Robot adaptive calibration method based on load change, electronic device and system

Abstract

The application provides a robot adaptive calibration method based on load change, an electronic device and a system. The method comprises the following steps: in response to the change of the end load, controlling each joint of the robot to move according to a preset excitation trajectory during a work gap, and obtaining joint motion data of a target joint, inertial measurement data of a robot wrist and six-dimensional moment data of a robot end; based on the above data, using a recursive least square method to identify a linear load inertia parameter vector of the end load; controlling the robot to stop moving and obtaining working condition data; according to the linear load inertia parameter vector and the working condition data, obtaining a kinematic parameter correction amount of the target joint, a sensor reference offset amount and a mapping confidence degree through a structure deformation mapping model; and based on the kinematic parameter correction amount and / or the sensor reference offset amount, calibrating a calibration object determined based on the mapping confidence degree. Thus, online response to load change, high-precision adaptive calibration without interrupting production are realized.

Description

Technical Field

[0001] This application relates to the field of robot calibration technology, and in particular to a robot adaptive calibration method, electronic device and system based on load variation. Background Technology

[0002] In fields such as industrial automation, flexible manufacturing, and human-robot collaboration, robot systems are widely used for tasks such as flexible assembly, material handling, end-effector tooling switching, and precision operations. During actual operation, the mass, center of mass, and inertia tensor of the robot's end-effector load frequently change due to workpiece replacement, tool switching, or changes in the object being held. This dynamic load variation causes elastic deformation of the robot's structural links, adjustments to joint compliance, and shifts in sensor mounting references. This leads to the gradual invalidation of offline calibrated kinematic parameters (such as link length and joint zero position) and extrinsic parameters (such as the pose relationship between the vision sensor and the end-effector), resulting in decreased kinematic accuracy and even affecting operational safety.

[0003] In related technologies, two main approaches are used to address the decrease in kinematic accuracy caused by load variations: kinematic compensation based on joint encoders and kinematic parameter calibration. The kinematic compensation method based on joint encoders obtains the joint angle compensation amount by looking up tables or error models using feedback data from the joint encoders; the command angle is then corrected based on this compensation amount. However, this approach relies solely on joint angle information and cannot detect changes in link elastic deformation and joint compliance caused by end-effector torques, resulting in a single compensation dimension and limited accuracy. The kinematic parameter calibration method, such as using a laser tracker for offline calibration, acquires the actual end-effector pose using the laser tracker to inversely solve for kinematic parameter errors and correct the kinematic model parameters, thereby improving model accuracy. However, this method is complex, time-consuming, and requires production interruption, making it unable to respond online to dynamic load changes and difficult to meet the requirements of continuous operation cycle time.

[0004] Therefore, there is an urgent need for an adaptive calibration solution that can respond to load changes online, without the need for external equipment or production interruption, and with high accuracy. Summary of the Invention

[0005] This application provides a robot adaptive calibration method, electronic device, and system based on load changes, which can achieve the effect of online response to load changes, without external equipment or production interruption, and with high-precision adaptive calibration.

[0006] Firstly, this application provides a robot adaptive calibration method based on load variation, comprising:

[0007] In response to changes in the load at the robot's end effector, the robot's joints are controlled to move according to a preset joint space excitation trajectory during the work interval. During the movement, the joint motion data of the target joint, the inertial measurement data of the robot's wrist, and the six-dimensional torque data of the robot's end effector are acquired.

[0008] Based on joint motion data, inertial measurement data, and six-dimensional torque data, the linear load inertial parameter vector of the robot end effector is identified online using the recursive least squares method.

[0009] Control the robot to stop moving and acquire the robot's operating data, including the end effector identifier and the operating status data of the main load-bearing joints;

[0010] Based on the linear load inertia parameter vector and working condition data, the kinematic parameter correction, sensor reference offset, and mapping confidence of the target joint are obtained through the structural deformation mapping model; among them, the sensor reference offset is the same for different target joints.

[0011] Based on the mapping confidence level, the calibration object is determined, and the calibration object is calibrated based on the kinematic parameter correction and / or sensor reference offset.

[0012] In one possible implementation, based on joint motion data, inertial measurement data, and six-dimensional torque data, a recursive least squares method is used to identify the linear load inertial parameter vector of the robot's end effector load online, including:

[0013] Based on joint motion data and inertial measurement data, a regression matrix of the target joint is constructed. The regression matrix represents the linear mapping relationship between the six-dimensional torque data and the linear load inertial parameter vector to be identified.

[0014] When the regression matrix satisfies the preset solvable conditions, the recursive least squares method is used to identify the linear load inertia parameter vector of the robot end effector online based on the regression matrix and the six-dimensional torque data.

[0015] In one possible implementation, the solvable conditions include: the rank of the regression matrix is ​​equal to the dimension of the linear load inertia parameter vector, and the condition number of the regression matrix is ​​less than a preset threshold; the condition number is the ratio of the maximum singular value to the minimum singular value of the regression matrix.

[0016] In one possible implementation, the robot adaptive calibration method based on load variation further includes:

[0017] When the regression matrix does not meet the preset solvability conditions, perform the following operations:

[0018] If the parameters of the excitation trajectory are adjusted, then joint motion data, inertial measurement data, and six-dimensional torque data will be re-acquired based on the adjusted excitation trajectory.

[0019] If the sampling time is extended, joint motion data, inertial measurement data, and six-dimensional torque data will continue to be collected during the extended sampling time.

[0020] Repeat the above steps until the regression matrix satisfies the solvable condition.

[0021] In one possible implementation, determining the calibration object based on the mapping confidence level includes:

[0022] If the mapping confidence is less than the confidence threshold, then the calibration object is determined to be the kinematic parameters;

[0023] If the mapping confidence is greater than or equal to the confidence threshold, then the calibration objects are determined to include kinematic parameters, sensor references, and visual camera extrinsic parameters;

[0024] Among them, the sensor reference characterizes the transformation relationship between the coordinate system of the six-dimensional torque sensor at the robot's end effector and the robot's base coordinate system; the vision camera extrinsic parameter characterizes the transformation relationship between the coordinate system of the vision camera and the robot's base coordinate system.

[0025] In one possible implementation, the kinematic parameters include link length and joint zero position; when the calibration object includes kinematic parameters, the calibration object is calibrated based on the kinematic parameter correction and / or sensor reference offset, including:

[0026] Based on the kinematic parameter correction amount corresponding to the target joint, update the link length and joint zero position of the target joint; wherein, the updated link length is the sum of the current link length and the length correction amount in the kinematic parameter correction amount, and the updated joint zero position is the sum of the current joint zero position and the zero position offset in the kinematic parameter correction amount.

[0027] In one possible implementation, when the calibration object includes kinematic parameters, sensor references, and visual camera extrinsics, the calibration object is calibrated based on kinematic parameter corrections and / or sensor reference offsets, including:

[0028] Acquire visual data collected by a vision camera during the robot's movement;

[0029] Using kinematic parameter corrections and sensor baseline offsets as prior information, and combining joint motion data, inertial measurement data, six-dimensional torque data, and visual data, a weighted least squares problem is constructed that includes joint residuals, IMU pre-integration residuals, torque residuals, and visual residuals.

[0030] The Gauss-Newton method is used to solve the weighted least squares problem, and the optimized kinematic parameter corrections and camera extrinsic parameter corrections are obtained.

[0031] The kinematic parameters of the target joint are corrected based on the optimized kinematic parameter correction amount, the visual camera extrinsic parameters are corrected based on the optimized camera extrinsic parameter correction amount, and the sensor reference is corrected based on the sensor reference offset.

[0032] In one possible implementation, the robot adaptive calibration method based on load variation further includes:

[0033] The corrected kinematic parameters are written into the robot's motion control unit, which is used to perform forward kinematics calculations based on the corrected kinematic parameters.

[0034] The corrected visual camera extrinsic parameters are written into the robot's perception fusion unit, which is used for visual localization based on the corrected visual camera extrinsic parameters.

[0035] The corrected sensor reference is written into the robot's force control unit, which is used to perform force conversion based on the corrected sensor reference.

[0036] In one possible implementation, when the calibration object includes kinematic parameters, sensor references, and visual camera extrinsics, the calibration object is calibrated based on kinematic parameter corrections and / or sensor reference offsets, including:

[0037] Acquire visual data collected by a vision camera during the robot's movement;

[0038] If the quality of the visual data does not meet the quality requirements, the kinematic parameter correction and sensor reference offset will be used as prior information. Combined with joint motion data, inertial measurement data and six-dimensional torque data, a weighted least squares problem including joint residuals, IMU pre-integration residuals and torque residuals will be constructed. The quality requirements include that the number of continuously trackable visual feature points is less than the first threshold, or the number of effective visual frames is less than the second threshold.

[0039] The Gauss-Newton method is used to solve the weighted least squares problem, and the optimized kinematic parameter corrections are obtained.

[0040] The kinematic parameters of the target joint are corrected based on the optimized kinematic parameter correction amount, and the sensor reference is corrected based on the sensor reference offset.

[0041] In one possible implementation, there are multiple target joints. Based on joint motion data, inertial measurement data, and six-dimensional torque data, a recursive least squares method is used to identify the linear load inertial parameter vector of the robot's end effector load online, including:

[0042] Based on the joint motion data, inertial measurement data and six-dimensional torque data of each target joint, the linear load inertial parameter vector estimate of each target joint is independently identified by the recursive least squares method.

[0043] The estimated linear load inertia parameter vectors corresponding to each target joint are fused to obtain the linear load inertia parameter vector of the robot end effector load.

[0044] Secondly, this application provides a robot adaptive calibration device based on load variation, comprising:

[0045] The control module is used to respond to changes in the load at the robot's end effector, control each joint of the robot to move according to a preset joint space excitation trajectory during the work interval, and acquire joint motion data of the target joint, inertial measurement data of the robot's wrist, and six-dimensional torque data of the robot's end effector during the movement.

[0046] The identification module is used to identify the linear load inertial parameter vector of the robot end effector online based on joint motion data, inertial measurement data, and six-dimensional torque data, using the recursive least squares method.

[0047] The control module is also used to control the robot to stop moving and to acquire the robot's working condition data, including the end effector identifier and the operating status data of the main load-bearing joints.

[0048] The mapping module is used to map the kinematic parameter correction, sensor reference offset, and mapping confidence of the target joint based on the linear load inertia parameter vector and working condition data through a structural deformation mapping model; wherein, the sensor reference offset is the same for different target joints.

[0049] The calibration module is used to determine the calibration object based on the mapping confidence level, and to calibrate the calibration object based on the kinematic parameter correction and / or sensor reference offset.

[0050] In one possible implementation, the identification module is specifically used to: construct a regression matrix of the target joint based on joint motion data and inertial measurement data, wherein the regression matrix represents the linear mapping relationship between the six-dimensional torque data and the linear load inertial parameter vector to be identified; when the regression matrix satisfies the preset solvable condition, the recursive least squares method is used to identify the linear load inertial parameter vector of the robot end effector online based on the regression matrix and the six-dimensional torque data.

[0051] In one possible implementation, the solvable conditions include: the rank of the regression matrix is ​​equal to the dimension of the linear load inertia parameter vector, and the condition number of the regression matrix is ​​less than a preset threshold; the condition number is the ratio of the maximum singular value to the minimum singular value of the regression matrix.

