Robotic control method and system for adaptive optimization of hand-eye calibration

Abstract

The application relates to the technical field of robots, and discloses a robot control method and system for adaptive optimization of hand-eye calibration, wherein the method comprises the following steps: acquiring calibration instructions and work constraint data of a target robot; adopting an adaptive optimization path strategy based on online quality feedback and a preset calibration error compensation strategy, performing hand-eye calibration matrix calculation of the target robot according to the calibration instructions and the work constraint data, and obtaining a calibration result; acquiring a work instruction of the target robot; controlling the target robot to work according to the calibration result, and compensating the calibration result according to a preset work drift compensation strategy in the process of work of the target robot. Through the synergistic effect of adaptive optimization of the acquisition process, error compensation of the calibration result and drift compensation in the work stage, the robot can maintain high-precision and stable work capability under complex working conditions.

Description

Technical Field

[0001] This application relates to the field of robotics technology, and in particular to an adaptive optimization robot control method and system for hand-eye calibration. Background Technology

[0002] In the field of robotics, hand-eye calibration is a crucial step in achieving precise robot operations. Its purpose is to establish the coordinate transformation relationship between the vision sensor and the robot's end effector (or robot base coordinate system). Traditional hand-eye calibration methods typically rely on pre-set fixed paths for data acquisition, lacking real-time evaluation and feedback mechanisms for image quality during sampling. This makes them ill-suited to handle dynamic interference factors such as changes in ambient lighting, occlusion, and reflections, leading to unstable calibration accuracy. Furthermore, existing technologies often overlook hand-eye relationship drift caused by mechanical wear, temperature variations, and other factors during long-term robot operation, making it difficult to guarantee continuous accuracy and stability during operations. Summary of the Invention

[0003] Based on this, it is necessary to address the technical problems of unstable calibration accuracy caused by the fixed sampling path of the existing hand-eye calibration method, and the inability to guarantee the accuracy and stability of the robot's hand-eye relationship during long-term operation, which makes it difficult to guarantee the accuracy and stability of the operation. Therefore, an adaptive optimization robot control method and system for hand-eye calibration is proposed.

[0004] In a first aspect, an adaptive optimization robot control method for hand-eye calibration is provided, the method comprising:

[0005] Obtain the calibration instructions and operational constraint data of the target robot;

[0006] An adaptive optimization path strategy based on online quality feedback and a preset calibration error compensation strategy are adopted. The hand-eye calibration matrix of the target robot is calculated according to the calibration instructions and the operation constraint data to obtain the calibration result.

[0007] Obtain the working instructions of the target robot;

[0008] In response to the work instruction, the target robot is controlled to perform the operation according to the calibration result, and the calibration result is compensated according to the preset operation drift compensation strategy during the operation of the target robot.

[0009] The adaptive optimization path strategy based on online quality feedback, the calibration error compensation strategy, and the job drift compensation strategy work together to form a closed-loop control system across all stages.

[0010] In a second aspect, an adaptive optimization robot control system for hand-eye calibration is provided, the system comprising: a system controller and a target robot, the system controller being communicatively connected to the target robot, and the system controller being configured to implement the steps of the adaptive optimization robot control method for hand-eye calibration as described above.

[0011] Thirdly, a target robot is provided, the robot including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the adaptively optimized robot control method for hand-eye calibration as described above.

[0012] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the adaptive optimization robot control method for hand-eye calibration as described above.

[0013] The adaptive optimization robot control method and system for hand-eye calibration presented in this application have at least the following beneficial effects:

[0014] Compared to existing hand-eye calibration methods that use fixed sampling paths, lack real-time evaluation and feedback mechanisms for image quality during sampling, and ignore long-term drift, this application introduces a dual-sided collaborative mechanism combining an "adaptive optimization path strategy based on online quality feedback" and a "preset calibration error compensation strategy" for hand-eye calibration, achieving comprehensive optimization of the calibration process. The adaptive optimization path strategy based on online quality feedback allows for dynamic adjustment of the sampling path according to real-time image quality during data acquisition, effectively avoiding adverse factors such as illumination interference and occlusion, ensuring higher accuracy and consistency of the source data used in calibration calculations, thereby improving the accuracy of the hand-eye calibration matrix solution. Furthermore, by combining this with a preset calibration error compensation strategy, systematic errors in the calibration results are corrected, eliminating residual deviations introduced by algorithmic or environmental factors, thus obtaining more accurate and robust initial calibration results. Moreover, in the subsequent operation phase, this application introduces a job drift compensation strategy during the robot's execution of work instructions, providing real-time compensation for hand-eye relationship drift during operation, avoiding the problem of calibration accuracy decaying over time. Therefore, this application achieves high-precision and stable operation capability of the robot under complex working conditions through the synergistic effect of three measures: "adaptive optimization of the acquisition process", "error compensation of calibration results" and "drift compensation during the operation phase". Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] in:

[0017] Figure 1 This is an application environment diagram of an adaptive optimization robot control method for hand-eye calibration in one embodiment;

[0018] Figure 2 This is a flowchart of an adaptive optimization robot control method for hand-eye calibration in one embodiment;

[0019] Figure 3 This is a block diagram of an adaptively optimized robot control system for hand-eye calibration in one embodiment.

[0020] Main component description:

[0021] 1. Control terminal; 2. Target robot; 3. System controller. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Please see Figure 1 As shown, Figure 1 This diagram illustrates the application environment of an adaptive optimization robot control method for hand-eye calibration provided in an embodiment of this application. Figure 1 The application environment shown includes a robot control terminal 1 and a target robot 2. Robot control terminal 1 is communicatively connected to target robot 2, sending control commands to target robot 2 and receiving status data and visual acquisition data from target robot 2. In actual deployment, robot control terminal 1 can be integrated into the target robot 2, existing as its host controller or industrial control computer; alternatively, robot control terminal 1 can also be a physical entity independent of target robot 2 (i.e.,...). Figure 3 The system controller 3 in the application is, for example, a host computer, server or cloud control platform, and the embodiments of this application do not specifically limit this.

[0024] In the hand-eye calibration scenario applicable to the embodiments of this application, the target robot 2 can be a multi-axis robotic arm system such as an industrial robot, humanoid robot, or collaborative robot. To achieve hand-eye calibration, a target vision device (not shown in the figure) needs to be configured, and a calibration object (not shown in the figure) needs to be set. The target vision device can be an industrial camera or a vision sensor.

[0025] Depending on the hand-eye configuration, this application scenario can be further divided into two typical modes:

[0026] Mode 1 (Eye on Hand): The target vision device is mounted on the end effector of the robotic arm of the target robot 2 and can move with the movement of the robotic arm; the calibration object is fixedly set at a predetermined position outside the target robot 2, for example, a high-precision calibration plate fixed on a worktable. During the calibration process, the target robot 2 changes the pose of the end effector of the robotic arm by moving, so that the target vision device can acquire images of the fixed calibration object from different angles.

[0027] Mode 2 (Eye Outside Hand): The target vision device is fixedly installed at a predetermined position outside the target robot 2, for example, mounted on a bracket above the work area; the calibration object is installed at the end of the robotic arm of the target robot 2, for example, using a tool or special calibration fixture at the end of the robotic arm as the calibration object. During the calibration process, the target robot 2 changes the pose of the end of the robotic arm through movement, allowing the fixed target vision device to acquire images of the moving calibration object from different angles.

[0028] In any of the above modes, robot control terminal 1 is configured to: acquire calibration instructions and task constraint data of the target robot; employ an adaptive optimization path strategy based on online quality feedback and a preset calibration error compensation strategy to calculate the hand-eye calibration matrix of the target robot according to the calibration instructions and the task constraint data, thereby obtaining the calibration result; acquire the work instructions of the target robot; respond to the work instructions, control the target robot to perform tasks according to the calibration result, and compensate the calibration result according to the preset task drift compensation strategy during the task performance. By introducing a dual-sided collaborative mechanism combining the "adaptive optimization path strategy based on online quality feedback" and the "preset calibration error compensation strategy" for hand-eye calibration, comprehensive optimization of the calibration process is achieved. The adaptive optimization path strategy based on online quality feedback enables dynamic adjustment of the sampling path according to real-time image quality during data acquisition, thereby effectively avoiding adverse factors such as illumination interference and occlusion, ensuring higher accuracy and consistency of the source data participating in the calibration calculation, and thus improving the accuracy of the hand-eye calibration matrix solution. Building upon this foundation, a pre-defined calibration error compensation strategy is further incorporated to systematically correct the calibration results, eliminating residual biases introduced by algorithmic or environmental factors, thereby obtaining more accurate and robust initial calibration results. Furthermore, in subsequent operation phases, this application introduces a job drift compensation strategy during the robot's execution of work instructions, providing real-time compensation for hand-eye relationship drift during operation, thus avoiding the problem of calibration accuracy decaying over time. Therefore, through the synergistic effect of three measures—adaptive optimization of the acquisition process, error compensation of calibration results, and drift compensation during the operation phase—this application enables the robot to maintain high-precision and stable operational capabilities under complex working conditions.

[0029] Furthermore, the technical solution described in this application possesses excellent system integration and deployment convenience. The adaptive optimization path strategy based on online quality feedback, the preset calibration error compensation strategy, and the preset job drift compensation strategy can be encapsulated as independent functional modules and seamlessly integrated with existing robot control systems and Manufacturing Execution Systems (MES) through standard communication interfaces. Enterprises can deploy the technical solution of this application on existing equipment without large-scale modifications to existing production lines, reducing technology upgrade costs. Simultaneously, the data interaction capability between the method described in this application and the production management system allows calibration triggering conditions (such as job accuracy deviations) to directly originate from process parameters issued by the MES. Calibration results and compensation data can also be written back into the MES, forming a closed-loop data flow from production instructions to equipment execution, providing key technical support for building an intelligent, self-optimizing manufacturing system.

[0030] This application is particularly suitable for industrial automation scenarios with high precision, high intensity, and complex environments, such as automotive welding production lines, precision assembly of 3C (computers, communication products, consumer electronics and related industries), and aerospace component manufacturing, which have stringent requirements for operational precision and long-term operational stability.

[0031] Please see Figure 2 As shown, Figure 2 A flowchart illustrating an adaptive optimization robot control method for hand-eye calibration provided in this application embodiment includes the following steps:

[0032] S1: Obtain the calibration instructions and operational constraint data of the target robot;

[0033] S2: Using an adaptive optimization path strategy based on online quality feedback and a preset calibration error compensation strategy, the hand-eye calibration matrix of the target robot is calculated according to the calibration instructions and the operation constraint data to obtain the calibration result;

[0034] S3: Obtain the working instructions of the target robot;

[0035] S4: In response to the work instruction, control the target robot to perform the operation according to the calibration result, and compensate the calibration result according to the preset operation drift compensation strategy during the operation of the target robot;

[0036] The adaptive optimization path strategy based on online quality feedback, the calibration error compensation strategy, and the job drift compensation strategy work together to form a closed-loop control system across all stages.

[0037] The target robot is the robot that the adaptive optimization robot control method for hand-eye calibration proposed in this application needs to control.

[0038] Job constraint data refers to the physical limitations and safety boundaries that a robot must meet during calibration or operation. Examples include the range of motion of each joint, the boundaries of the workspace, the minimum safe distance from surrounding fixed equipment or obstacles, and speed limits in specific postures. Sources of job constraint data include pre-set kinematic parameters, virtual walls set by offline programming software, and obstacle location information acquired in real time through an environmental perception system.

[0039] Calibration command: This refers to the command signal that initiates and guides the hand-eye calibration process. It typically includes parameters such as the identification information of the calibration object, the selection of the calibration mode, and the desired calibration accuracy level. For example, the user-inputted "Start hand-eye calibration" command, or the command actively triggered by the method described in this application.

[0040] Step S1 specifically includes: This application triggers the calibration process based on preset calibration start conditions, which include one or more of the following: a calibration signal input by the user, a calibration signal triggered by the adjustment of the target vision device, a calibration signal triggered by the adjustment of the target robot, the difference between the current calibration result's operating temperature and the calibration temperature being greater than a preset temperature value, the current calibration result's operating accuracy being greater than a preset accuracy, and an active prevention start condition; After triggering calibration, the system obtains a calibration instruction, which includes information such as the identifier of the calibration object and the expected calibration accuracy, and simultaneously obtains operating constraint data from the target robot's controller or production management system.

[0041] An adaptive optimization path strategy based on online quality feedback refers to a control mechanism that dynamically adjusts subsequent sampling paths during the data acquisition process of hand-eye calibration based on the quality assessment results of real-time acquired images. For example, in an eye-on-hand configuration, when the robot moves to a preset sampling point to photograph the calibration object, if the image is blurred due to environmental reflection, the robot predicts and plans the next observation pose that will yield a better image in real time based on the current image quality score, the high-quality data already acquired, and the task constraint data. The sampling path is then adjusted based on the determined new observation pose.

