compensated point cloud;

[0053] The hand-eye transformation matrix obtained in step S2, the translational linear model obtained in step S3, and the MLP network obtained in step S4 are used as the complete set of parameters for subsequent online scanning. Online scanning is achieved through two processes: basic spatial transformation and precise compensation for residual errors.

[0054] The basic spatial transformation involves using the offline hand-eye transformation matrix to transform the original point cloud in the visual sensor coordinate system to the robot base coordinate system, thus completing the initial spatial alignment of the point cloud.

[0055] The residual error is accurately compensated as follows: the six-dimensional pose of the robot end effector in real time is read, and the translation compensation amount is obtained in real time through the translation linear model; the six-dimensional pose of the robot end effector in real time is concatenated with the translation compensation amount and then input into the MLP network to obtain the rotation compensation amount; the translation compensation amount and the rotation compensation amount are merged into a complete six-dimensional registration matrix, which is superimposed on the point cloud applied to the initial spatial alignment to complete the error compensation, and the compensated point cloud is output.

[0056] In this embodiment, steps S1-S4 are all completed offline by the offline modeling module, requiring only a single run; step S5 is an online real-time process, replicating the core processing logic of the offline modeling module, sequentially calling the hand-eye transformation matrix, the translation linear model, and the MLP network to complete error compensation. The offline and online processes are structurally independent but have interconnected parameters, allowing for direct embedding into existing robot 3D vision scanning systems.

[0057] like Figure 3 and Figure 4 As shown, from Figure 3 In (b), obvious point cloud layering errors can be visually observed; Figure 4 The layering error in the calibration board is significantly suppressed. Without the error compensation method of this invention, the point cloud of the calibration board corner points after linear hand-eye calibration conversion has a layering deviation of about 40mm. After using the progressive error compensation processing of this invention, the layering error of the calibration board corner points in multiple frames is suppressed to within 3mm, the point cloud layering misalignment problem is greatly improved, and the compensated point cloud can fall into the effective convergence range of the ICP registration algorithm.

[0058] Specific Implementation Method Two: This implementation method is an embodiment of the progressive error compensation method for robot vision point cloud registration described in Specific Implementation Method One.

[0059] Optimal state 1:

[0060] In this embodiment, the control parameters are set as follows: the sampling space range of the calibration plate is consistent with the scanning space range of the workpiece, and the calibration plate is as close as possible to the area where the workpiece to be tested is located.

[0061] When collecting training samples, a checkerboard calibration plate is placed around the object to be tested to ensure that the pose space and working distance traversed by the robotic arm basically coincide with the subsequent formal scanning of the workpiece; the pose of the training set uniformly covers all joint spaces that will be used in the subsequent scanning.

[0062] The progressive error compensation method constructs a complete compensation system that learns the mapping relationship between the six-dimensional pose of the robot's end effector and the error in this space. When the training sampling space and the actual scanning space coincide, the fitting effect of the multi-source coupled nonlinear hierarchical error is optimal, and the compensation accuracy is the highest. Since the translation compensation amount is used as the conditional input of the MLP network in this method, if the spatial deviation between the two is too large, the prediction deviation of the translation reference will directly degrade the input condition quality of the MLP network, resulting in a significant decrease in the compensation effect.

[0063] Optimal state 2:

[0064] The control indicators set in this embodiment are: number of training samples and trajectory sampling density;

[0065] During the training and data collection phase, the robotic arm trajectory traverses the work area evenly to collect a sufficient number of spatially evenly distributed calibration pose samples (such as 30-60 sets of effective corner point samples), without the need to increase the number of samples indefinitely; at the same time, a smooth and continuous motion trajectory is adopted to avoid high-speed and sudden stops of the robotic arm.