[0052] In one possible implementation, the identification module is further configured to: when the regression matrix does not meet the preset solvable conditions, perform the following operations: if the parameters of the excitation trajectory are adjusted, then re-acquire joint motion data, inertial measurement data, and six-dimensional torque data based on the adjusted excitation trajectory; if the sampling time is extended, then continue to acquire joint motion data, inertial measurement data, and six-dimensional torque data within the extended sampling time; repeat the above operations until the regression matrix meets the solvable conditions.

[0053] In one possible implementation, the calibration module is specifically used to: determine the calibration object as kinematic parameters if the mapping confidence is less than a confidence threshold; and determine the calibration object as kinematic parameters, sensor reference, and vision camera extrinsic parameters if the mapping confidence is greater than or equal to the confidence threshold. The sensor reference characterizes the transformation relationship between the coordinate system of the six-dimensional torque sensor at the robot's end effector and the robot's base coordinate system; the vision camera extrinsic parameters characterize the transformation relationship between the coordinate system of the vision camera and the robot's base coordinate system.

[0054] In one possible implementation, the kinematic parameters include link length and joint zero position; when the calibration object includes kinematic parameters, the calibration module is further configured to: update the link length and joint zero position of the target joint according to the kinematic parameter correction amount corresponding to the target joint; wherein, the updated link length is the sum of the current link length and the length correction amount in the kinematic parameter correction amount, and the updated joint zero position is the sum of the current joint zero position and the zero position offset in the kinematic parameter correction amount.

[0055] In one possible implementation, when the calibration objects include kinematic parameters, sensor references, and visual camera extrinsics, the calibration module is further configured to: acquire visual data collected by the robot through the visual camera during its movement; use the kinematic parameter corrections and sensor reference offsets as prior information, and combine them with joint motion data, inertial measurement data, six-dimensional torque data, and visual data to construct a weighted least squares problem including joint residuals, IMU pre-integration residuals, torque residuals, and visual residuals; solve the weighted least squares problem using the Gauss-Newton method to obtain optimized kinematic parameter corrections and camera extrinsic parameter corrections; correct the kinematic parameters of the target joints based on the optimized kinematic parameter corrections, correct the visual camera extrinsics based on the optimized camera extrinsic parameter corrections, and correct the sensor references based on the sensor reference offsets.

[0056] In one possible implementation, the robot adaptive calibration device based on load variation further includes a processing module for writing the corrected kinematic parameters into the robot's motion control unit, which performs forward kinematics calculations based on the corrected kinematic parameters; writing the corrected visual camera extrinsic parameters into the robot's perception fusion unit, which performs visual positioning based on the corrected visual camera extrinsic parameters; and writing the corrected sensor reference into the robot's force control unit, which performs force conversion based on the corrected sensor reference.

[0057] In one possible implementation, when the calibration objects include kinematic parameters, sensor references, and visual camera extrinsics, the calibration module is further configured to: acquire visual data collected by the visual camera during the robot's movement; if the quality of the visual data does not meet the quality requirements, use the kinematic parameter correction and sensor reference offset as prior information, and combine them with joint motion data, inertial measurement data, and six-dimensional torque data to construct a weighted least squares problem including joint residuals, IMU pre-integration residuals, and torque residuals; the quality requirements include that the number of continuously trackable visual feature points is less than a first threshold, or the number of effective visual frames is less than a second threshold; solve the weighted least squares problem using the Gauss-Newton method to obtain the optimized kinematic parameter correction; correct the kinematic parameters of the target joint according to the optimized kinematic parameter correction, and correct the sensor reference according to the sensor reference offset.

[0058] In one possible implementation, there are multiple target joints, and the identification module is further used to: independently identify the estimated value of the linear load inertial parameter vector corresponding to each target joint based on the joint motion data, inertial measurement data and six-dimensional torque data of each target joint using the recursive least squares method; and fuse the estimated values ​​of the linear load inertial parameter vector corresponding to each target joint to obtain the linear load inertial parameter vector of the robot end effector.

[0059] Thirdly, this application provides a robot adaptive calibration system based on load variation, including: an electronic device, and a six-dimensional torque sensor, an inertial measurement unit, and a joint encoder that are communicatively connected to the electronic device;

[0060] A six-dimensional torque sensor is installed at the end effector of the robot to collect six-dimensional torque data at the end effector.

[0061] An inertial measurement unit, located on the robot's wrist, is used to collect inertial measurement data from the robot's wrist.

[0062] A joint encoder is installed on a target joint of a robot to collect joint motion data of the target joint.

[0063] Electronic equipment for performing the first aspect of the load-varying robot adaptive calibration method based on six-dimensional torque data, inertial measurement data, and joint motion data.

[0064] Fourthly, this application provides an electronic device, including: a memory and a processor;

[0065] The memory stores instructions that the computer executes;

[0066] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0067] Fifthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible embodiments of the first aspect.

[0068] In a sixth aspect, this application provides a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0069] The robot adaptive calibration method, electronic device, and system based on load variation provided in this application first respond to changes in the end-effector load by controlling each joint to execute a preset joint space excitation trajectory during work intervals. Simultaneously, joint motion data, wrist inertial measurement data, and end-effector six-dimensional torque data are collected during the movement. This process does not interrupt the production cycle, achieving decoupling of the calibration task from the production task. Based on this, the recursive least squares method is used to process the aforementioned multi-source data online, identifying the linear inertial parameter vector of the load in real time. The mass, center of mass, and inertia information of the load can be quickly obtained without relying on external measurement equipment. Subsequently, after the robot stops moving, working condition data including the end-effector identifier and the operating status of the main load-bearing joints are acquired and input together with the identified linear inertial parameter vector of the load into the structural deformation mapping model, thereby obtaining the kinematic parameter correction amount of the target joint, the globally unique sensor reference offset, and the mapping confidence. The calibration object is adaptively determined through the mapping confidence, and targeted calibration is performed on the selected calibration object based on the kinematic parameter correction amount and / or sensor reference offset. The entire calibration process is triggered by load changes and is executed entirely online. It does not rely on expensive external calibration equipment and does not interrupt normal production operations. At the same time, through multi-source data fusion and adaptive decision-making, it effectively compensates for link deformation, joint flexibility changes and sensor reference drift caused by load, significantly improving the kinematic accuracy and operational reliability of the robot under dynamic load conditions. Attached Figure Description

[0070] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0071] Figure 1 A flowchart illustrating the robot adaptive calibration method based on load variation provided in this application embodiment. Figure 1 ;

[0072] Figure 2 A flowchart illustrating the robot adaptive calibration method based on load variation provided in this application embodiment. Figure 2 ;

[0073] Figure 3 A schematic diagram of the structure of the robot adaptive calibration system based on load variation provided in the embodiments of this application. Figure 1 ;

[0074] Figure 4 A schematic diagram of the structure of the robot adaptive calibration system based on load variation provided in the embodiments of this application. Figure 2 ;

[0075] Figure 5 A schematic diagram of the structure of the robot adaptive calibration device based on load variation provided in the embodiments of this application. Figure 1 ;

[0076] Figure 6 A schematic diagram of the structure of the robot adaptive calibration device based on load variation provided in the embodiments of this application. Figure 2 ;

[0077] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0078] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0079] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application.

[0080] The terms “first,” “second,” etc., used in this application’s specification are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, products, or apparatus.

[0081] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0082] In related technologies, two main approaches are used to address the decrease in kinematic accuracy caused by load variations, but both approaches have inherent drawbacks that are difficult to overcome:

[0083] The first method is kinematic compensation based on joint encoders. The drawback of this approach is that it relies solely on joint angle information as a single data source, completely failing to detect the elastic deformation of the links and changes in joint compliance caused by end-effector torque. When the robot grasps different loads, the gravity and inertial forces at the end effector are transmitted to each joint through the arm, causing varying degrees of bending in the links and additional torsional deformation in the reducer. These errors are completely invisible at the joint encoder level. Therefore, this approach is essentially an open-loop feedforward compensation method with a single compensation dimension and incomplete coverage of error sources. Its accuracy drops sharply under dynamic load conditions, failing to meet the requirements of precision operations.

[0084] The second method is kinematic parameter calibration, typically using a laser tracker for offline calibration. The limitations of this approach are: laser trackers are expensive, complex to operate, and require specialized technicians; the calibration process usually takes tens of minutes or even hours, during which the robot must completely halt production. More importantly, this is an offline, static calibration method; the calibration results only reflect parameter deviations under the current conditions. Once subsequent load changes or prolonged operation leads to wear and tear on the mechanism, the kinematic accuracy will decrease again, requiring re-calibration and disrupting the production cycle. Therefore, this method cannot respond to dynamic load changes online and is ill-suited to the demands of flexible manufacturing for continuous operation and rapid changeover.

[0085] Therefore, there is an urgent need for an adaptive calibration solution that can respond to load changes online, without the need for external equipment or production interruption, and with high accuracy.

[0086] In view of this, the robot adaptive calibration scheme based on load variation provided in this application uses load variation as the trigger condition for the calibration process, embedding the calibration task into the gaps in production operations, thereby achieving dynamic response without interrupting normal production. At the data acquisition level, this scheme breaks through the limitations of traditional methods that rely on a single joint angle, simultaneously collecting three types of data: joint motion, wrist inertial measurement, and six-dimensional end-effector torque, constructing a complete information link from load inertial parameters to structural deformation effects. At the parameter identification level, recursive least squares method is used to process multi-source data online, calculating the load's mass, center of mass, and inertia information in real time, eliminating dependence on expensive external calibration equipment. At the deformation compensation level, the identified load parameters and real-time operating condition data (such as tool identification and joint operating status) are input into the structural deformation mapping model, directly outputting kinematic parameter corrections and sensor reference offsets, realizing an end-to-end mapping between load variation and structural deformation. At the calibration decision level, a confidence mechanism is introduced to adaptively select the calibration range. The entire solution takes "online perception - real-time identification - mapping compensation - adaptive decision-making" as its main technical line, forming a complete closed-loop calibration architecture.

[0087] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0088] First, the execution subject and system architecture of the robot adaptive calibration method based on load variation provided in this application embodiment are briefly described. This robot adaptive calibration method based on load variation is applied to an electronic device, meaning the execution subject is an electronic device, which can be considered the robot's control and processing unit. Specifically, this electronic device can be an industrial computer equipped with a high-performance processor or an embedded intelligent computing platform, for example, having more than 32GB of memory and solid-state storage to store the structural deformation mapping model and various data during the calibration process. Regarding the system architecture, for example, the robot body can adopt a six-axis cooperative robotic arm or a seven-axis redundant robotic arm, with a rated load range of 5 to 20 kg. In terms of sensing system configuration, a six-dimensional torque sensor is rigidly connected to the robot's end effector flange or wrist via an M8 or M12 aviation connector to collect six-dimensional torque data at the end effector. Its measurement range is, for example, ±100 Nf and ±10 Nm, with a sampling rate of at least 500 Hz. An inertial measurement unit (IMU) can be mounted on the robot's wrist via shock-absorbing pads to collect linear acceleration and angular velocity data from the wrist, with a sampling rate between 500 Hz and 1000 Hz and built-in temperature compensation. A multi-turn absolute encoder, for example, is installed on each joint of the robot, with a resolution of at least 19 bits, to collect joint motion data (such as angles and angular velocities). A vision camera (which can be fixedly mounted near the robot's end effector) is connected to electronic devices via a USB 3.0 or GigE interface to collect visual data during movement. The sensors and electronic devices can be connected via an EtherCAT bus or other high-speed communication link, with a communication cycle controlled between 1 and 4 milliseconds to ensure synchronous acquisition and real-time processing of multi-source data. The electronic device executes the robot adaptive calibration method based on load variation provided in this application to jointly process six-dimensional torque data, inertial measurement data, joint motion data and visual data, thereby realizing online robot calibration under load variation.