[0042] Preset calibration error compensation strategy: The pre-set calibration error compensation strategy refers to a method of systematically correcting the error of the currently calculated hand-eye calibration matrix based on historical calibration results. This strategy analyzes the calibration results of the target robot within a preset first time window, uses a pre-trained prediction model to identify the drift trend of hand-eye relationship with factors such as cumulative runtime or temperature, and then generates a compensation matrix to correct the current calibration matrix.

[0043] The hand-eye calibration matrix is ​​a mathematical expression describing the spatial transformation relationship between the coordinate system of the target vision device and the coordinate system of the robot's end effector (or robot base coordinate system), typically presented as a 4×4 homogeneous transformation matrix. This matrix transforms the pose detected by the target vision device into a coordinate system of motion commands that the robot can execute, and is a core parameter for vision-guided robot precision operation. The data source for the hand-eye calibration matrix is ​​the optimal solution obtained by calculating multiple sets of robot poses and corresponding image feature points using a hand-eye calibration algorithm.

[0044] In the eye-on-hand configuration, the hand-eye calibration matrix describes the spatial transformation relationship between the coordinate system of the target vision device fixed at the end of the robot arm and the coordinate system of the robot's end effector. Since the target vision device moves with the robot arm, this matrix converts the pose of the calibrated object detected by the target vision device in the camera coordinate system into the pose in the robot's end effector coordinate system. Then, through the robot's forward kinematics, the pose of the calibrated object in the robot's base coordinate system is calculated, providing a basis for the robot arm's grasping or manipulation. In this configuration, the calibrated object is fixed in the workspace outside the robot. Solving the hand-eye calibration matrix essentially establishes a constant transformation relationship between the camera coordinate system of the target vision device and the end effector coordinate system. Its mathematical form is usually expressed as the classic hand-eye calibration equation AX = XB, where A is the pose transformation matrix of the robot's end effector between different sampling points, B is the pose transformation matrix of the calibrated object in the camera coordinate system between different sampling points, and X is the hand-eye calibration matrix to be solved.

[0045] In an eye-outside-hand configuration, the hand-eye calibration matrix describes the spatial transformation relationship between the coordinate system of the target vision device, which is fixedly mounted in the robot's external workspace, and the robot's base coordinate system. Since the target vision device's position remains constant, this matrix directly converts the pose of the calibration object detected by the target vision device in the camera coordinate system into the pose in the robot's base coordinate system, thereby guiding the robot's end effector to the target position. In this configuration, the calibration object is mounted on the robot's end effector and moves with it. Solving the hand-eye calibration matrix essentially establishes a constant transformation relationship between the fixed camera coordinate system and the robot's base coordinate system. Its mathematical form is usually an extended form of AX=ZB, where A is the robot end effector pose transformation matrix, B is the pose transformation matrix of the calibration object observed by the camera, X is the transformation relationship between the end effector and the calibration object to be solved (usually known), and Z is the hand-eye calibration matrix to be solved, representing the transformation relationship between the fixed camera and the robot base.

[0046] Step S2 specifically includes: first, generating an initial sampling path sequence and controlling the target robot to move along the sequence; upon reaching each sampling point, controlling the target vision device to photograph the calibration object and performing online quality evaluation on the photographed results; if the evaluation is qualified, recording the current data pair and continuing along the original path; if the evaluation is unqualified, dynamically updating the subsequent sampling path based on the current quality evaluation result, the collected data, and the task constraint data to guide the robot to move to a better observation pose, repeating the process until all preset or dynamically generated sampling points have been traversed; finally, calculating the hand-eye calibration matrix based on all qualified data, and combining it with a preset calibration error compensation strategy, using historical calibration results to systematically compensate for the error of the matrix calculated this time, obtaining the final calibration result. Thus, by introducing an adaptive optimization path strategy based on online quality feedback, the sampling path can be dynamically adjusted according to the real-time image quality during data acquisition, effectively avoiding adverse factors such as illumination interference and occlusion, making the source data participating in the calibration calculation more accurate and consistent, and improving the accuracy of the hand-eye calibration matrix calculation. Based on this, a pre-set calibration error compensation strategy is further combined to systematically correct the current calibration results using historical calibration results, eliminating residual biases introduced by algorithm or environmental factors, thereby obtaining more accurate and robust initial calibration results.

[0047] Work instructions are operational commands that direct robots to perform specific production tasks. They include information such as the target object, action type, motion path, and process parameters. For example, in an assembly scenario, a work instruction could be described as "grab the gearbox housing on the conveyor belt and place it on the positioning pin at the assembly station." The data sources for work instructions mainly include task queues issued by the upper-level production management system (MES / ERP), trajectory points programmed manually, and gripping points generated by real-time identification of workpiece pose transformation.

[0048] In the application environment of this application, the calibration object (e.g., a high-precision calibration board) and the target object that the target robot needs to operate in actual operation (e.g., a workpiece to be grasped) can coexist in the same physical space. The calibration object serves as a reference point in the hand-eye calibration process, used to calculate the coordinate transformation relationship between the target vision device and the target robot; the target object is the actual object to be operated during the operation phase.

[0049] In step S2, only the calibration object is considered. The adaptive optimization path strategy based on online quality feedback generates a sampling path sequence according to the pose of the calibration object, and drives the dynamic adjustment of the path by performing quality assessment on the image containing the calibration object. The presence of the target object does not affect the core logic of the calibration algorithm, but during the dynamic path correction process, the safe distance parameter with surrounding objects included in the job constraint data will use the target object or its container as an obstacle avoidance constraint to ensure the safety of the corrected path. After the calibration process is completed, the system enters the job stage (step S4), at which point the calibration object can be removed or no longer used, and the method of this application then performs jobs on the target object based on the calibration results.

[0050] For step S3, the specific steps include: by real-time monitoring or active querying of the work task queue from the upper control system or manual input, when a new work task is detected, the corresponding work instruction is obtained. The instruction contains the pose information of the target object, the work type and process parameters, and the instruction is parsed into a work sequence that the robot can execute.

[0051] A pre-defined job drift compensation strategy refers to a method that dynamically compensates for real-time hand-eye relationship drift caused by factors such as mechanical wear and temperature changes during the execution of work instructions by the target robot. For example, after the robot has been working continuously for several hours, the corresponding correction amount is obtained from a pre-defined compensation rule or compensation matrix based on the real-time monitored joint temperature or cumulative load.

[0052] In practical industrial applications, during long-term continuous operation, mechanical wear (such as increased backlash in reducers and joint loosening) and changes in ambient temperature (such as thermal expansion and contraction of the robotic arm due to day-night temperature differences in the workshop) can cause slow but continuous systematic drift in the hand-eye calibration matrix. This drift is trend-driven and cumulative, and cannot be eliminated by a single calibration. Without an effective monitoring and compensation mechanism, the accuracy of visual guidance will gradually decrease, ultimately affecting the quality of work. To solve this problem, step S4 specifically includes: monitoring the work data in real time when the target robot responds to work instructions and performs work tasks. The work data includes the operating status and environmental parameters; generating a drift compensation amount to correct the current hand-eye relationship based on the real-time acquired work data according to a preset work drift compensation strategy; using this compensation amount to dynamically correct the initial calibration result generated in step S2 in real time, and controlling the target robot to complete subsequent fine operations such as grasping or assembly based on the corrected hand-eye calibration matrix. By introducing a job drift compensation strategy during the execution of work instructions by the target robot, real-time compensation is provided for hand-eye relationship drift during operation, avoiding the problem of calibration accuracy decaying over time, thereby improving the long-term operational reliability and job quality of the target robot.

[0053] The adaptive optimization path strategy based on online quality feedback, the calibration error compensation strategy, and the job drift compensation strategy work together to form a closed-loop control system across all stages. Specifically, the adaptive optimization path strategy based on online quality feedback dynamically adjusts the sampling path according to image quality during the calibration acquisition stage to ensure the quality of source data; the calibration error compensation strategy systematically corrects the solution results using historical drift trends during the calibration calculation stage; and the job drift compensation strategy dynamically compensates for long-term drift based on real-time operating conditions during the job execution stage. Through a positive feedback mechanism of "prediction-guided acquisition, acquisition-supported compensation, and compensation-feedback-prediction," these three strategies form a closed loop from calibration to job, achieving a synergistic effect of "1+1+1>3."

[0054] This embodiment achieves comprehensive optimization of the calibration process by introducing a dual-sided collaborative mechanism combining an "adaptive optimization path strategy based on online quality feedback" and a "preset calibration error compensation strategy" for hand-eye calibration. The adaptive optimization path strategy based on online quality feedback dynamically adjusts the sampling path according to real-time image quality during data acquisition, effectively avoiding adverse factors such as illumination interference and occlusion, ensuring higher accuracy and consistency of the source data used in the calibration calculation, thereby improving the accuracy of the hand-eye calibration matrix solution. Furthermore, the preset calibration error compensation strategy systematically corrects the calibration results, eliminating residual deviations introduced by algorithmic or environmental factors, thus obtaining more accurate and robust initial calibration results. Moreover, in the subsequent operation phase, this application introduces a job drift compensation strategy during the robot's execution of work instructions to compensate for hand-eye relationship drift in real time, avoiding the problem of calibration accuracy decaying over time. Therefore, this application achieves high-precision and stable operation capability of the robot under complex working conditions through the synergistic effect of three measures: "adaptive optimization of the acquisition process", "error compensation of calibration results" and "drift compensation during the operation phase".

[0055] In one embodiment, the step of employing an adaptive optimization path strategy based on online quality feedback and a preset calibration error compensation strategy, and calculating the hand-eye calibration matrix of the target robot according to the calibration instructions and the task constraint data to obtain the calibration result includes:

[0056] S21: Obtain the calibration pose of the calibration object according to the calibration instruction, and generate a sampling path sequence according to the calibration pose;

[0057] S22: Control the target robot to move along the sampling path sequence;

[0058] S23: When the target robot moves along the sampling path sequence to reach the sampling point, control the target vision device to take a picture of the calibrated object, obtain the shooting result and the end pose of the target robot at a single sampling point, perform quality evaluation based on the shooting result, and obtain the evaluation result;

[0059] S24: If the evaluation result is qualified, the shooting result and the single sampling point end pose with the evaluation result being qualified are taken as qualified data pairs, and the process jumps to the step of controlling the target robot to move along the sampling path sequence to continue execution until there are no sampling points in the sampling path sequence that the target robot has not reached.

[0060] S25: If the evaluation result is unqualified, then based on the job constraint data, the calibration pose, the current evaluation result, the current shooting result, and the current single sampling point end pose, the sampling path sequence is updated to move the target robot to a better observation pose, and the process jumps to the step of controlling the target robot to move along the sampling path sequence to continue execution until there are no sampling points in the sampling path sequence that the target robot has not reached;

[0061] S26: Based on the calibration error compensation strategy, calculate the hand-eye calibration matrix of the target robot according to the calibration pose and each qualified data pair to obtain the calibration result;

[0062] Wherein, the target vision device is located at the end of the robotic arm of the target robot and the calibration object is located at a predetermined position outside the target robot, or the target vision device is located at a predetermined position outside the target robot and the calibration object is located at the end of the robotic arm of the target robot.

[0063] Calibration pose refers to the spatial position and orientation information of the calibration object in the robot's workspace, specifically including three-dimensional coordinates (X, Y, Z) and rotation angles (RX, RY, RZ) around each coordinate axis. For example, in the "eye in hand" configuration, the calibration object is a high-precision calibration board fixed on the worktable. Its pose can be calculated by the target vision device after the robot moves to an initial position where the calibration board can be clearly observed. The calibration pose is obtained from the initial observation pose after obtaining the calibration command in step S1, by controlling the robot to move to the initial observation pose, and by taking a picture of the calibration object and calculating it in real time using the known geometric features of the calibration board. This serves as the spatial reference for subsequently generating the sampling path sequence.

[0064] Sampling path sequence: This refers to a pre-defined trajectory consisting of a series of target pose points that the robot end effector (or robotic arm end effector) must traverse to collect the data required for hand-eye calibration. This sequence is generated based on the calibration pose and aims to provide full coverage observation of the calibration object from different distances and angles. For example, when the calibration board is located on the left side of the workbench at a suitable distance, step S21 will automatically plan a path that includes translational scanning and rotational posture changes based on this pose to ensure that the calibration board is always within the camera field of view of the target vision device.