[0066] Too few samples will prevent sufficient learning of the nonlinear error mapping; too many samples will prolong training time. Since our method uses a linear closed-form solution for translation (high sample efficiency) and the translation compensation amount as a conditional input significantly reduces the mapping space complexity of the rotation MLP, our method requires significantly fewer samples than a pure six-dimensional black-box model. This ensures that the hierarchical error compensation accuracy converges at a sampling density of 30-60 sets, while minimizing data acquisition and model training time, achieving an optimal balance between accuracy and operational efficiency.

[0067] Optimal state 3: Progressive update strategy;

[0068] The control parameters in this embodiment are set as follows: when the vision sensor, the end effector of the robotic arm is replaced, or the workspace is significantly changed, offline modeling (steps S1 to S4) must be re-executed.

[0069] During routine continuous scanning operations, the progressive model remains fixed; only when hardware is replaced or the workspace is significantly relocated, 30 to 60 sets of calibration data are re-acquired to quickly complete the solution of the linear translation matrix and retrain the rotation MLP network.

[0070] In this embodiment, the translation linear model and the MLP network jointly encode the error characteristics of the current hardware system. Changes in the translation reference will directly alter the conditional input distribution of the MLP network. If the model is not remodeled, the physical causal relationship will be disrupted, and the compensation accuracy cannot be guaranteed.

[0071] The error compensation method described in this invention solves the problem of existing technologies that solve translation and rotation simultaneously in the same model, failing to recognize the physical causal relationship that rotation error is a function of spatial translation position. This leads to translation results being contaminated by rotational nonlinear errors, or translation information being used merely as ordinary input features without exerting its constraint on rotational errors. This invention breaks the paradigm of simultaneous solution in mixed dimensions, constructing a cascaded sequential decoupled architecture where translation is solved independently first, and then the translation results are used as prior conditions input into the MLP network, allowing the MLP to only fit the rotational residual error. This fundamentally avoids the mutual interference between translation and rotation.

[0072] The error compensation method described in this invention solves the problems of existing technologies in the simultaneous translation and rotation solution paradigm, which, even with the introduction of neural networks to fit the six-dimensional error as a whole, still suffer from large model parameter counts, high training sample requirements, and long inference time, making it difficult to meet the requirements of real-time scanning in industrial settings. This invention transforms the translation component from a "fitting target" into a "prior condition," allowing the MLP network to fit only the three-dimensional rotation, significantly reducing both the number of network parameters and the inference time.

[0073] The error compensation method described in this invention abandons the high-cost optimization approach of relying on purchasing high-precision robots and high-end calibration equipment to suppress layering errors. It does not require the addition of precision hardware and only uses algorithms to achieve point cloud layering error compensation, resulting in lower implementation costs.

[0074] The error compensation method described in this invention solves the problem that point clouds still have significant initial layering deviations after linear model compensation, and that global ICP registration is prone to getting trapped in local optima and cannot be fully aligned. The point cloud compensated by this method can stably complete global 3D registration.

[0075] The error compensation method described in this invention does not force the assumption of global linear error propagation, and does not require the establishment of physical mechanism models for various types of errors. It autonomously learns the mapping relationship between robot joint pose and error compensation amount based on measured calibration data. The overall solution is lightweight, has strong real-time reasoning, and can adapt to various workpieces and various scanning conditions.

[0076] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0077] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A progressive error compensation method for robot visual point cloud registration, characterized by: The compensation method involves offline acquisition of the original calibration dataset, solving the hand-eye transformation matrix, and constructing an MLP network, followed by online scanning to achieve error compensation. This method is implemented through the following steps: Step 1: Collect the original calibration dataset; Step 2: Solve the hand-eye transformation matrix by performing hand-eye calibration iterative calculations on the calibration dataset described in Step 1, and generate a paired dataset with error labels; Step 3: Construct a translational linear model using the paired dataset described in Step 2; output the translational compensation amount through the translational linear model; Step 4: The translation compensation amount is spliced ​​with the six-dimensional pose of the robot's end effector as the input of the MLP network, and the rotation compensation amount is output through the MLP network. Step 5: Perform online real-time error compensation using the hand-eye transformation matrix described in Step 2, the translational linear model from Step 3, and the MLP network from Step 4.