[0089] Figure 1 A flowchart illustrating the robot adaptive calibration method based on load variation provided in this application embodiment. Figure 1 ,like Figure 1 As shown, this robot adaptive calibration method based on load variation includes:

[0090] S101. In response to changes in the load at the robot's end effector, control each joint of the robot to move according to a preset joint space excitation trajectory during the work interval, and acquire joint motion data of the target joint, inertial measurement data of the robot's wrist, and six-dimensional torque data of the robot's end effector during the movement.

[0091] In this embodiment, the electronic device triggers an adaptive calibration process after detecting a change in the robot's end-effector load. This end-effector load change is not limited to being triggered by tool change signals, object switching records, sudden changes in six-dimensional torque static bias, wrist inertial response changes, or production task switching commands. After the current task ends and the robot enters a reserved idle period, the electronic device sends joint space excitation trajectory commands to each drive joint, enabling the robot to complete the short-term excitation motion required for identification within a time window that does not interfere with the normal production cycle. The joint space excitation trajectory is used to control the movement of each joint of the robot during the task interval to excite the structural deformation characteristics caused by the end-effector load change and provide sufficient observation data for subsequent parameter identification. Specifically, it can employ a sinusoidal frequency sweep excitation trajectory, a segmented multi-frequency sinusoidal excitation trajectory, or a minimum condition number excitation trajectory. It should be understood that if the robot only performs small-range uniform motion, the inertial changes sensed by the six-dimensional torque sensor will be very weak, which will lead to unstable regression of the load inertial parameters. Therefore, the electronic equipment executes pseudo-random sinusoidal frequency sweep excitation trajectory or minimum condition number excitation trajectory during the work interval. For example, the excitation frequency is preferably 0.1Hz to 2.0Hz, the joint amplitude is preferably 2 degrees to 5 degrees, the sampling window is preferably 1.5s to 3.0s, the number of sampling points is not less than 500, and it covers at least the acceleration and deceleration segments of the three main joints.

[0092] In one possible embodiment, the excitation trajectory sets position, velocity, and acceleration boundaries on each target joint, and satisfies the robot's mechanical limits, drive current limits, and end-effector safety envelope requirements, so that the sampling process can be directly deployed on the production site.

[0093] Accordingly, the target joint in this embodiment can be a joint that is significantly affected by load changes, typically including wrist joints, elbow joints, or upstream joints related to the end-load transmission path. The electronic device can pre-configure the target joint set based on the robot's structural parameters and historical load statistics.

[0094] During the excitation motion execution, the joint encoder outputs joint motion data of the target joint according to a unified sampling clock. The joint motion data is used to describe the dynamic motion state of the target joint under the excitation trajectory and serves as the input for online identification of the linear load inertial parameter vector. Specifically, it can include joint angle, joint angular velocity, and joint angular acceleration. The joint angular velocity can be calculated incrementally by the encoder, and the joint angular acceleration can be obtained by filtering and differentiating the velocity sequence. The wrist inertial measurement unit synchronously outputs inertial measurement data, which is used to characterize the inertial motion state of the robot's wrist and participates in the online identification of the linear load inertial parameter vector. Specifically, it includes three-axis acceleration and three-axis angular velocity. The six-dimensional torque sensor synchronously outputs six-dimensional torque data, which reflects the force on the robot's end effector and is an important observation for online identification of the linear load inertial parameter vector. Specifically, it includes six components: three-dimensional force (Fx, Fy, Fz) and three-dimensional torque (Mx, My, Mz). To ensure that multi-source data can be directly used for subsequent identification, the electronic device performs unified timestamp alignment, zero-bias initial reading, outlier removal, and low-pass filtering on the three types of data. The filter cutoff frequency is set to correspond to the highest frequency of the excitation trajectory to ensure that the dynamic components corresponding to the load inertia characteristics are preserved.

[0095] Based on the above analysis, it can be seen that by organizing short-term joint excitation immediately after the load change and simultaneously collecting joint, inertial and force data, a joint observation set covering position, velocity, acceleration and end force response can be formed. This allows the impact of end load change on the structural chain to be explicitly excited and recorded, providing a convergent data foundation for subsequent online identification.

[0096] S102. Based on joint motion data, inertial measurement data, and six-dimensional torque data, the linear load inertial parameter vector of the robot end effector is identified online using the recursive least squares method.

[0097] In this embodiment, after receiving the time-aligned data stream output by S101, the electronic device establishes an identification model based on the robot's end-effector dynamics. It uses six-dimensional torque data as the observation output and inertial excitation terms generated from joint motion data and inertial measurement data as the identification input, thereby estimating the linear load inertial parameter vector of the end-effector load online. The linear load inertial parameter vector characterizes the physical properties of the robot's end-effector load, specifically including the load mass, the first moment of the center of mass, and inertial tensor components. The recursive least squares method is used to identify the linear load inertial parameter vector of the robot end effector online based on the collected joint motion data, inertial measurement data, and six-dimensional torque data. Its recursive update relationship can be expressed as the linear load inertial parameter estimation vector is continuously corrected with each new sampling batch. The gain matrix is ​​jointly determined by the current covariance matrix, the identification input, and the forgetting factor. The covariance matrix is ​​used to characterize the uncertainty of each parameter to be estimated, and the forgetting factor is used to balance the weights of historical data and current data. In one possible embodiment, the forgetting factor is set in the range of 0.95 to 0.995, so that the parameter estimation retains a certain historical stability while maintaining the ability to track recent load changes.

[0098] In practice, the electronic device can expand the inertial contribution of the end effector load to the output of the six-dimensional torque sensor into a form that is easy to process using the recursive least squares method, based on the robot's current posture, wrist linear acceleration, and angular velocity. It then performs a parameter update after reading new input and observation data in each sampling cycle, without requiring a shutdown for batch recalculation. To improve the stability of the identification results, the electronic device can also perform a rationality check on the results.

[0099] Based on the above analysis, it can be seen that by jointly constructing the identification process with joint motion, wrist inertia and end-effector six-dimensional torque, and continuously updating the linear load inertia parameter vector using the recursive least squares method, load parameters reflecting the current tool and workpiece combination state can be obtained in real time without stopping the robot to reconstruct the whole model. This provides direct input for subsequent structural deformation correction and transforms the calibration basis from fixed working condition parameters to a set of parameters that are updated in real time with changes in the working load.

[0100] S103. Control the robot to stop moving and acquire the robot's operating data, including the end effector identifier and the operating status data of the main load-bearing joints.

[0101] After completing the online identification in S102, the electronic device sends a stop command to the robot servo system, causing the robot to enter a stable holding state at the end position of the current excitation trajectory. After the velocity converges to zero or falls below a preset static threshold, operational data is collected. Switching the robot to a stopped state before extracting operational data avoids disturbances caused by motion transients to the mapping results, ensuring that the state variables subsequently input into the structural deformation mapping model correspond to a clear and reproducible load-bearing state.

[0102] Operating condition data describes the robot's current operational state and is input into the structural deformation mapping model along with the linear load inertia parameter vector. Operating condition data includes the end-effector identifier and the operational status data of the main load-bearing joints. The end-effector identifier identifies the type of end-effector currently installed or used by the robot, reflects the source of load changes, and can also distinguish the mass distribution, installation stiffness, and lever arm length of different tools. Specifically, the end-effector identifier can be a tool number, tool category code, or a unique identification code issued by the tool library management module. The main load-bearing joints are those that bear the main load. Structural deformation caused by load changes mainly occurs on joints bearing larger torques, not all joints. The determination of the main load-bearing joints can be based on the robot's structural stress analysis results, the rated torque percentage, or the load transfer path under the current posture. For example, for a six-axis robot, the second joint, third joint, and wrist first joint can be considered as the main load-bearing joint set under different operating conditions. Operational status data describes the current working state of the main load-bearing joints and participates in the mapping as part of the operating condition data. This data can include joint temperature, posture range, joint angle position at rest, recent motion duty cycle, average drive current, or state markers before and after brake engagement.

[0103] In one specific embodiment, the electronic device reads the current end-effector number through the tool management interface and reads the operating status data of the main load-bearing joint through the servo driver, temperature sampling module, and motion status buffer. Then, it processes the operating data according to the input format required by the structural deformation mapping model. If the system is configured with an environmental status input channel, the operating data may further include ambient temperature or mounting base status indicators to reflect additional operating conditions affecting structural deformation.

[0104] More specifically, the operating status data of the main load-bearing joint includes at least joint temperature and attitude range. Joint temperature is used to characterize the stiffness change caused by thermal conditions, reflecting the influence of reducer stiffness, lubrication status, and thermal drift on deformation; the electronic equipment can read the instantaneous temperature value at the current calibration time, or record it as a scalar input after performing a moving average of the temperature samples within the most recent 0.5 to 1.0 seconds. Attitude range is used to characterize the deformation difference under changes in the direction of gravity, and is a discrete state index; the electronic equipment first reads the current angle of the main load-bearing joint, then maps it to a finite range (e.g., range 1, range 2, or range 3) according to a preset segmentation rule, and uses the corresponding range number as a network input to distinguish the lever arm changes caused by different configurations under the same load.

[0105] Based on the above mechanism, this step incorporates tool type and joint load state, which are not related to the linear load inertia parameter vector, into the input of structural deformation mapping. This makes structural deformation estimation no longer dependent on a single load parameter, but conditionally corrected by combining the robot's current thermal state, attitude state, and tool configuration state. This allows for effective differentiation of different deformation compensation results corresponding to the same linear load inertia parameter vector under different working conditions.

[0106] S104. Based on the linear load inertia parameter vector and working condition data, the kinematic parameter correction, sensor reference offset, and mapping confidence of the target joint are obtained by mapping through the structural deformation mapping model; wherein, the sensor reference offset is the same for different target joints.

[0107] This step concatenates the linear load inertia parameter vector identified in S102 with the operating condition data collected in S103 to form a unified query input feature vector. ,in This is the linear load inertia parameter vector output by S102 (containing mass, first moment and inertia tensor components). The main load-bearing joint temperature, The main load-bearing joint posture range. This is the identifier for the end-point tool. The query vector is not a variable to be determined, but rather a combination of load estimation and operating condition information, used as a one-time input to the structural deformation mapping model, enabling the network to directly output the corresponding compensation results based on the current load and operating conditions.