[0065] Single-sampling-point end-effector pose: This refers to the precise spatial position and orientation of the robotic arm's end effector in the robot's base coordinate system at the current moment, recorded by the target robot's controller when the target robot moves to a certain sampling point during the process of the target vision device capturing the calibrated object. For example, in step S23, when the target robot reaches the sampling point and triggers the capture, the current joint angle values ​​are simultaneously read from the target robot's controller, and the six-degree-of-freedom pose data of the end effector is obtained through forward kinematics calculation. The single-sampling-point end-effector pose, as a key component of the qualified data pair, corresponds one-to-one with the capture results and is a necessary input parameter for the subsequent step S26 to solve the hand-eye calibration matrix equation (such as AX=XB).

[0066] Image capture results refer to the raw image data generated after the target vision device captures images of the calibration object at each sampling point. For example, in "eye on hand" mode, an industrial camera fixed to the end of a robotic arm captures an image of the calibration object located on a worktable, resulting in a high-resolution digital image containing the calibration object. The quality of the image capture results directly affects the accuracy of subsequent feature extraction and the reliability of the calibration results, and is the basis for quality assessment and hand-eye matrix calculation.

[0067] Quality assessment refers to the quantitative evaluation process of automatically analyzing the shooting results acquired at each sampling point using image processing algorithms to determine whether they meet the calibration requirements. This assessment typically includes multiple dimensions, such as sharpness detection to determine if motion blur or focus issues exist, illumination uniformity analysis to detect overexposure or reflective areas, and extracting features of the calibration object and calculating its contour integrity to determine if occlusion exists. The assessment results (such as a comprehensive score or a Boolean value indicating whether each indicator meets the standards) will serve as the basis for decisions in steps S24 and S25 regarding whether to continue along the original path or trigger dynamic path correction.

[0068] Evaluation Results: Used to characterize the usability of the current sampling point's image capture results. Evaluation results include a set of index values ​​and a conclusion. The conclusion of the evaluation result can be a binary value (pass / fail) or a multi-dimensional scoring vector. For example, if the image of the calibration object is severely glared due to excessive ambient light, resulting in low confidence in feature point extraction, the evaluation result is "fail." This result directly drives the logical branches of steps S24 and S25. When the evaluation result is passable, it is retained as the trigger data for step S25 and the original path continues; when the evaluation result is failable, it serves as one of the core inputs for triggering the path update in step S25, indicating the need to find a better observation pose.

[0069] A qualified data pair refers to a data unit consisting of a captured image that has been deemed "qualified" after quality assessment and its corresponding end-effector pose at a single sampling point. For example, in a successful acquisition, the robot's end-effector pose P1 and a clear, complete calibration board image I1 captured in that pose together constitute a qualified data pair (P1, I1). In step S24, when the assessment result is qualified, the program file implemented in this application stores these two parts of the data at that sampling point together. Multiple such qualified data pairs constitute the input dataset used in step S26 for the final calculation of the hand-eye calibration matrix, ensuring that the source data involved in the calculation has high accuracy and consistency.

[0070] Step S21 specifically includes: After receiving the calibration command, the method described in this application first performs an initial observation of the calibration object through the target vision device to obtain its initial pose in the robot's workspace as the calibration pose. Subsequently, the method described in this application uses the calibration pose as a reference center and combines the geometric dimensions of the calibration object with preset sampling density parameters to automatically generate a sampling path sequence covering multiple distances, angles, and spatial regions within the robot's reachable space. This sequence contains multiple ordered sampling points, each corresponding to a target pose of a robot end effector, which is used to guide the robot to move sequentially and acquire images as the shooting result.

[0071] The sampling path sequence is generated using a sampling generation algorithm based on spatial coverage and viewpoint quality pre-assessment. This algorithm uses the calibration pose as a reference center and, combined with the geometric dimensions of the calibration object, the field of view of the target vision device, and the working distance, constructs a set of candidate observation poses within the reachable space of the target robot. Specifically, the algorithm first uniformly samples distance and angle parameters in the normal direction of the calibration object and multiple deflection angle directions, forming an initial viewpoint set covering different distances and azimuth angles. Then, based on operational constraint data, viewpoints exceeding the robot's workspace boundary or colliding with surrounding equipment are eliminated. Finally, the viewpoint set is sorted according to a pre-set image quality pre-assessment model, prioritizing viewpoints with high expected feature clarity and low occlusion risk, and organizing them into an ordered sampling path sequence based on spatial proximity, ensuring efficient and comprehensive sampling. This algorithm balances the integrity of sampling coverage with the safety of path execution.

[0072] Understandably, taking pictures without planning a path often relies on the operator's experience or random movement, which can easily lead to missing key observation angles or repeated sampling in different areas, resulting in unreasonable data distribution. Step S21 generates an initial sampling path sequence based on the calibration pose, systematically covering spatial areas around the calibration object at different distances and angles, ensuring that the data participating in the calibration calculation have good geometric constraints, thereby improving calibration accuracy; at the same time, it avoids redundant sampling and improves calibration efficiency.

[0073] It is understood that the spatial range, center point, and distance parameters of the sampling path sequence generated in step S21 are automatically adjusted based on the calibration pose obtained in step S1. For example, when the placement position, distance, or geometric dimensions of the calibration object are different, the range, center point, and distance parameters of the generated initial scanning path are dynamically adapted to ensure that the sampling path sequence can effectively cover the spatial area where the calibration object is located, providing a good spatial reference for subsequent data acquisition.

[0074] Step S22 specifically includes: The method described in this application sends the sampling path sequence to the controller of the target robot one by one in the form of motion commands, driving the target robot to move sequentially to each sampling point in the sequence according to a preset order. During the movement, the method described in this application monitors the state and position of the target robot in real time to ensure that it accurately reaches the target pose, and pauses at each sampling point to wait for image acquisition to be triggered.

[0075] Step S23 specifically includes: after the target robot moves to the current sampling point and stabilizes, the method described in this application triggers the target vision device to perform an image acquisition on the calibration object to obtain the shooting result. At the same time, the current pose of the target robot's end effector is read from the target robot's controller as the end effector pose of the single sampling point. Subsequently, the method described in this application performs an online quality assessment of the shooting result, extracting indicators such as feature integrity, sharpness, contrast, and whether there is occlusion or reflection on the calibration board, and comprehensively generating the quality assessment result of the current sampling point.

[0076] Optionally, online quality assessment of the shooting results can be performed, specifically through multi-dimensional feature quantization and fusion. First, feature points (such as the center of a circle or corner points of a checkerboard) are extracted from the image of the calibrated object (i.e., the shooting result). The feature integrity index is calculated as the ratio of the actual number of detected feature points to the theoretical total number of feature points. Second, the sharpness index is evaluated using the image gradient energy function (such as the Tenengrad operator) or Laplacian variance, reflecting the sharpness of the edges of the calibrated object; low scores correspond to blurry images. Third, the contrast index is measured by the local contrast of the feature point region (such as the grayscale standard deviation); low contrast affects the sub-pixel accuracy of feature point localization. For occlusion and reflection, the spatial continuity of the feature point distribution is analyzed: if feature point loss exhibits regional clustering, it is determined to be occlusion; if the grayscale saturation of the feature point region is too high and the gradient is abnormal, it is determined to be reflection. After normalization, the above indicators are weighted and fused to generate a comprehensive quality score. This comprehensive quality score is then compared with a preset quality threshold. If the score is lower than the preset threshold, the evaluation result is deemed unqualified; otherwise, it is deemed qualified. The evaluation values ​​of the above indicators are then added to the indicator value set of the evaluation results.

[0077] The comprehensive judgment of the quality assessment adopts a weighted threshold fusion rule, and the specific steps are as follows:

[0078] (1) Quantification of indicators in each dimension:

[0079] Feature completeness index C_complete: The ratio of the actual number of feature points detected based on the shooting results to the total number of theoretical feature points.

[0080] Clarity index C_clarity: The mean value of the gradient magnitude in the captured image is calculated using the Tenengrad gradient operator and normalized to the [0,1] interval.

[0081] Contrast index C_contrast: The mean of the local grayscale standard deviation of the feature point region in the shooting result, normalized to the [0,1] interval.

[0082] Occlusion / Reflection Index C_occlusion: The spatial continuity analysis result of the missing feature points in the shooting results. If the missing area shows regional clustering and the area ratio exceeds the preset threshold (such as 10%), it is judged as occlusion or reflection, and the index is assigned a value of 0; otherwise, it is 1.

[0083] (2) Weighted fusion calculation:

[0084] Calculate the overall quality score: Q_total = α·C_complete + β·C_clarity + γ·C_contrast + δ·C_occlusion, where α, β, γ, and δ are preset weight coefficients that satisfy α+β+γ+δ=1, and the selectable values ​​are α=0.4, β=0.3, γ=0.2, and δ=0.1, respectively.

[0085] (3) Rules for determining qualification:

[0086] The evaluation result is deemed unqualified if any of the following conditions are met: the overall quality score Q_total is lower than a preset quality threshold (e.g., 0.75); the feature integrity index C_complete is lower than a preset integrity lower limit (e.g., 0.8); or the occlusion / reflection index C_occlusion is 0.

[0087] Otherwise, the conclusion of the evaluation result shall be deemed as qualified.

[0088] Step S24 specifically includes: when the evaluation result is qualified, combining the current shooting result with the corresponding single sampling point end pose to form a qualified data pair and storing it in a temporary data buffer. Subsequently, continue to execute step S22, control the robot to move to the next sampling point in the sampling path sequence, and repeat the acquisition and evaluation process until all sampling points in the sampling path sequence have been traversed.

[0089] Step S25 specifically includes: when the evaluation result is unqualified, the current data pair is not recorded. Based on the quality defect-driven Next Best View (NBV) planning algorithm, according to the unfavorable factors reflected in the current evaluation result (such as reflection, occlusion), the spatial distribution of the collected qualified data, the calibration pose, and the operation constraint data, the optimal observation planning algorithm predicts in real time an observation pose that can obtain a higher quality image, and inserts this pose as a new sampling point into the sampling path sequence. Subsequently, the system continues to execute step S22, controlling the robot to move to the next sampling point in the updated sequence.

[0090] Specifically, based on the current set of evaluation results, the specific types of quality defects (such as glare, occlusion, blur, and low contrast) are identified, and the areas where defects occur in the captured images are located. For glare, a target viewing angle that deviates from the specular reflection angle is calculated based on the geometric relationship between the surface normal of the calibrated object and the current light source direction. For occlusion, the approximate spatial range of the occluder in the robot's workspace is estimated based on the projected area of ​​the occluder in the captured image, combined with the known intrinsic parameters of the target vision device and the robot pose at the current sampling point, and an observation direction that can bypass the occluder is planned. For blur or low contrast, the working distance is adjusted to the ideal depth of field range of the calibration board. Centered on the calibration pose, within a conical region near the avoidance direction, multiple candidate observation poses are generated, combining the field of view of the target vision device and the working distance constraints. Taking the current single-sampling point end pose as the starting point of motion, check in turn whether each candidate pose meets the operation constraint data (such as joint limits, whether there is a collision risk in the path from the current pose to the candidate pose), and score it according to the "expected image quality improvement" and "path cost required to move from the current single-sampling point end pose to the candidate pose", and select the best one as the updated observation pose.

[0091] During the update of the sampled path sequence in step S25, the task constraint data serves as boundary conditions, limiting the feasible space for the better observation pose. When other objects (such as baskets or supports) exist in the scene, the safe distance parameter with surrounding objects included in the task constraint data will participate in the selection of candidate observation poses, ensuring that the updated path avoids quality defects such as reflection and occlusion while preventing collisions with other objects in the scene. Therefore, the adaptive path optimization strategy can dynamically adapt to the specific physical environment of the calibration object.

[0092] Candidate observation pose generation: A spherical coordinate system is constructed centered on the calibration pose. Multiple candidate observation directions are uniformly sampled on the sphere, with each candidate direction corresponding to a candidate observation pose. The position of the candidate observation pose is determined by the radius of the sphere and the direction angle. The radius of the sphere is constrained by the working distance range of the target vision device, and the direction angle is discretely sampled within the range of azimuth 0°~360° and pitch -60°~+60° with preset step sizes (e.g., azimuth step size 30°, pitch step size 15°). For each candidate observation pose, inverse kinematics is used to check whether it meets the operational constraint data (e.g., joint limits, accessibility), and unreachable poses are eliminated.