2. The progressive error compensation method for robot vision point cloud registration according to claim 1, characterized in that: In step one, the industrial robotic arm equipped with a vision sensor is controlled to move along a set trajectory. Using a checkerboard calibration board as a reference, multiple sets of six-dimensional poses of the robot's end effector, raw point clouds output by the vision sensor, and checkerboard corner point images are collected simultaneously. The measured coordinates of the checkerboard corner points in the vision sensor coordinate system under each set of sampled poses are obtained, i.e., the measured coordinates of the lower corner points of the vision sensor. All six-dimensional poses of the robot's end effector, measured coordinates of the lower corner points of the vision sensor, and coordinates of the checkerboard corner points are saved to form the original calibration dataset.

3. The progressive error compensation method for robot vision point cloud registration according to claim 2, characterized in that: The specific implementation process of step two is as follows: Step 2: For the coordinates of multiple pairs of checkerboard corner points and the corresponding six-dimensional poses of the robot end effector in the original calibration dataset, solve the hand-eye transformation matrix of the vision sensor relative to the end flange of the robotic arm through hand-eye calibration iterative calculation. Step 22: Traverse all sampling records of the original calibration dataset. For each set of sampling records, use the hand-eye transformation matrix to transform and project the measured coordinates of the lower corner of the vision sensor to the robot base coordinate system to obtain the measured coordinates of the corner in the robot base coordinate system. Steps 2 and 3: Select one set of visual sensor corner coordinates under the sampling pose as a unified reference from all the sampling records of the original calibration dataset. Then, perform coordinate transformation and alignment on the corner coordinates of all other sampling poses in the robot base coordinate system to this reference. After the corner point set is aligned to the reference, singular value decomposition (SVD) is performed to solve the rigid body transformation matrix, and translation error labels and rotation error labels are obtained. A labeled error pairing dataset with one-to-one correspondence between the robot end effector's six-dimensional pose, translation error label, and rotation error label is constructed.

4. The progressive error compensation method for robot vision point cloud registration according to claim 3, characterized in that: In step three, the robot end-effector six-dimensional pose is used as input, and the translation error label in the paired dataset is used as the fitting target. Least squares linear regression is used to establish a linear mapping from the robot end-effector pose to the translation compensation amount, obtain the translation linear model, and output the translation compensation amount. ； In the formula, This is the amount of translation compensation; The translation compensation amount is respectively Translation compensation components in the X, Y, and Z axes; Let be the linear regression coefficient matrix to be solved; This represents the six-dimensional pose vector of the robot's end effector. This is the bias term for linear regression.

5. The progressive error compensation method for robot vision point cloud registration according to claim 4, characterized in that: In step four, the translation compensation amount is concatenated with the six-dimensional pose of the robot's end effector to construct a nine-dimensional input feature vector. ; in, These are the six-dimensional components of the robot's six-dimensional pose at the end effector; These represent the translation compensation components along the X, Y, and Z coordinate axes, respectively; the MLP network outputs the rotation compensation amount. ;in, , , These are rotational compensation amounts. Rotational compensation components around the X, Y, and Z axes.

6. A progressive error compensation method for robot vision point cloud registration according to any one of claims 1-5, characterized in that: In step five, error compensation is completed through basic spatial transformation and precise compensation of residual errors. The specific implementation process is as follows: Basic spatial transformation: Using a hand-eye transformation matrix, the original point cloud in the visual sensor coordinate system is transformed to the robot base coordinate system, completing the initial spatial alignment of the point cloud; Accurate compensation for residual errors: Read the six-dimensional pose of the robot end effector in real time, and obtain the translation compensation amount in real time through the translation linear model; The six-dimensional pose of the real-time robot end effector is concatenated with the translation compensation amount and then input into the MLP network to obtain the rotation compensation amount. The translation compensation and rotation compensation are combined into a complete six-dimensional registration matrix, which is then superimposed on the point cloud for initial spatial alignment to complete the error compensation.

Images