[0108] The structural deformation mapping model can employ a pre-trained lightweight neural network, specifically a three-layer multilayer perceptron structure. The network input layer receives the aforementioned query feature vector, the hidden layers contain fully connected layers and nonlinear activation layers, and the output layer is divided into three branches: a joint correction branch outputs the kinematic parameter correction for the target joint, a sensor offset branch outputs a common sensor reference offset, and a confidence branch outputs the mapping confidence. The kinematic parameter correction is used to compensate for structural deformation caused by load changes, and may include link length correction, joint zero-position offset correction, etc. The sensor reference offset is used to correct the offset relationship between the robot's end effector six-dimensional torque sensor coordinate system and the robot's base coordinate system to compensate for changes in the installation reference caused by load variations, and this offset is the same value for all target joints. The mapping confidence is a scalar between 0 and 1, reflecting the reliability of the kinematic parameter correction and sensor reference offset output by the model under the current load parameters and operating conditions.

[0109] During offline training, standard loading experiments were performed on prototypes with different load parameters, joint temperatures, posture ranges, and tool ID combinations. External high-precision measurement equipment, such as laser trackers, was used to acquire the actual end-effector displacement, joint zero-position deviation, and sensor mounting reference offset. These were then used to calculate kinematic parameter corrections and sensor reference offsets as supervision labels, enabling the network to learn the nonlinear mapping relationship between the working conditions and compensation quantities. During online operation, the electronic device inputs the real-time concatenated query feature vector into the network, directly obtaining the kinematic parameter corrections, sensor reference offsets, and mapping confidence under the current working condition. The sensor reference offset is physically based on the overall mounting reference change of the end-effector coordinate system relative to the robot's base coordinate system, and is not set separately for individual target joints. Therefore, the sensor reference offsets corresponding to different target joints are the same. This design ensures the consistency of sensor reference offsets in multi-joint scenarios, avoiding reference conflicts caused by independent joint corrections. The mapping confidence is directly provided by the network output layer. The confidence increases when the current input feature falls within the central region of the training sample distribution, and decreases when the working condition deviates from the known samples.

[0110] Through the above mechanism, this step inputs the real-time identified load parameters and the working condition data collected in the stopped state into the structural deformation mapping model. It directly generates targeted corrections to the kinematic chain and end sensor references from the current load and working condition, so that the structural deformation and reference drift caused by variable load can be uniformly estimated within the same framework, and provides a quantitative basis for subsequent calibration strategies through mapping confidence.

[0111] S105. Based on the mapping confidence level, determine the calibration object, and calibrate the calibration object based on the kinematic parameter correction and / or sensor reference offset.

[0112] This step adaptively determines the range of this calibration based on the mapping confidence level, and performs calibration on the calibration object based on the corresponding correction amount.

[0113] Mapping confidence is used to characterize the reliability of the output results of the structural deformation mapping model. A high mapping confidence indicates high reliability of the model output, and the electronic device will employ a more comprehensive calibration range, applying available corrections (including kinematic parameter corrections and sensor reference offsets) to the corresponding calibration object. For example, the electronic device will accumulate or replace the kinematic parameter corrections corresponding to the target joint into the current kinematic parameter values, and superimpose a globally uniform sensor reference offset onto the current sensor reference transformation relationship.

[0114] When the mapping confidence is low, it indicates that the reliability of the model output is limited. The electronic device will adopt a more conservative calibration range and apply only some corrections, such as only applying kinematic parameter corrections and temporarily not applying sensor reference offsets, in order to avoid introducing unreliable corrections.

[0115] In another implementation, if the mapping confidence is too low to support any meaningful update, the electronic device can keep all current parameters unchanged and only temporarily store the mapping result in the cache, leaving it for processing after subsequent data accumulation.

[0116] Through the above mechanism, this step adaptively selects the calibration range based on the mapping confidence level, so that the calibration process performs full calibration when the model is reliable, and performs conservative calibration or postpones calibration when the model is unreliable, thereby ensuring the calibration effect while maintaining the reliability of the calibration process.

[0117] In this embodiment, responding to changes in the end-effector load, each joint is controlled to execute a preset joint space excitation trajectory during work breaks. Simultaneously, joint motion data, wrist inertial measurement data, and end-effector six-dimensional torque data are collected during the movement. This process does not interrupt the production cycle, achieving decoupling between calibration and production tasks. Based on this, the recursive least squares method is used to process the aforementioned multi-source data online, identifying the linear inertial parameter vector of the load in real time. The mass, center of mass, and inertia information of the load can be quickly obtained without relying on external measurement equipment. Subsequently, after the robot stops moving, working condition data including the end-effector identifier and the operating status of the main load-bearing joints are acquired and input together with the identified linear inertial parameter vector of the load into the structural deformation mapping model, thereby obtaining the kinematic parameter correction amount of the target joint, the globally unique sensor reference offset, and the mapping confidence. The calibration object is adaptively determined through the mapping confidence, and the selected calibration object is calibrated specifically based on the kinematic parameter correction amount and / or the sensor reference offset. The entire calibration process is triggered by load changes and is executed entirely online. It does not rely on expensive external calibration equipment and does not interrupt normal production operations. At the same time, through multi-source data fusion and adaptive decision-making, it effectively compensates for link deformation, joint flexibility changes and sensor reference drift caused by load, significantly improving the kinematic accuracy and operational reliability of the robot under dynamic load conditions.

[0118] In some embodiments, based on joint motion data, inertial measurement data, and six-dimensional torque data, a recursive least squares method is used to identify the linear load inertial parameter vector of the robot's end effector online. This includes: constructing a regression matrix of the target joint based on joint motion data and inertial measurement data, whereby the regression matrix represents the linear mapping relationship between the six-dimensional torque data and the linear load inertial parameter vector to be identified; and when the regression matrix satisfies a preset solvable condition, using a recursive least squares method, based on the regression matrix and the six-dimensional torque data, to identify the linear load inertial parameter vector of the robot's end effector online.

[0119] The preset solvability conditions are used to characterize whether the regression matrix has sufficient identifiability. Specifically, this can be determined by the consistency between the matrix rank and the dimension of the linear load inertia parameter, as well as whether the condition number is lower than the threshold, so that online identification is only performed when the data is sufficient and the matrix is ​​stable.

[0120] For example, the electronic device constructs a regression matrix in real time using a sliding window and monitors the rank and condition number of the constructed regression matrix to determine whether the currently acquired data sufficiently excites all the inertial parameters of the load. Each row of the regression matrix corresponds to a sampling time, and its number of columns matches the number of linear load inertial parameters to be identified. Within each sliding window (sampling period), the regression matrix is ​​calculated based on joint motion data and inertial measurement data.

[0121] Although the load mass, center of mass, and inertia tensors have different physical meanings, their effects on the six-dimensional torque can all be expressed as a linear regression form. Therefore, the observations within the sliding window can be represented as: In the formula, The observation vector is obtained by stacking the six-dimensional torque data at each moment in chronological order. If necessary, additional load torques calculated from joint currents can also be added. The total dimension of the observation vector is the number of sampling points within the sliding window multiplied by the number of torque components selected in each sampling (usually 6, i.e., three-dimensional force and three-dimensional torque). For the regression matrix, The number of linear load inertia parameters (i.e., dimensions, for example) ); For example, the linear load inertia parameter vector to be identified. ,in For load quality, , , The coordinates of the load's center of mass in the six-dimensional torque sensor coordinate system (the measurement reference point of the sensor fixed at the end flange) are given. to These are the independent components of the inertia tensor. The reason for using... Instead of using directly , , This is because the first moment maintains a linear relationship with the observation, which facilitates stable solution.

[0122] To clarify the regression matrix The correspondence between each element and the physical parameters to be identified (linear load inertia parameters) can... Expand by column. To include Bar scale observations (i.e., observation vectors) The dimension is Taking a sliding window as an example, H can be represented as:

[0123]

[0124] In this matrix, each row corresponds to a scalar observation (e.g., a component of the six-dimensional torque data taken at a certain sampling time), and each column is associated with a specific load physical parameter. Specifically, the first column... Indicates the first In the observation, load mass The contribution coefficient to this observation. ; Columns 2 to 4 ( ) respectively correspond to the first moment Contribution coefficient; Columns 5 to 10 ( These correspond to the six inertia tensor components respectively. The contribution coefficient.

[0125] Observation of the first scalar For example, it and its corresponding coefficient row vector Satisfy linear relationship ,in This represents the 10-dimensional load parameter vector to be identified. It should be noted that the regression matrix... and each of its rows It is not an unknown quantity to be determined, but rather a quantity derived from the first... The known coefficient matrix is ​​obtained by directly calculating the known data (including joint motion data, linear acceleration, angular velocity, angular acceleration, and gravity direction in the sensor coordinate system) at each sampling time. This is achieved by... Each column corresponds one-to-one with a specific physical parameter. This construction method allows electronic equipment to intuitively monitor the linear correlation between columns. When one or more columns show an approximately linear correlation with other columns over a long period, it can be determined which type of load parameter has not been fully excited under the current excitation trajectory, thus providing a quantitative basis for adjusting the excitation trajectory or supplementing data acquisition.

[0126] Further, examine the regression matrix. If the preset solvable conditions are met, and if so, the recursive least squares algorithm is initiated to update the estimated linear load inertia parameters successively according to the sampling time. Specifically, the electronic equipment uses the regression matrix... Each row serves as a coefficient vector, and the observation vector is formed by stacking measured six-dimensional torque data. The corresponding components are used as observations. For each new sampling point, the current prediction residual is calculated, and the parameter estimate from the previous moment is corrected accordingly, causing the estimate to gradually converge as data accumulates. When the difference between two consecutive estimates is less than a preset convergence threshold multiple times, the electronic device determines that the identification process has converged and uses the current estimation result as the final linear load inertia parameter vector. Output.

[0127] This application's embodiments achieve the dual effects of data source decoupling and quality self-checking by limiting the construction of a regression matrix based on joint motion and inertial data and initiating recursive least squares identification only when solvable conditions are met. The regression matrix is ​​independent of the noisy six-dimensional torque data, improving numerical stability; the regression matrix is ​​solved after being filtered by solvable conditions, reducing the impact of ill-conditioned data on the identification results. The combination of these two aspects ensures that online identification is performed only under conditions of sufficient information and numerical robustness, guaranteeing the reliability, accuracy, and real-time performance of load inertial parameter estimation, and providing reliable input for subsequent deformation mapping.

[0128] Specifically, in some embodiments, the solvable conditions include: the rank of the regression matrix is ​​equal to the dimension of the linear load inertia parameter vector, and the condition number of the regression matrix is ​​less than a preset threshold; the condition number is the ratio of the maximum singular value to the minimum singular value of the regression matrix.

[0129] The condition number is used to measure the ill-conditioning of a matrix. When the ratio between the maximum singular value and the minimum singular value is too large, it indicates that the matrix is ​​ill-conditioning and is prone to making the recursive least squares solution sensitive to noise.

[0130] For example, singular value decomposition is performed on the regression matrix to obtain the maximum and minimum singular values, and the condition number is calculated accordingly. The electronic device compares the rank of the regression matrix with the dimension of the linear load inertia parameter vector. When the two are equal and the condition number is less than a preset threshold, it determines that the sampled data within the current sliding window meets the solvability condition, and inputs the regression matrix and the six-dimensional torque data into the recursive least squares method for online identification. The preset threshold can be set to a fixed value (e.g., 1000) based on the robot model, sensor noise level, and load variation range to ensure that the regression matrix has sufficient independent information and numerical stability.