[0093] Based on the set of index values ​​from the evaluation results, the quality defect type of the current shooting result is identified: if the defect type is reflection, the specular reflection direction is calculated based on the geometric relationship between the surface normal of the calibration object and the current light source direction, and the angle between the observation direction of the candidate observation pose and the specular reflection direction is constrained to be greater than a preset reflection avoidance angle (e.g., 30°); if the defect type is occlusion, the approximate spatial range of the occluder in the robot's workspace is estimated by back projection based on the spatial distribution of the feature point missing area in the shooting result, combined with the camera intrinsic parameters of the target vision device and the current single sampling point end pose, and the angle between the observation direction of the candidate observation pose and the pointing direction of the occluder is constrained to be less than a preset occlusion avoidance threshold, so that the calibration object avoids the occluded area; if the defect type is blur or low contrast, the distance between the position of the candidate observation pose and the calibration pose is constrained to fall within the ideal depth of field range of the target vision device, and the angle between the observation direction and the surface normal of the calibration object is constrained to be less than a preset maximum tilt angle (e.g., 45°).

[0094] A weighted multi-objective optimization function is constructed to score each candidate observation pose: Score = w1·Q_pred + w2·(1 - C_norm), where Q_pred is the expected image quality score, pre-evaluated based on factors such as the distance between the candidate observation pose and the calibration pose, the observation angle, and the estimated illumination; C_norm is the normalized value of the path cost from the current single-sampling point's end pose to the candidate observation pose (path cost is calculated based on the joint space Euclidean distance or motion time); w1 and w2 are preset weight coefficients (e.g., w1=0.7, w2=0.3). The candidate observation pose with the highest Score is selected as the superior observation pose and inserted into the sampling path sequence, located at the next position after the current sampling point.

[0095] It is understandable that the "adaptive optimization path strategy based on online quality feedback" does not adopt an open-loop mode of "collecting all data first and then planning the path uniformly," nor is it a completely fixed "planning first and then executing" mode. Instead, it constructs a closed-loop control mode of "planning-execution-evaluation-correction." Specifically, the sampling path sequence generated in step S21 based on the calibration pose aims to ensure the basic coverage and efficiency of sampling. Subsequently, during the execution of steps S23 to S25, a dynamic decision is made on whether to continue using the original path or make corrections through real-time image quality evaluation. When the evaluation result is unqualified, the NBV planning algorithm driven by quality defects predicts a better observation pose and updates the path in real time based on the collected data. This closed-loop mode allows the systematic nature of the initial planning to be balanced with the flexibility of adaptive adjustment, thereby improving calibration efficiency while effectively avoiding the impact of dynamic adverse factors such as ambient light interference and occlusion on the final calibration accuracy.

[0096] Step S26 specifically includes: after traversing all sampling points, summarizing all qualified data pairs, combining the calibration pose, and using a hand-eye calibration algorithm to calculate the hand-eye calibration matrix as the first matrix. Subsequently, according to the calibration error compensation strategy, historical calibration data within a preset first time window is obtained from the historical calibration dataset, input into the pre-trained prediction model for drift trend prediction, generating a calibration compensation matrix, and using the calibration compensation matrix to compensate the first matrix to obtain the final calibration result of this calibration.

[0097] This embodiment first generates an initial sampling path based on the actual pose of the calibration board in step S21, ensuring basic coverage and systematicity of the sampling and avoiding blind planning. Subsequently, through online quality assessment in step S23, the scheme can perceive the image quality of each sampling point in real time, thereby achieving dynamic control of the data acquisition process in steps S24 and S25: high-quality data is retained, while low-quality data affected by factors such as lighting interference and occlusion triggers path correction. This closed-loop mechanism of "acquiring, evaluating, and adjusting simultaneously" enables proactive avoidance of adverse factors in the environment and adaptive exploration of better observation perspectives. Therefore, compared with the traditional method of fixed-path sampling, this scheme can acquire more qualified data pairs with higher quality and richer perspectives. Since the source data participating in the hand-eye calibration matrix calculation in step S26 has higher accuracy and consistency, and the final result is also systematically corrected by combining calibration error compensation strategies, the calculation accuracy and robustness of the hand-eye calibration matrix are improved. Therefore, even under complex working conditions such as changes in lighting and the presence of occlusion, the technical solution of this application can still stably obtain accurate hand-eye calibration results, laying a solid foundation for subsequent high-precision robot operations.

[0098] In one embodiment, the step of calculating the hand-eye calibration matrix of the target robot based on the calibration error compensation strategy, according to the calibration pose and each of the qualified data pairs, to obtain the calibration result includes:

[0099] S261: Calculate the hand-eye calibration matrix of the target robot based on the calibration pose and each of the qualified data pairs to obtain the first matrix;

[0100] S262: According to a preset first time window, obtain data from the historical calibration dataset corresponding to the target robot as the first data;

[0101] S263: Input the first data into the pre-trained first prediction model to predict the drift trend of the hand-eye calibration matrix and obtain the first prediction result;

[0102] S264: Generate a calibration compensation matrix based on the first prediction result, as the second matrix;

[0103] S265: Based on the second matrix, perform calibration error compensation on the first matrix to obtain the calibration result.

[0104] The first matrix refers to the initial hand-eye transformation matrix obtained directly by the hand-eye calibration algorithm based on the currently collected qualified data and the calibration pose during the hand-eye calibration process.

[0105] The first time window refers to the pre-defined time range selection criteria when extracting calibration data from historical calibration datasets. It limits the time interval of historical calibration results used for drift trend analysis, typically by looking back over a preset period (e.g., 10 days, 30 days, 90 days) or a preset number of calibrations (e.g., the last 10 calibrations) from the current moment. The data source for the first time window is preset configuration parameters, which can be dynamically adjusted according to the robot's operating conditions and accuracy requirements.

[0106] First data refers to the set of historical records related to hand-eye calibration, selected from the historical calibration dataset corresponding to the target robot according to a preset first time window. First data includes the hand-eye transformation matrix obtained from each calibration calculation, calibration timestamp, calibration quality score, and operating parameters (such as cumulative running time, ambient temperature, etc.).

[0107] The calibration quality score specifically includes: After completing the hand-eye calibration matrix solution, the calibration quality score is calculated based on the following dimensions: (1) Reprojection error: The feature points in the qualified data pairs are back-projected to the image space through the solved hand-eye calibration matrix, and the average pixel error between the actual detected feature points is calculated; the smaller the error, the higher the calibration accuracy. (2) Number of qualified data pairs: The total number of qualified data pairs participating in the solution. The more data, the stronger the statistical significance, and the higher the stability of the calibration results. (3) Data distribution diversity: The distribution of the robot end pose covered by all qualified data pairs in space is evaluated, including distance coverage range, angle coverage range, etc.; the more uniform the distribution, the stronger the generalization ability of the calibration results to extrapolate the scene. (4) Average image quality score: The arithmetic mean of the image quality scores of the single sampling points corresponding to all qualified data pairs, reflecting the basic quality level of the source data. After normalizing the above-mentioned indicators, a weighted fusion method is used to calculate the comprehensive score. For example, the reprojection error index is converted into an inverse score (the smaller the error, the higher the score), and usually has the highest weight; the number of qualified data pairs is normalized according to the preset target value, and has the second highest weight; the diversity of data distribution is calculated based on the ratio of the covered space volume to the theoretical maximum covered volume, and has a moderate weight; the average image quality score is directly adopted as the normalized mean, and has a relatively low weight.

[0108] Historical calibration dataset: This refers to a structured collection of information used to store data and results from all previous hand-eye calibration processes of the target robot. The historical calibration dataset includes the hand-eye transformation matrix calculated for each calibration, the calibration timestamp, the calibration quality score (such as reprojection error, number of qualified data pairs, data distribution diversity, etc.), and the operating parameters during calibration (such as cumulative runtime, ambient temperature, workload, etc.). The data for the historical calibration dataset is automatically stored after each hand-eye calibration process, along with relevant operating data, forming a historical data resource that can be called upon and fused for subsequent calibration processes, used for drift trend prediction, calibration error compensation, and fusion calibration calculations.

[0109] The first prediction model refers to a machine learning or time series analysis model trained on historical calibration data to predict the drift trend of the hand-eye calibration matrix. This model can capture the patterns of hand-eye relationship changes with factors such as cumulative runtime, ambient temperature, and work intensity, and output the theoretical drift amount for the future or current time. The first prediction model employs any one of the following: multiple linear regression, time series model, or Gaussian process regression model, and performs parameter fitting based on the historical calibration dataset through supervised training. The inputs to the first prediction model are "each component of the historical calibration data, cumulative runtime, and ambient temperature," and the output is "the predicted drift amount (Δx, Δy, Δz, Δθ) for the current time." The first prediction model is trained by minimizing the mean squared error between the predicted drift amount and the historical actual drift amount.

[0110] Each historical calibration record in the first data is extracted through feature engineering into the following input vector x = [Δt, T_env, T_joint, L_cum, Q_score], where Δt is the cumulative runtime (in hours) from the calibration completion time of the first data to the current prediction time; T_env is the ambient temperature (in °C) at the calibration time corresponding to the first data; T_joint is the average temperature of the robot joints (in °C) at the calibration time corresponding to the first data; L_cum is the cumulative workload (in tons per cycle or dimensionless normalized value) at the calibration time corresponding to the first data; and Q_score is the calibration quality score (between 0 and 1) corresponding to the first data. The input vector x serves as the input to the first prediction model.

[0111] The output of the first prediction model is the drift of each component of the hand-eye calibration matrix, represented as a 6-dimensional vector y = [Δx, Δy, Δz, Δrx, Δry, Δrz], where Δx, Δy, and Δz are the drift of the translation component (unit: mm); and Δrx, Δry, and Δrz are the drift of the rotation component (unit: degrees or radians).

[0112] The first prediction result refers to the predicted value of the hand-eye calibration matrix drift trend output by the model after inputting the first data into the pre-trained first prediction model. The first prediction result reflects the theoretical offset of the current hand-eye relationship relative to the baseline state, taking into account historical drift patterns.

[0113] The calibration compensation matrix is ​​a transformation matrix generated based on the first prediction result and used to correct the currently calculated hand-eye calibration matrix. This matrix is ​​expressed in 4×4 homogeneous transformation form and aims to eliminate residual biases introduced by systematic factors such as mechanical wear and temperature drift.

[0114] Calibration error compensation refers to the process of correcting the currently calculated first matrix using a calibration compensation matrix generated based on historical drift trends. This compensation is achieved through matrix multiplication, that is, multiplying the first matrix by the second matrix (or performing corresponding operations according to the drift direction) to obtain the final calibration result after eliminating systematic residual biases.

[0115] Step S261 specifically includes: summarizing all qualified data pairs accumulated in step S24, where each qualified data pair contains a qualified image and its corresponding robot single-sampling point end-effector pose. Based on the hand-eye configuration mode (eye on the hand or eye outside the hand), a corresponding hand-eye calibration algorithm (such as AX=XB or AX=ZB) is selected. Using the calibration pose as a known reference, a system of equations is constructed and solved to obtain the first matrix describing the relationship between the current vision and the robot coordinates.

[0116] Step S262 specifically includes: reading a preset first time window parameter, which defines the range of historical data to be traced back (e.g., the most recent 90 days or the most recent 10 calibration records). Then, accessing the historical calibration dataset corresponding to the target robot, filtering out all calibration records with timestamps within the time window, each record containing the hand-eye transformation matrix calculated from historical calibration, the cumulative runtime at the calibration time, the ambient temperature, and the calibration quality score, and integrating the filtering results into the first data.

[0117] Step S263 specifically includes: using the first data as input features and feeding it into a pre-trained first prediction model. This model, based on time series regression or machine learning algorithms, fits the drift trend function of each component of the hand-eye matrix as a function of factors such as runtime and temperature. The model outputs the theoretical drift amount at the current moment or under the current calibration conditions as the first prediction result.

[0118] Step S264 specifically includes: converting the first prediction result (i.e., the predicted drift amount) into a homogeneous transformation matrix form. Specifically, if the first prediction result predicts a drift of Δx, Δy, or Δz in the translation vector, or a drift of Δθ in the rotation angle, then the corresponding translation compensation matrix or rotation compensation matrix is ​​generated. These compensation components are combined into a complete 4×4 homogeneous transformation matrix, which serves as the calibration compensation matrix, i.e., the second matrix.

[0119] Step S265 specifically includes: performing matrix multiplication between the first matrix calculated in step S261 and the second matrix generated in step S264. The direction of compensation depends on the physical meaning of the drift trend: if the second matrix represents the drift from the reference state to the current state, the final calibration result can be obtained by multiplying the first matrix by the inverse of the second matrix (or by performing corresponding operations according to the model definition). The hand-eye transformation matrix obtained after compensation is the final calibration result after eliminating systematic residual bias.