[0131] In terms of working principle, the rank condition ensures that the regression matrix contains sufficient independent observation information, so that the linear load inertia parameter vector can be uniquely or approximately uniquely identified; the condition number constraint is used to suppress ill-conditioned solution problems caused by excessive feature correlation. When the two work together, the recursive least squares method can update the parameter estimates under relatively stable numerical conditions, thereby outputting a reliable linear load inertia parameter vector.

[0132] In this embodiment, by limiting the solvability condition to a rank equal to the parameter dimension and a condition number less than a preset threshold, the uniqueness and stability of parameter identification are dually guaranteed. The rank condition ensures that the regression matrix has full rank, fundamentally eliminating the possibility of multiple solutions for the parameters; the condition number restriction ensures that the regression matrix is ​​not ill-conditioned, making the parameter estimation insensitive to noise from six-dimensional torque measurements and preventing the amplification of small errors. Together, these two constitute a dual verification mechanism for data quality, initiating identification only when the excitation is sufficient and the values ​​are robust, effectively improving the reliability, accuracy, and engineering practicality of load inertia parameter identification.

[0133] Correspondingly, in some embodiments, the robot adaptive calibration method based on load variation further includes: when the regression matrix does not meet the preset solvability conditions, performing the following operations: if the parameters of the excitation trajectory are adjusted, then re-collect joint motion data, inertial measurement data, and six-dimensional torque data based on the adjusted excitation trajectory; if the sampling time is extended, then continue to collect joint motion data, inertial measurement data, and six-dimensional torque data during the extended sampling time; repeat the above operations until the regression matrix meets the solvability conditions.

[0134] For example, when it is detected that the regression matrix does not meet the solvability condition, such as the regression matrix... The condition number being too large or the rank being insufficient is essentially because the current motion trajectory has failed to fully excite the inertial effect of the load across all degrees of freedom, or the sampled data is insufficient to cover all dimensions of the parameter space. When the regression matrix is ​​not of rank... Irreversible, non-unique parametric solutions; when the condition number is too large, In cases of pathological conditions, minute measurement noise can be amplified significantly, leading to severe distortion in parameter estimation. This embodiment provides corresponding compensation mechanisms for both scenarios. If the rank is insufficient, it typically indicates a lack of motion excitation in certain directions. In this case, the parameters of the excitation trajectory should be adjusted to introduce richer joint motion modes (e.g., increasing the swing amplitude of the main load-bearing joint or changing the direction of motion) so that the contribution coefficients of the load's center of mass shift and inertia tensor in all directions can be effectively excited. If the condition number is too large but the rank is full, it indicates a strong linear correlation between the contribution coefficients of each parameter. In this case, the diversity of data samples can be increased by extending the sampling time to reduce the interference of random noise on parameter estimation, or the numerical characteristics of the regression matrix can be improved from the source by switching to the excitation trajectory with the minimum condition number.

[0135] If the parameters of the excitation trajectory need to be adjusted, the trajectory amplitude, frequency, or waveform type can be modified (e.g., switching from pseudo-random frequency sweep to a minimum condition number trajectory) to give the excitation trajectory a more complete excitation characteristic for the joint angular velocity, joint angular acceleration, and attitude changes of the target joint. Under the updated trajectory, the motion of each joint is recontrolled to simultaneously collect joint motion data, inertial measurement data, and six-dimensional torque data. If the sampling time needs to be extended, while keeping the current excitation trajectory unchanged, the above data is collected within the extended sampling time window (e.g., from 500 sampling points to 800), allowing the new samples to supplement the original dataset and expand the effective observation range of the regression matrix. After each re-acquisition or continued acquisition, the regression matrix is ​​reconstructed and its solvability is checked again until the rank of the regression matrix reaches the dimension of the linear load inertial parameter vector and the condition number is below a preset threshold. At this point, data acquisition ends, and recursive least squares identification begins.

[0136] In other words, the electronic device is only allowed to proceed to the subsequent recursive least squares identification step when both of the above conditions (the rank of the regression matrix reaches the dimension of the linear load inertia parameter vector and the number of conditions is lower than the preset threshold) are met simultaneously. If either condition is not met, it indicates that the current excitation trajectory has failed to fully excite all the load inertia parameters or that the regression matrix is ​​seriously ill. The electronic device will refuse to perform parameter updates and return to the data acquisition stage to adjust the excitation trajectory or extend the sampling time until the data quality meets the solvable conditions.

[0137] In this embodiment, two remedial measures—adjusting the excitation trajectory parameters and extending the sampling time—provide an automatic processing path for regression matrices that are not rank or ill-conditioned. The electronic device iteratively executes a closed-loop control process of "judgment-adjustment-reacquisition-rejudgment" until the data quality meets the solvable conditions. This mechanism ensures the uniqueness and numerical stability of the load inertia parameter identification from the source, avoids invalid calculations caused by insufficient excitation or ill-conditioned regression matrices, and significantly improves the reliability, robustness, and automation level of the adaptive calibration process.

[0138] In some embodiments, determining the calibration object based on the mapping confidence level includes: if the mapping confidence level is less than a confidence threshold, then the calibration object is determined to be kinematic parameters; if the mapping confidence level is greater than or equal to the confidence threshold, then the calibration object is determined to include kinematic parameters, sensor reference, and visual camera extrinsic parameters; wherein, the sensor reference characterizes the transformation relationship between the coordinate system of the six-dimensional torque sensor at the robot end effector and the robot base coordinate system; and the visual camera extrinsic parameters characterize the transformation relationship between the coordinate system of the visual camera and the robot base coordinate system.

[0139] The confidence threshold can be obtained from historical calibration data and stored in the parameter table of the electronic device. Kinematic parameters include, for example, the robot's link length and joint zero position; sensor references include the translation and rotation of the six-dimensional torque sensor mounting coordinate system relative to the base coordinate system; and vision camera extrinsic parameters include the pose transformation matrix of the vision camera optical axis coordinate system relative to the base coordinate system. The sensor references and vision camera extrinsic parameters can be obtained by pose calculation using the same calibration plate or reference fixture in a stationary state of the robot. The pose matrix can be represented by a combination of quaternions and translation vectors. In practical applications, other commonly used matrix forms can also be selected, and this embodiment does not limit this.

[0140] In practical implementation, after obtaining the mapping confidence level, the electronic device compares it with a preset confidence threshold. Taking a confidence threshold of 0.8 as an example, when the confidence level is less than 0.8, it indicates that the model lacks sufficient confidence in the current mapping result. In this case, the electronic device defines the calibration object as containing only kinematic parameters. In this mode, for example, only the kinematic parameter corrections obtained from the mapping (such as link length corrections and joint zero-position offsets) are used to correct the target joint, while the use of sensor reference offsets to update the coordinate system transformation of the six-dimensional torque sensor is prohibited, and the extrinsic parameters of the vision camera are not calibrated. This conservative strategy avoids introducing incorrect extrinsic parameter corrections when the model is unreliable, ensuring the safe operation of the system. When the confidence level is greater than or equal to 0.8, it indicates that the model has high confidence in the mapping result. In this case, the electronic device defines the calibration object as a complete set including kinematic parameters, sensor references, and vision camera extrinsic parameters. In this mode, multi-source sensor data can be further integrated for joint optimization, outputting more accurate corrections and achieving comprehensive high-precision calibration.

[0141] In this embodiment, by adaptively shrinking or expanding the calibration range based on the mapping confidence, it can perform full compensation to obtain optimal accuracy when the model is reliable, and degenerate into conservative calibration to ensure system safety when the model is unreliable, thus achieving an effective balance between accuracy and robustness.

[0142] Furthermore, in some embodiments, the kinematic parameters include link length and joint zero position; when the calibration object includes kinematic parameters, the calibration object is calibrated based on the kinematic parameter correction amount and / or sensor reference offset, specifically including: updating the link length and joint zero position of the target joint according to the kinematic parameter correction amount corresponding to the target joint; wherein, the updated link length is the sum of the current link length and the length correction amount in the kinematic parameter correction amount, and the updated joint zero position is the sum of the current joint zero position and the zero position offset in the kinematic parameter correction amount.

[0143] In this context, the link length describes the equivalent distance between adjacent joint axes, and the joint zero point describes the deviation between the joint encoder zero point and the actual mechanical zero point. Since the elastic deformation of the link caused by load changes alters the equivalent value of the link length, and changes in joint compliance change the joint zero point, these two factors are transmitted step-by-step along the kinematic chain and amplified cumulatively at the end. Therefore, this is the physical basis for the requirement in this embodiment to correct the joint zero point and link length online under variable load conditions.

[0144] The length correction corresponds to "how much the connecting rod lengthens or shortens under load", and is used to compensate for the elastic deformation of the connecting rod caused by load changes; the zero offset corresponds to "how much angular offset the compliant deformation is equivalent to", and is used to compensate for the zero drift of the joint encoder or the torsional deformation of the transmission components.

[0145] For example, in a specific implementation, after obtaining the kinematic parameter correction amount corresponding to the target joint, the electronic device reads the link length currently stored in the kinematic parameter table. and joint zero position These two values ​​could be the factory-specified values ​​or the results of the last calibration update. Following an incremental update strategy, the electronic device will adjust the length correction amount. The updated link length is obtained by superimposing it onto the current link length. Simultaneously, zero offset will be applied. Superimpose onto the current joint zero position to obtain the updated joint zero position. The updated kinematic parameters are stored in the kinematic parameter cache for subsequent forward kinematics calculations, trajectory planning, and servo control. Through this cumulative update method, each calibration performs incremental corrections based on the current parameters, applicable to both the initial calibration (where the current value is the factory nominal value) and subsequent calibrations (where the current value is the result of the previous calibration). This enables continuous tracking and compensation for structural deformation and zero-position drift caused by load changes.

[0146] In this embodiment, the link length and joint zero position are updated cumulatively according to the length correction and zero position offset, respectively, ensuring that the geometric model of the target joint remains consistent with the current mechanical state, and the kinematic solution results are closer to the actual pose. This strategy achieves targeted compensation for structural deformation and zero position drift caused by load, possesses continuous optimization capabilities, and adapts to dynamic operating environments with frequent load changes. It is computationally simple, facilitates online real-time execution, and meets the requirements of low latency and high reliability while ensuring calibration effectiveness.

[0147] In some embodiments, when the calibration object includes kinematic parameters, sensor references, and visual camera extrinsic parameters, the calibration object is calibrated based on the kinematic parameter correction and / or sensor reference offset. Specifically, this includes: acquiring visual data collected by the robot through the visual camera during its movement; using the kinematic parameter correction and sensor reference offset as prior information, and combining joint motion data, inertial measurement data, six-dimensional torque data, and visual data, constructing a weighted least squares problem including joint residuals, IMU pre-integration residuals, torque residuals, and visual residuals; solving the weighted least squares problem using the Gauss-Newton method to obtain optimized kinematic parameter corrections and camera extrinsic parameter corrections; correcting the kinematic parameters of the target joint based on the optimized kinematic parameter corrections, correcting the visual camera extrinsic parameters based on the optimized camera extrinsic parameter corrections, and correcting the sensor reference based on the sensor reference offset.