[0120] The first matrix, X_current, is the currently calculated hand-eye calibration matrix, describing the transformation relationship from the target vision device coordinate system to the robot end effector coordinate system (eye-on-hand mode) or to the robot base coordinate system (eye-outside-hand mode). The second matrix, X_comp, is the calibration compensation matrix generated based on the first prediction result, describing the correction transformation from the current drift state to the reference state. The compensation operation uses a right multiplication method, that is, the second matrix is ​​used as a right multiplication factor on the first matrix: X_calibrated = X_current · X_comp, where X_comp is defined as the transformation matrix from the current coordinate system to the reference coordinate system. When the drift is positively cumulative, the physical meaning of X_comp is "the transformation required to correct the current coordinate system back to the reference coordinate system," and the right multiplication order ensures that the compensation amount takes effect in the current coordinate system.

[0121] If the hand-eye configuration is set to eye-out mode, the calibration results describe the transformation between the fixed camera and the robot base. The compensation calculation also uses right multiplication to ensure consistency in mathematical form.

[0122] The accuracy of the first matrix obtained in step S261 of this embodiment is limited by the quality of the collected data and the residual error of the algorithm itself. Steps S262 to S264 introduce a drift trend prediction mechanism based on historical data: the system obtains the first data within the first time window from the historical calibration dataset, uses the first prediction model to identify the systematic drift pattern of hand-eye relationship with changes in running time, temperature, and other factors, and generates a calibration compensation matrix accordingly. This compensation matrix is ​​not a simple averaging of random errors, but an explicit modeling of systematic errors with trends and accumulation. In step S265, the system applies the compensation matrix to the first matrix, realizing a systematic correction of the current solution result. This mechanism eliminates the residual deviations introduced by long-term drifts such as mechanical wear and thermal deformation that cannot be overcome by a single calibration process, making the final calibration result more accurate and robust than the first matrix. Therefore, through the dual processing of "historical trend prediction + current result compensation", this application further improves the accuracy of hand-eye calibration in the error compensation stage of the calibration result, providing a more reliable coordinate transformation benchmark for subsequent robot operation stages.

[0123] In one embodiment, the step of calculating the hand-eye calibration matrix of the target robot based on the calibration pose and each of the qualified data pairs to obtain the first matrix further includes:

[0124] S2611: Obtain the historical calibration dataset of the target robot;

[0125] S2612: Calculate the hand-eye calibration matrix of the target robot based on the historical calibration dataset, the calibration pose, and each qualified data pair to obtain the first matrix.

[0126] Step S2611 specifically includes: Before performing the hand-eye calibration matrix calculation, first accessing the historical calibration dataset associated with the target robot and reading the stored calibration records. This database can be located on the local storage medium of the target robot's controller or deployed on a host server. Based on the identification information of the current calibration task, all historical calibration data belonging to the same robot body are filtered out, including the hand-eye transformation matrix calculated from each calibration, the corresponding calibration timestamp, calibration quality score, and working parameters, as the basic data for subsequent calculations.

[0127] Step S2612 specifically includes: taking all the qualified data pairs accumulated in step S24, the calibration poses obtained in step S21, and the historical calibration data in the historical calibration dataset obtained in step S2611 as input, and using a weighted hand-eye calibration algorithm that integrates historical information to solve the first matrix. Specifically, different weights can be assigned to the historical calibration matrix according to the quality scores of each calibration in the historical calibration dataset, or a Bayesian estimation method can be used to use the historical calibration results as prior information, and the equations constructed with the currently collected qualified data pairs can be solved together to obtain the first matrix that integrates historical drift patterns and current measured data.

[0128] Weighted fusion based on quality scores: The quality scores of each calibration in the historical calibration dataset are normalized and used as weights for each historical calibration matrix. When constructing the residual function of the hand-eye calibration equation system, the residual terms corresponding to the historical calibration matrices are multiplied by this weight. The current qualified data is assigned a baseline weight (e.g., 1) to the corresponding residual terms, forming a weighted least squares objective function. The fused first matrix is ​​obtained through optimization. Historical calibration results with higher weights contribute more to the solution, while those with lower weights are suppressed.

[0129] The fusion method based on Bayesian estimation involves using the calibration results and their covariance matrices from previous calibration datasets as prior distributions, representing prior knowledge of the calibration parameters. The calibration equations constructed from currently collected qualified data pairs are used as likelihood functions, representing current observation information. Bayes' theorem is then employed to combine the prior and likelihood to obtain the posterior distribution, and the maximum point of the posterior distribution is taken as the solution to the first matrix. Mathematically, this is equivalent to introducing a regularization term into the solution of the calibration equations, ensuring that the solution fits the current data without deviating from historical statistical patterns.

[0130] This embodiment obtains the historical calibration dataset in step S2611, thereby utilizing historical information accumulated from previous calibrations, rather than relying solely on the qualified data pairs collected in the current single sampling. In step S2612, the historical calibration data is introduced as prior information into the calculation process of the hand-eye calibration matrix. The historical calibration data contains long-term statistical patterns of hand-eye relationships changing with factors such as runtime and temperature, as well as confidence information of the quality scores from previous calibrations. Jointly solving the first matrix with the historical calibration dataset and the current qualified data pairs is equivalent to introducing regularization constraints or Bayesian priors into the solution of the calibration equations. This mechanism effectively suppresses the interference of random noise or local outliers that may exist in the current single sampling on the calibration results, so that the calculated first matrix not only reflects the measured information at the current moment, but also takes into account the long-term stable trend of the system. Therefore, compared with the method of using only the current qualified data pairs for calibration calculation, this scheme improves the stability and accuracy of the first matrix by integrating the historical calibration dataset, providing a more reliable initial benchmark for subsequent calibration error compensation.

[0131] In one embodiment, the step of acquiring the calibration instructions and operational constraint data of the target robot includes:

[0132] S11: Based on the preset calibration start conditions, obtain the calibration instructions and operation constraint data of the target robot;

[0133] The calibration activation conditions include one or more of the following: a calibration signal input by the user, a calibration signal triggered by the adjustment of the target vision device, a calibration signal triggered by the adjustment of the target robot, the difference between the current operating temperature and the calibration temperature being greater than a preset temperature value, the operating accuracy corresponding to the current calibration result being greater than a preset accuracy, and an active prevention activation condition.

[0134] Preset calibration start conditions: These refer to a pre-configured set of judgment conditions used to automatically or manually trigger the hand-eye calibration process. This set of conditions includes various trigger types: user-input calibration signals (e.g., an operator clicking the "Start Calibration" button via the human-machine interface); calibration signals triggered by adjustments to the target vision device (e.g., after a camera is reinstalled due to a collision or maintenance); calibration signals triggered by adjustments to the target robot (e.g., after the robotic arm end effector is disassembled and reinstalled); the difference between the current calibration result's operating temperature and the calibration temperature is greater than the preset temperature value (e.g., an ambient temperature change exceeding 5°C causing thermal deformation); the current calibration result's operating accuracy is greater than the preset accuracy (e.g., a decrease in vision-guided grasping success rate exceeding the process allowable range); and proactive prevention start conditions (e.g., preventative calibration automatically triggered based on drift trend prediction). The data sources for the preset calibration start conditions are system configuration files, process parameters issued by the production management system, and the evaluation time generated based on the prediction model in steps S111 to S115.

[0135] Step S11 specifically includes: real-time monitoring of various trigger signals in the preset calibration start conditions. When any trigger condition is met, a calibration command is automatically generated or received, which includes the identification information of the calibration object and the expected calibration accuracy level. Simultaneously, current operational constraint data is obtained from the target robot's controller or production management system. The calibration command and operational constraint data are integrated and used as input for the subsequent step S2 to initiate the hand-eye calibration process.

[0136] This embodiment, by setting preset calibration initiation conditions, enables the system to automatically trigger the hand-eye calibration process in various scenarios. When the user actively initiates calibration, it meets the need for manual intervention; when the target vision device or target robot undergoes physical changes, it can promptly detect and trigger recalibration, avoiding accuracy failure due to changes in relative position; when the ambient temperature changes beyond a preset threshold or the current operational accuracy exceeds the process allowable range, it proactively initiates calibration to ensure that the hand-eye relationship always meets operational requirements. In particular, the introduction of proactive prevention initiation conditions allows for autonomous judgment of calibration timing based on drift trend prediction, eliminating the need for manual intervention. Therefore, this solution, through multi-dimensional and multi-scenario calibration initiation condition design, achieves an intelligent upgrade from "passively waiting for manual calibration" to "actively monitoring and triggering on demand." It can automatically initiate calibration before accuracy is about to decline but before defects occur, improving the robot's automation level and operational reliability, while reducing manual maintenance costs and the risk of production losses due to untimely calibration.

[0137] In one embodiment, when the calibration initiation condition includes an active prevention initiation condition, and the method for acquiring calibration instructions and operational constraint data of the target robot based on the active prevention initiation condition includes:

[0138] S111: Obtain prediction instructions, respond to the prediction instructions, and obtain data from the historical calibration dataset corresponding to the target robot according to the preset second time window, as the second data;

[0139] S112: Input the second data into the pre-trained first prediction model to predict the drift trend of the hand-eye calibration matrix, obtain the second prediction result, and determine the evaluation time based on the second prediction result and the preset first drift threshold.

[0140] S113: Obtain the actual drift data of the hand-eye calibration matrix of the target robot according to the evaluation time, and make a calibration judgment based on the actual drift data and the preset second drift threshold to obtain the judgment result;

[0141] S114: If the judgment result is yes, then the calibration instructions and operation constraint data of the target robot are obtained, and the second prediction result is used as prior information to guide the generation of the sampling path sequence, so that the sampling path sequence has a higher sampling point density in the drift direction indicated by the second prediction result.

[0142] S115: If the judgment result is negative, then the next prediction instruction is generated based on the actual drift data, the second drift threshold and the second prediction result.

[0143] Prediction command: refers to the control signal generated internally by the method described in this application for initiating drift trend prediction and evaluation time determination. The prediction command can be periodically triggered by a pre-set timer, or dynamically generated by the method described in this application in step S115 based on the actual drift data, the second drift threshold, and the second prediction result, forming a closed-loop prediction and evaluation cycle.

[0144] The second time window refers to the time range set for selecting data from the historical calibration dataset when predicting drift trends based on proactive prevention activation conditions. This time window limits the time interval of the historical calibration data used for prediction, typically looking back a preset duration (e.g., 30 days, 90 days) or a preset number of calibrations (e.g., the last 10 calibrations) from the current moment. The data source for the second time window is the system's preset configuration parameters, which can be dynamically adjusted according to the target robot's operating conditions and accuracy requirements.

[0145] The proactive prevention activation condition is one of the preset calibration activation conditions. It refers to predicting the drift trend of the hand-eye calibration matrix based on historical calibration data and the first prediction model. When the predicted drift reaches a preset threshold, calibration is automatically triggered, thereby completing preventive maintenance before accuracy fails and realizing an intelligent upgrade from passive correction to proactive early warning.

[0146] Understandably, the first time window serves "calibration error compensation," aiming to improve the accuracy of the current calibration; the second time window serves "active prevention start conditions," aiming to determine when to initiate the next calibration. They are located at different points on the timeline and serve different closed-loop control stages.

[0147] The second data refers to all historical calibration data selected from the historical calibration dataset corresponding to the target robot according to a preset second time window. The second data includes the hand-eye transformation matrix obtained from each calibration calculation, calibration timestamp, calibration quality score, and operating parameters (such as cumulative running time, ambient temperature, etc.).

[0148] It is understandable that the first and second data are not fundamentally different in terms of data format and content; their difference lies in their application purposes. The first data is used for post-compensation of the current calibration results and serves as the input to the "calibration error compensation strategy." The second data is used for pre-prediction of future drift trends and serves as the input to the "active prevention activation conditions." Together, they constitute a two-way closed loop of "calibration stage error compensation" and "operation stage drift early warning" in this application.

[0149] The second prediction result refers to the predicted value of the hand-eye calibration matrix drift trend output by the model after inputting the second data into the pre-trained first prediction model. The second prediction result reflects the theoretical offset of the current hand-eye relationship relative to the baseline state, taking into account historical drift patterns.

[0150] The first drift threshold is a critical value used to determine whether the drift trend has reached the point where an evaluation time needs to be set. This threshold is preset in the form of the absolute value of the drift amount (such as the change in translation vector or the change in rotation angle). When the second prediction result exceeds the first drift threshold, an evaluation time is determined to perform actual drift detection. The data source for the first drift threshold is the configuration parameters preset by the system according to the process accuracy requirements.