[0148] The visual camera extrinsic parameter correction is used to characterize the optimization increment of the visual camera pose. Joint residuals refer to the deviation between the theoretical end-effector pose calculated using forward kinematics based on the corrected kinematic parameters (including updated link lengths and joint zero positions) and the nominal end-effector pose calculated using the nominal kinematic model based on the measured angles of the joint encoder. This residual constrains the correction of the kinematic parameters, enabling the corrected model to more accurately describe the robot's true geometric relationships. IMU pre-integration residuals are also included. The acceleration and angular velocity measured by the IMU between two keyframes are pre-integrated to obtain the relative pose and velocity increment measurements of the end effector coordinate system from the start frame to the end frame. Simultaneously, based on robot joint encoder data and the corrected kinematic model, the theoretical values ​​of the relative pose and velocity increment of the end effector coordinate system within the same time period are calculated. The pre-integration residual is the deviation between the pre-integrated measurements and the kinematically calculated values, specifically including three components: rotational residual, velocity residual, and position residual. This IMU pre-integration residual utilizes high-frequency motion information measured by the IMU to constrain and correct the pose estimation results of the pure kinematic model under dynamic conditions, and is a key step in the tightly coupled optimization of vision-inertia-kinematics. The torque residual refers to the torque calculated based on the identified load inertia parameter vector. The deviation between the theoretical six-dimensional torque predicted by the inverse dynamics model and the measured six-dimensional torque actually collected by the six-dimensional torque sensor, along with the joint motion data at the current moment, is considered. The torque residual is used to constrain the correction of kinematic parameters (link length, joint zero position) and camera extrinsic parameters. The visual residual refers to the deviation between the theoretical pixel coordinates of feature points in three-dimensional space and the observed pixel coordinates detected by the feature extraction algorithm in the actual image. The theoretical pixel coordinates are obtained by transforming the robot's base coordinate system to the camera coordinate system based on the currently estimated camera extrinsic parameters, and then projecting them onto the image plane through the camera intrinsic parameter model. The visual residual is used to constrain the estimation of camera extrinsic parameters and end effector pose.

[0149] For example, the electronic device first reads the joint motion data, inertial measurement data, and six-dimensional torque data collected and stored in stage S101, as well as the visual data collected by the robot through the vision camera during its movement. Then, the electronic device uses the kinematic parameter corrections and sensor reference offsets as prior information (i.e., as initial guesses and constraint boundaries for the optimization problem) to construct a weighted least squares problem together with the aforementioned multi-source data. This problem contains four types of residual terms: joint residuals... IMU pre-integration residuals Torque residual and visual residual Each residual term is assigned a weight according to the noise level of the corresponding sensor. to This forms the total cost function:

[0150]

[0151] The electronic device employs the Gauss-Newton method to iteratively solve the aforementioned weighted least squares problem. In each iteration, the residual function is linearized, the increment is calculated, and the state variables are updated until convergence is achieved, yielding the optimal fine-tuning result. This optimization process outputs the optimized kinematic parameter corrections and camera extrinsic parameter corrections. In this context, the sensor reference offset is used as prior information in the optimization but no new correction is generated. That is, the sensor reference is directly updated using the sensor reference offset output by S104.

[0152] Finally, the electronic device corrects the kinematic parameters of the target joint (link length and joint zero position) based on the optimized kinematic parameter correction; and corrects the visual camera extrinsic parameters based on the optimized camera extrinsic parameter correction, denoted as... , For visual camera extrinsics, The corrected extrinsic parameters of the visual camera are used; and the reference transformation relationship of the six-dimensional torque sensor is corrected based on the sensor reference offset output by S104.

[0153] This embodiment constructs a weighted least squares problem by fusing four types of data: joint, IMU, six-dimensional torque, and visual data. It then employs the Gauss-Newton method for iterative solution, jointly tightening the open-loop compensation obtained from structural deformation mapping with real-world observations during operation. This mechanism effectively suppresses measurement noise and modeling errors from a single sensor, making the calibration results of kinematic parameters and visual camera extrinsics more accurate and reliable, and improving the overall accuracy and robustness of multi-source information fusion calibration.

[0154] Considering that the quality of visual data is often affected by factors such as ambient lighting, texture richness, and motion blur during the joint optimization and calibration of robots, there may be insufficient continuously trackable feature points or an inadequate number of effective visual frames. In such cases, forcibly introducing visual residuals will not only fail to provide effective constraints but may also introduce incorrect optimization directions, leading to distorted camera extrinsic parameter estimation or even calibration failure.

[0155] To address this technical problem, this embodiment provides a joint optimization calibration scheme under visual degradation conditions. In some embodiments, when the calibration object includes kinematic parameters, sensor references, and visual camera extrinsic parameters, the calibration object is calibrated based on the kinematic parameter correction and / or sensor reference offset. This includes: acquiring visual data collected by the robot through the visual camera during movement; if the data quality of the visual data does not meet the quality requirements, the kinematic parameter correction and sensor reference offset are used as prior information, combined with joint motion data, inertial measurement data, and six-dimensional torque data, to construct a weighted least squares problem including joint residuals, IMU pre-integration residuals, and torque residuals; the quality requirements include the number of continuously trackable visual feature points being less than a first threshold, or the number of effective visual frames being less than a second threshold; solving the weighted least squares problem using the Gauss-Newton method to obtain the optimized kinematic parameter correction; correcting the kinematic parameters of the target joint according to the optimized kinematic parameter correction, and correcting the sensor reference according to the sensor reference offset.

[0156] For example, the electronic device performs a quality assessment on the acquired visual data. Specifically, the electronic device detects the number of continuously trackable visual feature points and the number of effective visual frames. If the number of continuously trackable visual feature points is less than a first threshold (e.g., 20), or the number of effective visual frames is less than a second threshold (e.g., 3 frames), the current visual data is determined to not meet the quality requirements. In this case, the electronic device uses the kinematic parameter corrections and sensor reference offsets output by the structural deformation mapping model as prior information. Combined with the joint motion data, inertial measurement data, and six-dimensional torque data collected and stored in stage S101, it constructs a weighted least squares problem that only includes joint residuals, IMU pre-integration residuals, and torque residuals, without introducing unreliable visual residuals. The total cost function of this optimization problem is:

[0157]

[0158] The electronic device uses the Gauss-Newton method to iteratively solve the aforementioned weighted least squares problem, obtaining optimized kinematic parameter corrections. Subsequently, the kinematic parameters of the target joint (link length and joint zero position) are corrected based on these optimized kinematic parameter corrections, and the reference transformation relationship of the six-dimensional torque sensor is updated based on the sensor reference offset. During this process, because the visual data quality does not meet requirements, the camera extrinsic parameters are not updated temporarily, maintaining their current values ​​to avoid erroneous calibration due to visual degradation.

[0159] Through the above mechanism, this embodiment automatically downgrades to optimized calibration based solely on joint, IMU, and torque data when visual data quality is insufficient. This ensures the stability of the calibration process while effectively preventing visual degradation from contaminating the overall calibration results.

[0160] Furthermore, in some embodiments, the robot adaptive calibration method based on load variation further includes: writing the corrected kinematic parameters into the robot's motion control unit, which is used to perform forward kinematics calculation based on the corrected kinematic parameters; writing the corrected visual camera extrinsic parameters into the robot's perception fusion unit, which is used to perform visual positioning based on the corrected visual camera extrinsic parameters; and writing the corrected sensor reference into the robot's force control unit, which is used to perform force conversion based on the corrected sensor reference.

[0161] For example, after calibration, the electronic device writes the corrected kinematic parameters into the robot's motion control unit. The corrected kinematic parameters include the link lengths and joint zero points of each target joint. The motion control unit is responsible for performing forward kinematics calculations based on these parameters, that is, calculating the actual pose of the robot's end effector in Cartesian space based on the angle values ​​fed back from the encoders of each joint. By writing the corrected kinematic parameters, the motion control unit can eliminate calculation errors caused by link elastic deformation and zero-point drift due to load changes, making the forward kinematics calculation results closer to the true end effector pose.

[0162] Simultaneously, the electronic device writes the corrected visual camera extrinsic parameters into the robot's perception fusion unit. The perception fusion unit is responsible for visual localization based on these extrinsic parameters, such as converting the coordinates of detected feature points in the image into three-dimensional spatial coordinates in the robot's base frame, or for error calculation in visual servo control. By writing the corrected camera extrinsic parameters, the perception fusion unit can compensate for camera pose drift caused by load changes or installation reference offsets, ensuring the accuracy of visual localization.

[0163] Furthermore, the electronic equipment writes the corrected sensor reference into the robot's force control unit. The force control unit is responsible for force conversion based on this reference, that is, converting the six-dimensional torque directly measured by the sensor (expressed in the sensor coordinate system) into equivalent force and torque in the robot's base system or tool system, for force feedback control in tasks such as force-controlled assembly and grinding. By writing the corrected sensor reference, the force control unit can eliminate sensor mounting reference offset caused by load deformation, ensuring the accuracy of force conversion.

[0164] Through the above parameter injection mechanism, this embodiment deploys the calibration results to the three functional units of motion control, perception fusion and force control, so that the robot's subsequent forward kinematics calculation, visual positioning and force conversion are all performed based on the updated parameters, and the system accuracy after load change is fully restored from multiple dimensions.

[0165] In some embodiments, there are multiple target joints. Based on joint motion data, inertial measurement data, and six-dimensional torque data, the linear load inertial parameter vector of the robot's end effector is identified online using the recursive least squares method. This includes: independently identifying the estimated value of the linear load inertial parameter vector corresponding to each target joint using the recursive least squares method based on the joint motion data, inertial measurement data, and six-dimensional torque data of each target joint; and fusing the estimated values ​​of the linear load inertial parameter vector corresponding to each target joint to obtain the linear load inertial parameter vector of the robot's end effector.

[0166] For example, a robot has multiple target joints (e.g., the first to sixth joints of a six-axis industrial robot), each equipped with a joint encoder capable of collecting its own joint motion data. In a specific implementation, a regression model corresponding to the motion state of each target joint is constructed. The joint motion data, inertial measurement data, and six-dimensional torque data at the sampling time are input into a recursive least squares method for online updating, yielding estimates of the linear load inertial parameter vector for each joint. This is because the load inertial parameter vector to be identified in the observation equations of different target joints... This describes the same end-effector load (mass, center of mass, inertia). Theoretically, the estimated linear load inertia parameter vectors obtained by independently solving each target joint should tend to be consistent. However, due to differences in measurement noise, modeling errors, and data excitation levels among the target joints, there will be slight deviations between the actual estimated values. The estimated values ​​can be aligned according to time synchronization, and then fused based on residual magnitude, data confidence level, or weighting coefficients to obtain a unified linear load inertia parameter vector for the robot's end-effector load. The fusion process can employ weighted averaging or minimum variance estimation based on the covariance matrix to ensure consistent output of the independent identification results of multiple target joints within the same parameter space.

[0167] During operation, when the load on the robot's end effector changes, the excitation motion of each target joint synchronously introduces a load-related dynamic response. The recursive least squares method can continuously update the parameter estimates based on the newly acquired data. The parameter vectors independently identified by each target joint are then fused to form the final output. Since the fusion result integrates observation information from multiple target joints under different postures and stress conditions, it can serve as a unified load representation for subsequent kinematic corrections, sensor baseline corrections, and operational condition mapping.

[0168] In this embodiment, the linear load inertia parameter vector of the robot end effector is obtained by the collaborative identification of multiple target joints. This allows the estimation results of other target joints to compensate for the limitations of a single target joint due to attitude or local noise interference, thereby improving the stability and consistency of parameter identification and providing more reliable load parameter input for subsequent calibration.

[0169] Based on the foregoing embodiments, combined with Figure 2 The adaptive calibration method for robots based on load variation is further described in detail.

[0170] Figure 2 A flowchart illustrating the robot adaptive calibration method based on load variation provided in this application embodiment. Figure 2 .like Figure 2 As shown, this robot adaptive calibration method based on load variation specifically includes the following steps:

[0171] S201. In response to changes in the load at the robot's end effector, control each joint of the robot to move according to a preset joint space excitation trajectory during the work interval.

[0172] That is, after the load is switched, the robot enters the micro-perturbation calibration stage, where electronic devices control the robot's joints to execute low-amplitude excitation trajectories to ensure that the safety of the object being held is not affected.

[0173] S202. During the movement, acquire joint motion data of the target joint, inertial measurement data of the robot wrist, six-dimensional torque data of the robot end effector, and visual data collected by the vision camera.

[0174] It synchronously receives data reported by the joint encoder, six-dimensional torque sensor, vision camera and IMU.

[0175] S203. Based on joint motion data and inertial measurement data, construct the regression matrix of the target joint.

[0176] For example, when constructing the regression matrix, the electronic device first calculates the pose, angular velocity, and angular acceleration of the end effector in the sensor coordinate system using forward kinematics based on the joint angle, angular velocity, and angular acceleration at the current sampling moment. Simultaneously, the linear acceleration and angular velocity measured by the IMU are transformed to the sensor coordinate system, and the true acceleration of the load's center of mass is obtained by combining this with the direction of gravity. Based on this, the contribution coefficients of the load mass, first moment, and inertia tensor components to the components of the current six-dimensional torque are analytically calculated using the Newton-Euler inverse dynamics formula. These coefficients are then arranged in order to form one row of the regression matrix. This process is repeated for each sampling moment within the sliding window, and the coefficient vectors of all rows are stacked to obtain the complete regression matrix.

[0177] S204. Determine whether the regression matrix satisfies the preset solvable conditions.

[0178] Specifically, the solvable conditions include: the rank of the regression matrix is ​​equal to the dimension of the linear load inertia parameter vector, and the condition number of the regression matrix is ​​less than a preset threshold; the condition number is the ratio of the maximum singular value to the minimum singular value of the regression matrix.

[0179] If satisfied, proceed with S206;

[0180] If not satisfied, then execute S205.

[0181] S205. Adjust the parameters of the excitation trajectory or extend the sampling time.

[0182] Then execute S202 until the regression matrix meets the preset solvable conditions.

[0183] S206. Using the recursive least squares method, the linear load inertia parameter vector of the robot end effector is identified online based on the regression matrix and six-dimensional torque data.

[0184] S207. Control the robot to stop moving and acquire the robot's operating data.

[0185] The operating data includes end-effector identification and the operating status data of the main load-bearing joint.

[0186] S208. Based on the linear load inertia parameter vector and working condition data, the kinematic parameter correction, sensor reference offset, and mapping confidence of the target joint are obtained by mapping through the structural deformation mapping model.

[0187] S209. Determine the calibration object based on the mapping confidence level.

[0188] If the mapping confidence is less than the confidence threshold, the calibration object is determined to be kinematic parameters, and S210 is executed;

[0189] If the mapping confidence is greater than or equal to the confidence threshold, the calibration objects are determined to include kinematic parameters, sensor references, and visual camera extrinsics, and S211 is executed.

[0190] S210. Update the link length and joint zero position of the target joint based on the kinematic parameter correction amount corresponding to the target joint.

[0191] S211. Based on joint motion data, inertial measurement data, six-dimensional torque data, and visual data, multi-source joint optimization is performed to obtain the optimized kinematic parameter correction and the optimized camera extrinsic parameter correction.

[0192] Specifically, the kinematic parameter correction and sensor reference offset are used as prior information. Combined with joint motion data, inertial measurement data, six-dimensional torque data and visual data in S202, a weighted least squares problem including joint residuals, IMU pre-integration residuals, torque residuals and visual residuals is constructed. The Gauss-Newton method is used to solve the weighted least squares problem to obtain the optimized kinematic parameter correction and camera extrinsic parameter correction.

[0193] S212. Correct the kinematic parameters of the target joint according to the optimized kinematic parameter correction amount, correct the visual camera extrinsic parameters according to the optimized camera extrinsic parameter correction amount, and correct the sensor reference according to the sensor reference offset.

[0194] S213. Inject the corrected kinematic parameters, visual camera extrinsic parameters, and sensor references into the corresponding functional units.

[0195] Specifically, the corrected kinematic parameters are written into the robot's motion control unit, the corrected visual camera extrinsic parameters are written into the robot's perception fusion unit, and the corrected sensor references are written into the robot's force control unit.

[0196] S214. Perform closed-loop verification and rollback on the robot after calibration injection.

[0197] After parameter injection, the electronic device drives the robot to repeatedly perform positioning verification on multiple preset test poses to check the effectiveness of the calibration parameters. Specifically, the electronic device selects 5 to 10 test poses distributed in the workspace, repeats the operation on each pose at least 10 times, and calculates the end-effector position error after each successful positioning. ,in For the first The test pose is the first The position obtained by solving in the end coordinate system at the next arrival. This is the average position of all repeated results for this test pose. Simultaneously, the sensor fusion unit calculates the average reprojection error. It is used to check the visual geometric consistency after the camera extrinsic parameters are updated.

[0198] The electronic device compares the verification results with preset indicators: when the mean of the repeatability error is less than 0.1 mm, the three-fold standard deviation is less than 0.3 mm, and the mean reprojection error is less than 1.0 pixel, the calibration parameters are deemed valid, and this group of parameters is registered as a verified parameter group. If any indicator fails to meet the requirements, the current calibration parameters are deemed invalid, the electronic device automatically rolls back to the previous group of verified parameters, and returns to S201 to re-execute the adaptive calibration process until a calibration result that meets the accuracy requirements is obtained. Through the above verification and rollback mechanism, this step ensures the reliability and safety of the calibration parameters injected into the system, avoiding the risks of subsequent operations due to calibration failure or insufficient accuracy.

[0199] The above describes the robot adaptive calibration method based on load variation. Next, we will describe the robot adaptive calibration system based on load variation.

[0200] Figure 3 A schematic diagram of the structure of the robot adaptive calibration system based on load variation provided in the embodiments of this application. Figure 1,like Figure 3 As shown, the robot adaptive calibration system 30 based on load variation includes: an electronic device 31, and a six-dimensional torque sensor 32, an inertial measurement unit 33, and a joint encoder 34 communicatively connected to the electronic device 31; wherein, the six-dimensional torque sensor 32 is disposed at the end of the robot and is used to collect six-dimensional torque data of the end of the robot; the inertial measurement unit 33 is disposed at the wrist of the robot and is used to collect inertial measurement data of the wrist of the robot; the joint encoder 34 is disposed at the target joint on the robot and is used to collect joint motion data of the target joint; the electronic device 31 is used to execute the aforementioned robot adaptive calibration method based on load variation based on the six-dimensional torque data, inertial measurement data, and joint motion data.

[0201] By placing a six-dimensional torque sensor at the robot's end effector, changes in force and torque caused by load variations can be directly sensed. Placing an inertial measurement unit (IMU) on the robot's wrist allows for simultaneous acquisition of attitude, acceleration, and angular velocity changes. Combined with joint motion data output from the joint encoder at the target joint, the electronic equipment can jointly analyze the end effector load state, wrist inertial response, and joint motion behavior, thereby performing more accurate adaptive calibration based on load changes. This allows for timely correction of relevant parameters when tools are changed or workpiece mass and center of gravity change, enabling the robot to maintain high positioning accuracy and operational stability under varying load conditions. This reduces the need for downtime recalibration, making it more suitable for continuous production scenarios such as flexible assembly, material handling, and human-robot collaboration.

[0202] like Figure 4 As shown, in some embodiments, the robot adaptive calibration system 30 based on load variation further includes a vision camera 35 for acquiring visual data during robot movement. The electronic device 31 may also include a perception fusion unit, a motion control unit, and a force control unit. The perception fusion unit performs visual positioning based on corrected visual camera extrinsic parameters; the force control unit performs force conversion based on corrected sensor references; and the motion control unit performs forward kinematics calculations based on corrected kinematic parameters.

[0203] Figure 5 A schematic diagram of the structure of the robot adaptive calibration device based on load variation provided in the embodiments of this application. Figure 1 ,like Figure 5 As shown, the robot adaptive calibration device 50 based on load variation includes: a control module 51, an identification module 52, a mapping module 53, and a calibration module 54. Wherein:

[0204] The control module 51 is used to respond to changes in the load at the robot end effector, control each joint of the robot to move according to a preset joint space excitation trajectory during the work interval, and acquire joint motion data of the target joint, inertial measurement data of the robot's wrist, and six-dimensional torque data of the robot end effector during the movement.

[0205] The identification module 52 is used to identify the linear load inertial parameter vector of the robot end effector online based on joint motion data, inertial measurement data and six-dimensional torque data, using the recursive least squares method.

[0206] The control module 51 is also used to control the robot to stop moving and to acquire the robot's working condition data, including the end tool identifier and the running status data of the main load-bearing joints.

[0207] The mapping module 53 is used to map the kinematic parameter correction, sensor reference offset, and mapping confidence of the target joint based on the linear load inertia parameter vector and working condition data through the structural deformation mapping model; wherein, the sensor reference offset is the same for different target joints.

[0208] The calibration module 54 is used to determine the calibration object based on the mapping confidence level, and to calibrate the calibration object based on the kinematic parameter correction and / or sensor reference offset.

[0209] In one possible implementation, the identification module 52 is specifically used to: construct a regression matrix of the target joint based on joint motion data and inertial measurement data, wherein the regression matrix represents the linear mapping relationship between the six-dimensional torque data and the linear load inertial parameter vector to be identified; when the regression matrix satisfies the preset solvable condition, the recursive least squares method is used to identify the linear load inertial parameter vector of the robot end load online according to the regression matrix and the six-dimensional torque data.

[0210] In one possible implementation, the solvable conditions include: the rank of the regression matrix is ​​equal to the dimension of the linear load inertia parameter vector, and the condition number of the regression matrix is ​​less than a preset threshold; the condition number is the ratio of the maximum singular value to the minimum singular value of the regression matrix.

[0211] In one possible implementation, the identification module 52 is further configured to: when the regression matrix does not meet the preset solvable conditions, perform the following operations: if the parameters of the excitation trajectory are adjusted, then re-acquire joint motion data, inertial measurement data, and six-dimensional torque data based on the adjusted excitation trajectory; if the sampling time is extended, then continue to acquire joint motion data, inertial measurement data, and six-dimensional torque data during the extended sampling time; repeat the above operations until the regression matrix meets the solvable conditions.

[0212] In one possible implementation, the calibration module 54 is specifically used to: if the mapping confidence is less than the confidence threshold, determine that the calibration object is kinematic parameters; if the mapping confidence is greater than or equal to the confidence threshold, determine that the calibration object includes kinematic parameters, sensor reference, and visual camera extrinsic parameters; wherein, the sensor reference characterizes the transformation relationship between the coordinate system of the six-dimensional torque sensor at the robot end and the robot base coordinate system; and the visual camera extrinsic parameters characterize the transformation relationship between the coordinate system of the visual camera and the robot base coordinate system.