[0151] Evaluation time: This refers to the time point at which the system calculates the actual drift data acquisition based on the second prediction result and the first drift threshold. This time point is usually set an advance amount before the predicted drift amount reaches the second drift threshold to ensure sufficient time to complete recalibration before accuracy fails.

[0152] Actual drift data refers to the actual hand-eye calibration matrix drift measured at the time of evaluation, obtained through a simplified calibration or accuracy verification process. Actual drift data is used to verify the accuracy of the prediction results and is compared with a second drift threshold to determine whether a full calibration is necessary.

[0153] The second drift threshold is a critical value used to determine whether a complete hand-eye calibration is required. This threshold is preset as the absolute value of the drift amount (such as the change in translation vector or rotation angle). When the actual drift data exceeds the second drift threshold, recalibration is deemed necessary. The data source for the second drift threshold is a configuration parameter preset by the system based on process accuracy requirements. Typically, the second drift threshold is less than or equal to the maximum allowable accuracy error of the process.

[0154] First Drift Threshold: The first drift threshold is a warning threshold used to trigger actual drift verification, and its physical meaning is "the maximum allowable predicted drift amount". This threshold is expressed in the form of Euclidean distance of a 6-dimensional drift vector: D_pred = sqrt(Δx² + Δy² + Δz² + k_rot·(Δrx² + Δry² + Δrz²)), where k_rot is the scale normalization coefficient of the rotation component and the translation component (e.g., k_rot = 100, which equates 1 degree rotation to 1 mm translation). The value of the first drift threshold is set according to the robot's operational accuracy requirements, usually taking 50% to 70% of the maximum allowable error in the process, for example, set to 0.5 mm (translation equivalent error).

[0155] Second drift threshold: The second drift threshold is the execution threshold used to trigger the full calibration, and its physical meaning is "the maximum allowable actual drift amount". This threshold is also expressed in the form of Euclidean distance of a 6-dimensional drift vector. The value is set according to the robot's operation accuracy requirements, usually taking 80% to 100% of the maximum allowable error of the process, for example, set to 1.0 mm (translation equivalent error).

[0156] The first drift threshold and the second drift threshold satisfy the following relationship: the warning threshold is less than the execution threshold. The margin between them (the result of the second drift threshold minus the first drift threshold is greater than 0) ensures that there is sufficient time window between the warning trigger and the actual accuracy deviation to complete the evaluation time determination, actual drift data acquisition, and the complete calibration process. This margin is dynamically set based on parameters such as robot drift rate and calibration time, and is typically the drift increment corresponding to the calibration time.

[0157] Judgment result: This refers to the binary decision result obtained by comparing the actual drift data with the preset second drift threshold. If the actual drift data is greater than or equal to the second drift threshold, the judgment result is "yes", triggering the calibration process; if the actual drift data is less than the second drift threshold, the judgment result is "no", generating the next prediction instruction.

[0158] Step S111 specifically includes: after receiving the prediction instruction, reading the preset second time window parameters. Subsequently, accessing the historical calibration dataset corresponding to the target robot, filtering out all historical calibration data whose timestamps are within the second time window, and integrating the filtering results into the second data.

[0159] Step S112 specifically includes: using the second data as input features and feeding it into the pre-trained first prediction model. The model outputs a prediction curve showing the change in hand-eye calibration matrix drift over a future period. The system compares this prediction curve with a preset first drift threshold, calculates the time point when the drift reaches the first drift threshold, and subtracts a preset lead time (such as 24 hours or 50 work cycles) from this time point to determine the evaluation time, ensuring sufficient margin for verification and calibration before actual accuracy failure.

[0160] Step S113 specifically includes: when the clock reaches the evaluation time, the method described in this application triggers a simplified accuracy verification process (such as acquiring a small number of calibration images during the work interval and calculating the current hand-eye matrix, quickly calculating the current hand-eye matrix, and comparing it with the reference matrix to obtain the actual drift data), and obtaining the current actual hand-eye calibration matrix drift data. The actual drift data is compared with a preset second drift threshold: if the actual drift data is greater than or equal to the second drift threshold, the judgment result is "yes"; if the actual drift data is less than the second drift threshold, the judgment result is "no".

[0161] In step S114, specifically, when the judgment result is yes, the method described in this application determines that the actual drift of the current hand-eye relationship has reached the level requiring recalibration, automatically generates a calibration command, and obtains the current operation constraint data from the target robot's controller or production management system. Unlike conventional calibration processes, this step, while triggering calibration, uses the second prediction result generated in step S112 (i.e., the predicted drift trend value of the hand-eye calibration matrix output by the first prediction model) as prior information and transmits it to the sampling path sequence generation module in step S21.

[0162] In step S21, the algorithm for generating the sampling path sequence does not employ a uniform spatial sampling strategy. Instead, it optimizes the sampling path sequence based on the drift direction and magnitude indicated by the second prediction result. Specifically, the second prediction result includes the predicted drift trends of each component of the hand-eye calibration matrix, such as the predicted drift amounts of the translation components (Δx, Δy, Δz) or rotation components (Δrx, Δry, Δrz). The algorithm for generating the sampling path sequence normalizes the absolute values ​​of the predicted drift amounts of each component in the second prediction result and uses this normalized value as a sampling density weighting factor in the corresponding spatial dimension.

[0163] For example, if the second prediction result predicts a significant drift in the translation component of the hand-eye calibration matrix in the Z-axis direction (i.e., a large absolute value of the predicted drift Δz), then the method described in this application increases the density of sampling points in the Z-axis direction when generating the sampling path sequence. For example, more sampling points with distance gradients are set in the normal direction of the calibration object to collect image data containing rich depth variations. If a significant drift in the rotation component around the X-axis direction is predicted (i.e., a large absolute value of the predicted drift Δrx), then the density of rotation sampling points around the X-axis is increased in the pose space. For example, more sampling poses with angle gradients are set in the pitch direction of the calibration object.

[0164] Through the above mechanism, the sampling point density of the sampling path sequence in the drift-sensitive dimension is enhanced, resulting in higher information gain for the qualified data pairs collected in subsequent steps S23 to S25 in the key drift direction. Thus, this step achieves a deep coupling between the "active prevention initiation condition" and the "adaptive optimization path strategy": the predicted drift trend (second prediction result) serves as prior knowledge, actively guiding the sampling path in the calibration stage to focus on the spatial region with the richest information, enabling the subsequent calibration error compensation stage to obtain the most targeted source data, thereby achieving higher calibration accuracy with the same number of samplings, forming a positive feedback synergy of "prediction-collection-compensation".

[0165] A production management system is a comprehensive information platform used to plan, schedule, monitor, and optimize the production process, typically including a Manufacturing Execution System (MES) and an Enterprise Resource Planning (ERP) system. The production management system is responsible for issuing work instructions to the robot control terminal. These instructions include information such as the target object's identifier, job type, process parameters, and expected cycle time. Simultaneously, the system can provide real-time feedback on the robot's operating status, job completion status, and quality data. In the technical solution of this application, the production management system, as the data source for judging job accuracy in calibration start conditions and as the upstream provider of work instructions and job constraint data, provides business-level input support for achieving automated triggering of hand-eye calibration and job drift compensation.

[0166] Step S115 specifically includes: when the judgment result is "no", the method described in this application determines that the current actual drift has not yet reached the recalibration threshold. The method described in this application uses the actual drift data as feedback to correct the second prediction result, re-estimate the drift trend, and calculate the trigger time of the next prediction command (for example, determine the execution time of the next prediction command based on the remaining drift margin and the predicted drift rate), generate a new prediction command, and form a closed-loop monitoring and prediction cycle.

[0167] When the judgment result is negative (i.e., the actual drift data is less than the second drift threshold), the method of this application performs the following steps: The first prediction model is updated online using the residual. For Gaussian process regression models, the model hyperparameters can be re-optimized by using the current actual drift data and its corresponding operating condition features as new training samples, or incremental learning can be performed using a sparse Gaussian process. For multiple linear regression models, the regression coefficients can be updated using recursive least squares. Based on the corrected drift trend, the time required for the drift amount to reach the first drift threshold is re-estimated. The specific calculation formula is: t_next = (D_pred - D_actual) / v_drift, where v_drift is the current drift rate calculated based on the corrected drift trend (unit: mm / hour or equivalent unit); D_pred is the first drift threshold; D_actual is the actual drift data, which refers to the actual hand-eye calibration matrix drift amount measured by performing a simplified calibration or accuracy verification process when the evaluation time arrives. This data is used to verify the accuracy of the prediction results and is compared with a second drift threshold to determine whether a full calibration is needed. The current time plus t_next is used as the trigger time for the next prediction instruction. If t_next is less than a preset minimum interval (e.g., 1 hour), the minimum interval is used as the next trigger time to avoid excessively frequent prediction evaluations. The generated next prediction instruction includes a trigger timestamp and the identifier of the currently corrected model parameters, and automatically triggers a new prediction evaluation cycle when the preset timer expires.

[0168] This embodiment achieves preventative calibration under proactive prevention activation conditions through a closed-loop prediction and evaluation mechanism constructed in steps S111 to S115. Steps S111 and S112 utilize historical calibration datasets and a first prediction model to predict the drift trend of the hand-eye calibration matrix in advance, and determine the evaluation time based on a first drift threshold, thereby predicting when the accuracy is about to exceed the tolerance. Step S113 acquires the actual drift data when the evaluation time arrives and compares it with a second drift threshold, verifying the accuracy of the prediction and avoiding misjudgments that may occur if only the prediction result is relied upon. When the actual drift data exceeds the second drift threshold, step S114 automatically triggers the complete calibration process to complete the correction before the accuracy fails; when the actual drift data has not yet reached the threshold, step S115 dynamically adjusts the timing of the generation of the next prediction instruction based on the measured data, forming an adaptive monitoring frequency. Therefore, this embodiment upgrades hand-eye calibration from passive "post-correction" to proactive "preventive maintenance," enabling automatic calibration before accuracy declines but before defects occur, thus improving the automation level and operational reliability of robot control.

[0169] In one embodiment, the active prevention activation condition based on drift trend prediction is deeply coupled with the adaptive optimization path strategy based on online quality feedback and the subsequent calibration error compensation strategy to guide the strategy optimization of the next calibration process and improve the overall calibration efficiency. Specifically, when the active prevention activation condition (as described in steps S111 to S115) determines that calibration needs to be performed, the method of this application not only generates calibration instructions, but also passes the second prediction result (i.e., the predicted drift trend) as prior information to the calibration process in step S2. When generating the sampling path sequence in step S21, the method of this application optimizes the sampling path sequence according to the drift direction and drift magnitude indicated in the second prediction result. For example, if the second prediction result predicts that the translation component of the hand-eye calibration matrix has a significant drift in the Z-axis direction, the method of this application increases the sampling point density with rich depth changes in the Z-axis direction when generating the sampling path sequence; if the rotation component is predicted to drift in a specific axis direction, the method increases the attitude sampling points in that rotation axis direction. The algorithm for generating the sampling path sequence uses the second prediction result as a weighting factor for the spatial sampling density, thereby maximizing the information gain of subsequent data acquisition in the drift-sensitive direction.

[0170] Through the above mechanism, this embodiment achieves positive feedback synergy among "prediction-collection-compensation": First, the drift prediction (second prediction result) in the active prevention initiation condition indicates the "key observation area" for adaptive path planning, making the collected qualified data pairs have higher information content in the drift-sensitive dimension; Second, based on these high-quality and highly targeted data pairs, the first matrix calculated, after calibration error compensation based on historical drift trends (steps S261-S265), can obtain a more accurate final calibration result; Finally, this accurate calibration result, as part of the input data for the next prediction model (first prediction model), will further improve the accuracy of subsequent drift trend prediction, thus forming a self-optimizing closed loop. Compared with the open-loop mode of "blindly collecting data first and then uniformly compensating," or the quasi-closed-loop mode of "only triggering calibration based on prediction results but still using fixed path sampling," this embodiment deeply embeds the prediction results into the dynamic planning of the sampling path, making the "information collection" and "error correction" in the calibration process no longer isolated serial steps, but an integrated process that guides and reinforces each other. The synergistic effect of this integrated process enables the proposed solution to achieve higher calibration accuracy with the same number of samplings, or to significantly reduce the number of samplings required while achieving the same calibration accuracy. This improves calibration efficiency and enhances the robustness of calibration results to long-term drift.

[0171] In one embodiment, the step of compensating the calibration result according to a preset job drift compensation strategy during the operation of the target robot includes:

[0172] S41: During the operation of the target robot, the operation drift compensation matrix is ​​obtained as the third matrix according to the preset time interval, and the calibration result is compensated according to the third matrix.

[0173] The third matrix is ​​a calibration compensation matrix generated based on the work data corresponding to the work instruction and the equipment parameters of the target robot.