[0213] In one possible implementation, the kinematic parameters include link length and joint zero position; when the calibration object includes kinematic parameters, the calibration module 54 is further configured to: update the link length and joint zero position of the target joint according to the kinematic parameter correction amount corresponding to the target joint; wherein, the updated link length is the sum of the current link length and the length correction amount in the kinematic parameter correction amount, and the updated joint zero position is the sum of the current joint zero position and the zero position offset in the kinematic parameter correction amount.

[0214] In one possible implementation, when the calibration objects include kinematic parameters, sensor references, and visual camera extrinsics, the calibration module 54 is further configured to: acquire visual data collected by the robot through the visual camera during its movement; use the kinematic parameter correction and sensor reference offset as prior information, and combine them with joint motion data, inertial measurement data, six-dimensional torque data, and visual data to construct a weighted least squares problem including joint residuals, IMU pre-integration residuals, torque residuals, and visual residuals; solve the weighted least squares problem using the Gauss-Newton method to obtain optimized kinematic parameter corrections and camera extrinsic parameter corrections; correct the kinematic parameters of the target joints according to the optimized kinematic parameter corrections, correct the visual camera extrinsics according to the optimized camera extrinsic parameter corrections, and correct the sensor references according to the sensor reference offsets.

[0215] like Figure 6 As shown, in one possible implementation, the robot adaptive calibration device 50 based on load variation further includes a processing module 55, which is used to write the corrected kinematic parameters into the robot's motion control unit, and the motion control unit is used to perform forward kinematics calculation based on the corrected kinematic parameters; write the corrected visual camera extrinsic parameters into the robot's perception fusion unit, and the perception fusion unit is used to perform visual positioning based on the corrected visual camera extrinsic parameters; and write the corrected sensor reference into the robot's force control unit, and the force control unit is used to perform force conversion based on the corrected sensor reference.

[0216] In one possible implementation, when the calibration object includes kinematic parameters, sensor references, and visual camera extrinsics, the calibration module 54 is further configured to: acquire visual data collected by the visual camera during the robot's movement; if the quality of the visual data does not meet the quality requirements, use the kinematic parameter correction and sensor reference offset as prior information, and combine them with joint motion data, inertial measurement data, and six-dimensional torque data to construct a weighted least squares problem including joint residuals, IMU pre-integration residuals, and torque residuals; the quality requirements include that the number of continuously trackable visual feature points is less than a first threshold, or the number of effective visual frames is less than a second threshold; solve the weighted least squares problem using the Gauss-Newton method to obtain the optimized kinematic parameter correction; correct the kinematic parameters of the target joint according to the optimized kinematic parameter correction, and correct the sensor reference according to the sensor reference offset.

[0217] In one possible implementation, there are multiple target joints, and the identification module 52 is further configured to: independently identify the estimated value of the linear load inertial parameter vector corresponding to each target joint based on the joint motion data, inertial measurement data and six-dimensional torque data of each target joint using the recursive least squares method; and fuse the estimated values ​​of the linear load inertial parameter vector corresponding to each target joint to obtain the linear load inertial parameter vector of the robot end effector.

[0218] The robot adaptive calibration device based on load variation provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0219] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 7 As shown, the electronic device 31 provided in this embodiment includes at least one processor 311 and a memory 312. Optionally, the computing device 31 further includes a communication component 313. The processor 311, the memory 312, and the communication component 313 are connected via a bus 314.

[0220] In a specific implementation, at least one processor 311 executes computer execution instructions stored in memory 312, causing at least one processor 311 to perform the above-described method.

[0221] The specific implementation process of processor 311 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0222] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0223] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0224] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0225] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0226] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0227] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0228] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0229] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0230] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0231] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0232] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0233] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0234] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A robot adaptive calibration method based on load variation, characterized in that, include: In response to changes in the load at the robot's end effector, the robot's joints are controlled to move according to a preset joint space excitation trajectory during work breaks. During the movement, the joint motion data of the target joint, the inertial measurement data of the robot's wrist, and the six-dimensional torque data of the robot's end effector are acquired. Based on the joint motion data, the inertial measurement data, and the six-dimensional torque data, the linear load inertial parameter vector of the robot end effector is identified online using the recursive least squares method. Control the robot to stop moving and acquire the robot's operating data, which includes the end effector identifier and the operating status data of the main load-bearing joints; Based on the linear load inertia parameter vector and the working condition data, the kinematic parameter correction, sensor reference offset, and mapping confidence of the target joint are obtained through the structural deformation mapping model; wherein, the sensor reference offset is the same for different target joints. Based on the mapping confidence, the calibration object is determined, and the calibration object is calibrated based on the kinematic parameter correction and / or the sensor reference offset. The step of determining the calibration object based on the mapping confidence includes: If the mapping confidence is less than the confidence threshold, then the calibration object is determined to be a kinematic parameter; If the mapping confidence is greater than or equal to the confidence threshold, then the calibration object is determined to include the kinematic parameters, sensor reference, and visual camera extrinsic parameters; The sensor reference represents the transformation relationship between the coordinate system of the six-dimensional torque sensor at the robot's end effector and the robot's base coordinate system; the vision camera extrinsic represents the transformation relationship between the coordinate system of the vision camera and the robot's base coordinate system.

2. The robot adaptive calibration method based on load variation according to claim 1, characterized in that, Based on the joint motion data, the inertial measurement data, and the six-dimensional torque data, the linear load inertial parameter vector of the robot's end effector is identified online using the recursive least squares method, including: Based on the joint motion data and the inertial measurement data, a regression matrix of the target joint is constructed. The regression matrix represents the linear mapping relationship between the six-dimensional torque data and the linear load inertial parameter vector to be identified. When the regression matrix satisfies the preset solvable conditions, the recursive least squares method is used to identify the linear load inertia parameter vector of the robot end effector online based on the regression matrix and the six-dimensional torque data.

3. The robot adaptive calibration method based on load variation according to claim 2, characterized in that, The solvable conditions include: the rank of the regression matrix is ​​equal to the dimension of the linear load inertia parameter vector, and the condition number of the regression matrix is ​​less than a preset threshold; the condition number is the ratio of the maximum singular value to the minimum singular value of the regression matrix.

4. The robot adaptive calibration method based on load variation according to claim 2, characterized in that, Also includes: When the regression matrix does not meet the preset solvable conditions, perform the following operations: If the parameters of the excitation trajectory are adjusted, the joint motion data, the inertial measurement data, and the six-dimensional torque data are re-acquired based on the adjusted excitation trajectory; If the sampling time is extended, the joint motion data, the inertial measurement data, and the six-dimensional torque data will continue to be collected during the extended sampling time. Repeat the above operation until the regression matrix satisfies the solvable condition.

5. The robot adaptive calibration method based on load variation according to any one of claims 1 to 4, characterized in that, The kinematic parameters include link length and joint zero position; when the calibration object includes the kinematic parameters, the calibration of the calibration object based on the kinematic parameter correction and / or the sensor reference offset includes: Based on the kinematic parameter correction amount corresponding to the target joint, update the link length and joint zero position of the target joint; wherein, the updated link length is the sum of the current link length and the length correction amount in the kinematic parameter correction amount, and the updated joint zero position is the sum of the current joint zero position and the zero position offset in the kinematic parameter correction amount.

6. The robot adaptive calibration method based on load variation according to any one of claims 1 to 4, characterized in that, When the calibration object includes the kinematic parameters, the sensor reference, and the visual camera extrinsic parameters, calibrating the calibration object based on the kinematic parameter correction and / or the sensor reference offset includes: Acquire visual data collected by the visual camera during the robot's movement; Using the kinematic parameter correction and the sensor reference offset as prior information, and combining the joint motion data, the inertial measurement data, the six-dimensional torque data, and the visual data, a weighted least squares problem is constructed that includes joint residuals, IMU pre-integration residuals, torque residuals, and visual residuals. The weighted least squares problem is solved using the Gauss-Newton method to obtain the optimized kinematic parameter corrections and camera extrinsic parameter corrections. The kinematic parameters of the target joint are corrected according to the optimized kinematic parameter correction amount, the visual camera extrinsic parameters are corrected according to the optimized camera extrinsic parameter correction amount, and the sensor reference is corrected according to the sensor reference offset amount.

7. The robot adaptive calibration method based on load variation according to claim 6, characterized in that, Also includes: The corrected kinematic parameters are written into the robot's motion control unit, which is used to perform forward kinematics calculations based on the corrected kinematic parameters. The corrected visual camera extrinsic parameters are written into the robot's perception fusion unit, which is used for visual localization based on the corrected visual camera extrinsic parameters. The corrected sensor reference is written into the force control unit of the robot, and the force control unit is used to perform force conversion based on the corrected sensor reference.

8. The robot adaptive calibration method based on load variation according to any one of claims 1 to 4, characterized in that, When the calibration object includes the kinematic parameters, the sensor reference, and the visual camera extrinsic parameters, the calibration of the calibration object based on the kinematic parameter correction and / or the sensor reference offset includes... Acquire visual data collected by the visual camera during the robot's movement; If the quality of the visual data does not meet the quality requirements, the kinematic parameter correction and the sensor reference offset are used as prior information. Combined with the joint motion data, the inertial measurement data, and the six-dimensional torque data, a weighted least squares problem including joint residuals, IMU pre-integration residuals, and torque residuals is constructed. The quality requirements include that the number of continuously trackable visual feature points is less than a first threshold, or the number of effective visual frames is less than a second threshold. The weighted least squares problem is solved using the Gauss-Newton method to obtain the optimized kinematic parameter corrections. The kinematic parameters of the target joint are corrected according to the optimized kinematic parameter correction amount, and the sensor reference is corrected according to the sensor reference offset amount.

9. The robot adaptive calibration method based on load variation according to any one of claims 1 to 4, characterized in that, The number of target joints is multiple. Based on the joint motion data, the inertial measurement data, and the six-dimensional torque data, the linear load inertial parameter vector of the robot's end effector is identified online using a recursive least squares method, including: Based on the joint motion data, inertial measurement data and six-dimensional torque data of each target joint, the linear load inertial parameter vector estimate value corresponding to each target joint is independently identified by using the recursive least squares method. The estimated linear load inertia parameter vectors corresponding to each of the target joints are fused to obtain the linear load inertia parameter vector of the robot end effector load.

10. A robot adaptive calibration system based on load variation, characterized in that, include: Electronic equipment, and a six-dimensional torque sensor, an inertial measurement unit, and a joint encoder that are communicatively connected to the electronic equipment; The six-dimensional torque sensor is installed at the end of the robot and is used to collect six-dimensional torque data of the end of the robot. The inertial measurement unit is disposed on the robot wrist and is used to collect inertial measurement data of the robot wrist; The joint encoder is installed on the target joint of the robot and is used to collect joint motion data of the target joint. The electronic device is used to execute the robot adaptive calibration method based on load variation according to any one of claims 1 to 9, based on the six-dimensional torque data, the inertial measurement data, and the joint motion data.

11. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the robot adaptive calibration method based on load variation as described in any one of claims 1 to 9.

Images