[0174] The operation data includes one or more of the following: operation duration, operation intensity, operation ambient temperature, operation vibration data, operation height, operation depth, and operation load data.

[0175] The preset time interval refers to the time period set by which the method described in this application periodically acquires the job drift compensation matrix during the target robot's operation. This time interval can be configured according to the target robot's work intensity, environmental change rate, and accuracy requirements, such as triggering a compensation update every 30 minutes, every 100 grasping actions completed, or every time a preset cumulative load is reached. The data source for the preset time interval is the system configuration file or process parameter settings.

[0176] Job drift compensation matrix: This is a transformation matrix used to correct hand-eye calibration results and offset real-time drift during operation. It is expressed in 4×4 homogeneous transformation form. This matrix is ​​dynamically generated based on the current job data and the target robot's equipment parameters. It aims to compensate for hand-eye relationship drift caused by factors such as accumulated job duration, changes in job intensity, fluctuations in ambient temperature, vibration interference, differences in job height and depth, and changes in load.

[0177] Operational data refers to the real-time set of parameters reflecting the operating status and working conditions of the target robot during the execution of work instructions. Operational data includes one or more of the following: operation duration (the time elapsed for this operation or the cumulative operation time), operation intensity (such as motion speed, acceleration, and joint torque), operating environment temperature (robot body temperature or ambient temperature), operation vibration data (vibration amplitude collected by accelerometers), operation height (the height position of the end effector in the workspace), operation depth (such as the insertion depth during deep cavity operations), and operation load data (end effector load weight or joint load rate). The data sources for operational data are the internal state parameters collected in real time by the target robot's controller, as well as measurement data from external sensors (such as temperature sensors and accelerometers).

[0178] The work intensity refers to a comprehensive quantitative index characterizing the motion load and dynamic characteristics of the target robot, specifically calculated using one of the following methods:

[0179] Method 1 (based on joint torque): Collect the average torque value of each joint of the target robot within the current time window, and normalize the weighted sum of the torques of each joint to the [0,1] interval.

[0180] Method 2 (based on end-effector motion parameters): Collect the average velocity and average acceleration of the target robot's end effector within the time window. The calculation formula is: I_op = (v_avg / v_max + a_avg / a_max) / 2, where v_avg is the average velocity, v_max is the preset maximum velocity reference value, a_avg is the average acceleration, and a_max is the preset maximum acceleration reference value.

[0181] Method 3 (based on cumulative power): Collect the average power consumption of each joint motor of the target robot within the time window, accumulate and normalize: I_op = (Σ P_j) / P_max, where P_j is the average power of the j-th joint motor and P_max is the preset maximum total power reference value.

[0182] The data source for the work intensity is the joint torque, joint speed, joint acceleration, or motor power parameters collected in real time by the target robot's controller, and acquired through an embedded interface at a preset frequency (e.g., 10Hz).

[0183] The target robot's equipment parameters refer to the set of inherent parameters describing the target robot's physical and kinematic characteristics, used as the benchmark for compensation calculations when generating the operational drift compensation matrix. The target robot's equipment parameters include the robot model, reduction ratios of each joint, coefficient of thermal expansion, rated load, stiffness coefficient, and drift compensation coefficients fitted based on historical data. The data sources for the target robot's equipment parameters include robot factory specifications, equipment maintenance manuals, or data obtained through offline identification experiments.

[0184] Step S41 specifically includes: during the target robot's execution of a task in response to a work instruction, a timed task is initiated to periodically perform compensation updates at preset time intervals. At the end of each compensation cycle, the task drift compensation matrix corresponding to the current moment is obtained as the third matrix. This third matrix is ​​then multiplied by the calibration result (hand-eye calibration matrix) obtained in step S2 to obtain the compensated hand-eye transformation matrix. The method described in this application uses this compensated matrix for subsequent robot motion control and visual guidance until the next compensation cycle arrives and it is updated again.

[0185] Within each compensation cycle or at a specific time (such as when a calibration command is received), real-time operational data (such as operational duration, operational intensity, ambient temperature, vibration data, operational height, operational depth, and load data) is collected from the target robot's controller, and pre-stored equipment parameters of the target robot are read. The operational data and equipment parameters are input into a preset compensation calculation module. This module calculates the theoretical drift of the hand-eye relationship relative to the baseline state under the current operating conditions based on offline calibrated compensation curves, empirical formulas, or machine learning models, and converts this drift into a homogeneous transformation matrix to generate an operational drift compensation matrix.

[0186] A compensation curve is a function curve (such as a polynomial curve, spline curve, or exponential decay curve) that is formed by measuring the drift of each component of the hand-eye calibration matrix through offline experiments, under the condition of controlling one or a few key operating variables (such as ambient temperature, joint temperature, and cumulative running time), and fitting the law of drift change with the variable.

[0187] Empirical formulas are explicit equations derived from theoretical analysis (such as thermal expansion theory and mechanics of materials) combined with experimental data, describing the mathematical relationship between drift and multiple operating parameters. Empirical formulas usually have a clear physical background, and their coefficients are obtained by fitting offline experimental data.

[0188] Machine learning models are black-box or gray-box models trained using supervised learning methods and large amounts of offline, multi-condition drift data. These models are capable of fitting the complex nonlinear mapping relationship between inputs (condition parameters) and outputs (drift). Common models include neural networks, support vector regression, random forests, and Gaussian process regression.

[0189] The compensation calculation module employs a drift compensation model based on the fusion of multivariate regression and offline calibration curves. This module takes operational data (operation duration, ambient temperature, load data, operation height and depth, etc.) and equipment parameters (thermal expansion coefficient, stiffness coefficient, reduction ratio, etc.) as input. Through pre-conducted full-condition calibration experiments in a laboratory or production line shutdown state, it fits the mapping relationship between each drift factor and each component of the hand-eye matrix, which can be represented as a multiple linear regression equation or a radial basis function neural network. During real-time operation, the module substitutes the currently collected operational data into the model, calculates each drift component, and synthesizes it into a compensation amount in the form of a 4×4 homogeneous transformation matrix, achieving dynamic correction of the calibration results.

[0190] When generating the job drift compensation matrix, one or more of the aforementioned job data can be selected as input variables for compensation calculation, based on the accuracy requirements and sensor configuration of the actual application scenario. For example, in high-precision assembly scenarios, job duration, ambient temperature, and load data can be comprehensively collected to compensate for cumulative wear and thermal deformation; in deep cavity operation scenarios, job depth data can be additionally introduced to compensate for nonlinear drift introduced by changes in the robotic arm's posture. Through the flexible combination of multi-dimensional job data, accurate modeling and compensation of drift factors under different working conditions can be achieved.

[0191] The job drift compensation matrix is ​​generated using a multiple linear regression model, with the following specific configuration:

[0192] Input feature definition: The combination of the operation data and equipment parameters forms an input vector: z=[t_op,I_op,T_env,V_ib,H_wk,H_wh,L_load,α_exp,K_stiff], where t_op is the operation duration (in hours); I_op is the operation intensity; T_env is the operation ambient temperature (in °C); V_ib is the operation vibration data (effective value of acceleration, in m / s²); H_wk is the operation height (Z coordinate of the end effector, in mm); H_wh is the operation depth; L_load is the operation load data (end effect load weight, in kg); α_exp is the coefficient of thermal expansion in the robot equipment parameters (in μm / (m·℃)); K_stiff is the stiffness coefficient in the robot equipment parameters (in N / mm).

[0193] Output format definition: The compensation calculation module outputs the predicted drift value, which is converted into a job drift compensation matrix in the form of a 4×4 homogeneous transformation matrix. The formula for calculating the predicted drift value is: Δp=β_0+β_1·t_op+β_2·I_op+β_3·T_env+β_4·V_ib+β_5·H_wk+β_9·H_wh+β_6·L_load+β_7·α_exp+β_8·K_stiff, where Δp is a 6-dimensional drift vector [Δx,Δy,Δz,Δrx,Δry,Δrz]; β_0, β_1, β_2, β_3, β_4, β_5, β_6, β_7, β_8, and β_9 are regression coefficient matrices fitted through offline experiments;

[0194] Model generation method: The regression coefficient moments are obtained through offline calibration experiments: While the robot is offline, a single operating condition variable is controlled to change (e.g., temperature rises from 10℃ to 50℃), and hand-eye calibration is performed at each operating point. The drift of the calibration results relative to the baseline state is recorded. After collecting multiple sets of data, the regression coefficients are fitted using the least squares method. Once sufficient actual drift data has accumulated during online operation, the regression coefficient moments can be incrementally updated periodically to improve the adaptability of the compensation model.

[0195] This embodiment, through a real-time compensation mechanism constructed during operation, can continuously address hand-eye relationship drift caused by factors such as mechanical wear, temperature changes, and load fluctuations during the execution of work instructions by the target robot. A job drift compensation matrix is ​​dynamically generated based on real-time collected job data and equipment parameters, ensuring that the compensation amount accurately matches the actual drift characteristics under the current working conditions. Step S41 periodically updates the compensation matrix at preset time intervals to ensure the continuity and timeliness of the compensation effect. Compared to the traditional method of performing correction only once during the calibration phase, this embodiment introduces a dynamic compensation step during operation, making the hand-eye calibration results no longer static parameters, but an adaptive benchmark that can be adjusted in real time according to changes in working conditions. The real-time inclusion of factors such as job duration reflecting cumulative wear effects, ambient temperature affecting thermal deformation, load data relating to structural deflection, and job height and depth corresponding to kinematic errors under different postures, allows the compensation matrix to accurately offset the superimposed effects of multi-source drift. Therefore, through the triple synergy of "drift compensation during operation" with the aforementioned "adaptive optimization during calibration" and "error compensation for calibration results", this application enables the robot to maintain high precision and stable operation capability under complex working conditions.

[0196] In one embodiment, for a new robot or the first calibration scenario, the historical calibration dataset is empty or insufficient. To address the cold start problem, this application employs one or a combination of the following strategies:

[0197] Strategy 1 (Pre-calibration before leaving the factory): Before the robot leaves the factory, perform hand-eye calibration a preset number of times (e.g., 5 times) under standard environmental conditions. Store the calibration results and corresponding operating parameters (ambient temperature 20℃, zero load, zero cumulative running time) into the historical calibration dataset as the initial benchmark data.

[0198] Strategy 2 (Same-Model Migration): For robots of the same model, import historical calibration data (after normalization) from other robots of the same model as prior data to form an initial historical calibration dataset. The migration data must be labeled as "same-model migration data" and gradually replaced in subsequent actual calibration.

[0199] Strategy 3 (Gradual Activation): During the initial calibration, temporarily disable calibration error compensation based on historical data and calibration matrix calculation based on historical data. Only use the currently collected qualified data for calibration. Simultaneously, store the calibration results in the historical calibration dataset. When the number of records in the historical calibration dataset reaches a preset threshold (e.g., 3 records), automatically enable the calibration error compensation function; when it reaches a higher preset threshold (e.g., 10 records), automatically enable the calibration matrix calculation based on historical data.

[0200] Strategy 4 (Simulation Data Population): Based on robot dynamics simulation software, virtual calibration is performed under various working conditions (different temperatures, different loads, and different cumulative running times) to generate simulation calibration data, which is then used to populate the historical calibration dataset as a temporary data source during the cold start period.

[0201] It is understood that the adaptive optimization robot control method and system for hand-eye calibration proposed in this application have at least the following beneficial effects: Compared with the existing hand-eye calibration methods that have fixed sampling paths, lack real-time evaluation and feedback mechanisms for image quality during sampling, and ignore long-term drift, this application constructs a cross-stage, feedback-based closed-loop control system. Through the deep synergy of three measures—"adaptive optimization of the acquisition process," "error compensation for calibration results," and "drift compensation during the operation phase"—intelligent enhancement of the entire calibration and operation process is achieved. Specifically, the "adaptive optimization of the acquisition process" and "error compensation for calibration results" are not executed in isolation, but are coupled through an active prevention initiation condition based on drift trend prediction. The drift trend predicted by the active prevention initiation condition serves as prior knowledge, actively guiding the adaptive optimization path strategy to generate a non-uniform sampling path sequence that maximizes information gain in key drift dimensions during the calibration phase. This allows the subsequent error compensation stage to obtain the most targeted source data, thereby achieving higher calibration accuracy with the same number of samplings. Building upon this foundation, a drift compensation strategy is further integrated into the operation phase to provide real-time compensation for hand-eye relationship drift during operation, thus avoiding the problem of calibration accuracy decaying over time. Therefore, this application, through a positive feedback loop of "prediction triggering guided path optimization, path optimization improving calibration accuracy, calibration accuracy supporting operation compensation, and operation compensation feeding back into the prediction model," enables the robot to maintain high-precision and stable operation capabilities under complex working conditions. Its overall technical effect surpasses the simple summation of the functions of "adaptive path planning," "error compensation," and "drift compensation," achieving a synergistic effect of "1+1+1>3."

[0202] Please see Figure 3As shown, in one embodiment, an adaptive optimization robot control system for hand-eye calibration is provided, the system comprising: a system controller 3 and a target robot 2, the system controller 3 being communicatively connected to the target robot 2, the system controller 3 being configured to implement the steps of any of the above-described adaptive optimization robot control methods for hand-eye calibration.

[0203] System controller 3 refers to the computing and control unit used to execute the adaptive optimization robot control method for hand-eye calibration described in this application. System controller 3 is communicatively connected to the target robot and configured to implement the steps of the adaptive optimization robot control method for hand-eye calibration described in this application. In actual deployment, system controller 3 can be integrated into the target robot 2 as its host controller or industrial control computer, or it can be a physical entity independent of the target robot 2, such as a host computer, server, or cloud control platform. The data sources for system controller 3 include status data fed back by the target robot, image data acquired by the target vision device, historical calibration datasets, and task instructions issued by the production management system. By integrating adaptive optimization, error compensation, and drift compensation mechanisms, system controller 3, as the core of robot decision-making and execution, achieves closed-loop control of the entire calibration and operation process.

[0204] This embodiment achieves comprehensive optimization of the calibration process by introducing a dual-sided collaborative mechanism combining an "adaptive optimization path strategy based on online quality feedback" and a "preset calibration error compensation strategy" for hand-eye calibration. The adaptive optimization path strategy based on online quality feedback dynamically adjusts the sampling path according to real-time image quality during data acquisition, effectively avoiding adverse factors such as illumination interference and occlusion, ensuring higher accuracy and consistency of the source data used in the calibration calculation, thereby improving the accuracy of the hand-eye calibration matrix solution. Furthermore, the preset calibration error compensation strategy systematically corrects the calibration results, eliminating residual deviations introduced by algorithmic or environmental factors, thus obtaining more accurate and robust initial calibration results. Moreover, in the subsequent operation phase, this application introduces a job drift compensation strategy during the robot's execution of work instructions to compensate for hand-eye relationship drift in real time, avoiding the problem of calibration accuracy decaying over time. Therefore, this application achieves high-precision and stable operation capability of the robot under complex working conditions through the synergistic effect of three measures: "adaptive optimization of the acquisition process", "error compensation of calibration results" and "drift compensation during the operation phase".

[0205] In one embodiment, a robot is provided, the robot including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the following steps:

[0206] Obtain the calibration instructions and operational constraint data of the target robot;

[0207] An adaptive optimization path strategy based on online quality feedback and a preset calibration error compensation strategy are adopted. The hand-eye calibration matrix of the target robot is calculated according to the calibration instructions and the operation constraint data to obtain the calibration result.

[0208] Obtain the working instructions of the target robot;

[0209] In response to the work instruction, the target robot is controlled to perform the operation according to the calibration result, and the calibration result is compensated according to the preset operation drift compensation strategy during the operation of the target robot.

[0210] The adaptive optimization path strategy based on online quality feedback, the calibration error compensation strategy, and the job drift compensation strategy work together to form a closed-loop control system across all stages.

[0211] This embodiment achieves comprehensive optimization of the calibration process by introducing a dual-sided collaborative mechanism combining an "adaptive optimization path strategy based on online quality feedback" and a "preset calibration error compensation strategy" for hand-eye calibration. The adaptive optimization path strategy based on online quality feedback dynamically adjusts the sampling path according to real-time image quality during data acquisition, effectively avoiding adverse factors such as illumination interference and occlusion, ensuring higher accuracy and consistency of the source data used in the calibration calculation, thereby improving the accuracy of the hand-eye calibration matrix solution. Furthermore, the preset calibration error compensation strategy systematically corrects the calibration results, eliminating residual deviations introduced by algorithmic or environmental factors, thus obtaining more accurate and robust initial calibration results. Moreover, in the subsequent operation phase, this application introduces a job drift compensation strategy during the robot's execution of work instructions to compensate for hand-eye relationship drift in real time, avoiding the problem of calibration accuracy decaying over time. Therefore, this application achieves high-precision and stable operation capability of the robot under complex working conditions through the synergistic effect of three measures: "adaptive optimization of the acquisition process", "error compensation of calibration results" and "drift compensation during the operation phase".

[0212] In one embodiment, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program that, when executed by a processor, performs the following steps:

[0213] Obtain the calibration instructions and operational constraint data of the target robot;

[0214] An adaptive optimization path strategy based on online quality feedback and a preset calibration error compensation strategy are adopted. The hand-eye calibration matrix of the target robot is calculated according to the calibration instructions and the operation constraint data to obtain the calibration result.

[0215] Obtain the working instructions of the target robot;

[0216] In response to the work instruction, the target robot is controlled to perform the operation according to the calibration result, and the calibration result is compensated according to the preset operation drift compensation strategy during the operation of the target robot.

[0217] The adaptive optimization path strategy based on online quality feedback, the calibration error compensation strategy, and the job drift compensation strategy work together to form a closed-loop control system across all stages.

[0218] This embodiment achieves comprehensive optimization of the calibration process by introducing a dual-sided collaborative mechanism combining an "adaptive optimization path strategy based on online quality feedback" and a "preset calibration error compensation strategy" for hand-eye calibration. The adaptive optimization path strategy based on online quality feedback dynamically adjusts the sampling path according to real-time image quality during data acquisition, effectively avoiding adverse factors such as illumination interference and occlusion, ensuring higher accuracy and consistency of the source data used in the calibration calculation, thereby improving the accuracy of the hand-eye calibration matrix solution. Furthermore, the preset calibration error compensation strategy systematically corrects the calibration results, eliminating residual deviations introduced by algorithmic or environmental factors, thus obtaining more accurate and robust initial calibration results. Moreover, in the subsequent operation phase, this application introduces a job drift compensation strategy during the robot's execution of work instructions to compensate for hand-eye relationship drift in real time, avoiding the problem of calibration accuracy decaying over time. Therefore, this application achieves high-precision and stable operation capability of the robot under complex working conditions through the synergistic effect of three measures: "adaptive optimization of the acquisition process", "error compensation of calibration results" and "drift compensation during the operation phase".

[0219] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0220] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0221] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0222] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. An adaptive optimization robot control method for hand-eye calibration, characterized in that, The method includes: Obtain the calibration instructions and operational constraint data of the target robot; An adaptive optimization path strategy based on online quality feedback and a preset calibration error compensation strategy are adopted. The hand-eye calibration matrix of the target robot is calculated according to the calibration instructions and the operation constraint data to obtain the calibration result. Obtain the working instructions of the target robot; In response to the work instruction, the target robot is controlled to perform the operation according to the calibration result, and the calibration result is compensated according to the preset operation drift compensation strategy during the operation of the target robot. Among them, the adaptive optimization path strategy based on online quality feedback, the calibration error compensation strategy, and the operation drift compensation strategy work together to form a closed-loop control system across all stages. The step of employing an adaptive optimization path strategy based on online quality feedback and a preset calibration error compensation strategy, and calculating the hand-eye calibration matrix of the target robot according to the calibration instructions and the task constraint data to obtain the calibration result includes: The calibration pose of the calibration object is obtained according to the calibration instruction, and a sampling path sequence is generated according to the calibration pose; Control the target robot to move along the sampling path sequence; When the target robot moves along the sampling path sequence to reach the sampling point, the target vision device is controlled to capture the calibrated object, and the capture result and the end pose of the target robot at a single sampling point are obtained. The quality is evaluated based on the capture result to obtain the evaluation result. If the evaluation result is qualified, the captured image and the single sampling point end pose are taken as qualified data pairs, and the process jumps to the step of controlling the target robot to move along the sampling path sequence to continue execution until there are no sampling points in the sampling path sequence that the target robot has not reached. If the evaluation result is unqualified, the sampling path sequence is updated according to the job constraint data, the calibration pose, the current evaluation result, and the current single sampling point end pose to move the target robot to a better observation pose. Then, the process jumps to the step of controlling the target robot to move along the sampling path sequence and continues until there are no sampling points in the sampling path sequence that the target robot has not reached. Based on the calibration error compensation strategy, the hand-eye calibration matrix of the target robot is calculated according to the calibration pose and each qualified data pair to obtain the calibration result; Wherein, the target vision device is located at the end of the robotic arm of the target robot and the calibration object is located at a predetermined position outside the target robot, or the target vision device is located at a predetermined position outside the target robot and the calibration object is located at the end of the robotic arm of the target robot; The step of calculating the hand-eye calibration matrix of the target robot based on the calibration error compensation strategy, according to the calibration pose and each qualified data pair, to obtain the calibration result includes: Based on the calibration pose and each of the qualified data, the hand-eye calibration matrix of the target robot is calculated to obtain the first matrix; According to the preset first time window, data is obtained from the historical calibration dataset corresponding to the target robot as the first data; The first data is input into the pre-trained first prediction model to predict the drift trend of the hand-eye calibration matrix, and the first prediction result is obtained. A calibration compensation matrix is ​​generated based on the first prediction result, which serves as the second matrix. Based on the second matrix, the first matrix is ​​subjected to calibration error compensation to obtain the calibration result.

2. The adaptive optimization robot control method for hand-eye calibration according to claim 1, characterized in that, The step of calculating the hand-eye calibration matrix of the target robot based on the calibration pose and each of the qualified data to obtain the first matrix further includes: Obtain the historical calibration dataset of the target robot; The first matrix is ​​obtained by calculating the hand-eye calibration matrix of the target robot based on the historical calibration dataset, the calibration pose, and each qualified data pair.

3. The adaptive optimization robot control method for hand-eye calibration according to claim 1, characterized in that, The steps for obtaining the calibration instructions and operational constraint data of the target robot include: Based on preset calibration start conditions, obtain the calibration instructions and operational constraint data of the target robot; The calibration activation conditions include one or more of the following: a calibration signal input by the user, a calibration signal triggered by the adjustment of the target vision device, a calibration signal triggered by the adjustment of the target robot, the difference between the current operating temperature and the calibration temperature being greater than a preset temperature value, the operating accuracy corresponding to the current calibration result being greater than a preset accuracy, and an active prevention activation condition.

4. The adaptive optimization robot control method for hand-eye calibration according to claim 3, characterized in that, The calibration initiation conditions include: when the active prevention initiation conditions are met, the method for obtaining calibration instructions and operational constraint data of the target robot based on the active prevention initiation conditions includes: Obtain a prediction instruction, respond to the prediction instruction, and obtain data from the historical calibration dataset corresponding to the target robot according to a preset second time window, as the second data; The second data is input into the pre-trained first prediction model to predict the drift trend of the hand-eye calibration matrix, and a second prediction result is obtained. The evaluation time is determined based on the second prediction result and the preset first drift threshold. The actual drift data of the hand-eye calibration matrix of the target robot is obtained according to the evaluation time. The calibration is determined based on the actual drift data and the preset second drift threshold, and the determination result is obtained. If the judgment result is yes, then the calibration instructions and operation constraint data of the target robot are obtained, and the second prediction result is used as prior information to guide the generation of the sampling path sequence, so that the sampling path sequence has a higher sampling point density in the drift direction indicated by the second prediction result. If the judgment result is negative, then the next prediction instruction is generated based on the actual drift data, the second drift threshold, and the second prediction result.

5. The adaptive optimization robot control method for hand-eye calibration according to claim 1, characterized in that, The step of compensating the calibration result according to a preset job drift compensation strategy during the operation of the target robot includes: During the operation of the target robot, a job drift compensation matrix is ​​obtained as a third matrix according to a preset time interval, and the calibration result is compensated according to the third matrix. The third matrix is ​​a calibration compensation matrix generated based on the work data corresponding to the work instruction and the equipment parameters of the target robot. The operation data includes one or more of the following: operation duration, operation intensity, operation ambient temperature, operation vibration data, operation height, operation depth, and operation load data.

6. An adaptive optimization robot control system for hand-eye calibration, characterized in that, The system includes a system controller and a target robot, the system controller being communicatively connected to the target robot, and the system controller being configured to implement the steps of the adaptive optimization robot control method for hand-eye calibration as described in any one of claims 1 to 5.

7. A target robot, characterized in that, The robot includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the adaptive optimization robot control method for hand-eye calibration as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the adaptive optimization robot control method for hand-eye calibration as described in any one of claims 1 to 5.

Images