Real-time evaluation method for whole-body joint coordination of humanoid robot

Abstract

The invention provides a humanoid robot whole-body joint coordination real-time evaluation method, which comprises the following steps: acquiring real-time data of each joint of a humanoid robot, determining an included angle relationship between a mass center offset and a forward direction, and constructing an initial distribution architecture; carrying out filtering processing on the joint data according to the initial distribution architecture, and generating denoised torque and rotation angle data; determining a torque distribution set of each joint through the denoised torque and rotation angle data, and analyzing a complexity quantitative evaluation value of whole-body joint compensation; the torque distribution weight of each joint is dynamically adjusted based on the complexity quantitative evaluation value, and an optimized torque distribution scheme is generated; and according to the optimized torque distribution scheme, identifying an actual deviation value, carrying out torque redistribution on a joint combination with weak coordination, and determining a corrected torque distribution compensation scheme.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for real-time evaluation of the coordination of the whole-body joints of a humanoid robot. Background Technology

[0002] Humanoid robot research occupies a crucial position in modern technology, with applications spanning industrial production, medical assistance, and daily life services, demonstrating irreplaceable value, especially in complex tasks such as object handling. The coordination of a robot's joints directly determines the stability and efficiency of task execution, and is one of the core indicators for evaluating robot performance. However, achieving precise joint coordination in dynamic environments remains a key challenge that urgently needs to be overcome in this field. Current research and application methods often struggle to adapt to the diverse characteristics of objects when handling tasks, especially when dealing with non-homogeneous objects. Existing solutions focus more on the robot's own motion planning, neglecting the impact of uneven object distribution. This neglect prevents the robot from flexibly adjusting to the actual situation, leading to uneven joint load and potentially causing mechanical damage or task failure. Particularly when the object's center of gravity changes, relying solely on preset control strategies is insufficient to handle complex real-world scenarios. A deeper technical challenge lies in the fact that the changing angle between the object's center of gravity shift direction and the robot's forward direction places extremely high demands on joint coordination. The shift in the center of gravity directly affects the torque that each joint of a robot needs to bear, and changes in the angle between the joints amplify this effect, leading to increased uncertainty in torque distribution. If this change in angle cannot be accurately identified and adapted to, the robot cannot rationally distribute the compensating torque of each joint, thus affecting the overall stability of its movement. For example, when moving a box with its center of gravity shifted to one side, if the robot's forward direction forms a large angle with the direction of the center of gravity shift, some joints may need to bear a much higher load than expected, while other joints are inefficient. This imbalance not only increases energy consumption but may also lead to robot instability. Therefore, how to rationally adjust the torque distribution of each joint based on the dynamic changes in the angle between the direction of the center of gravity shift and the forward direction when moving non-homogeneous objects, in order to achieve coordinated operation of all joints, has become a critical problem that urgently needs to be solved. Summary of the Invention

[0003] This invention provides a method for real-time evaluation of the coordination of all joints in a humanoid robot, mainly including: Acquire real-time data of each joint of the humanoid robot, determine the angular relationship between the center of mass offset and the forward direction, and construct the initial allocation architecture; Based on the initial allocation architecture, the joint data is filtered to generate denoised torque and rotation angle data. The torque distribution set of each joint is determined by the denoised torque and rotation angle data, and the complex quantitative evaluation value of whole-body joint compensation is analyzed. Based on the complex quantitative evaluation value, the torque distribution weight of each joint is dynamically adjusted to generate an optimized torque distribution scheme. Based on the optimized torque distribution scheme, the actual deviation is identified, the torque is redistributed to the weakly coordinated joint combination, and the corrected torque distribution compensation scheme is determined.

[0004] Furthermore, the process of acquiring real-time data of each joint of the humanoid robot, determining the angular relationship between the center of mass offset and the forward direction, and constructing an initial allocation architecture includes: Real-time torque values ​​are obtained from sensors deployed at each joint, rotation angle data are obtained from the inertial measurement unit, the angle between the center of mass offset direction and the forward direction is calculated, and the center of mass offset is determined by multiplying the cosine of the angle by the offset distance. Cross-validation was performed using the centroid offset and the torque value; Based on the comparison between the verified centroid offset and the preset boundary, the urgency level is divided and a urgency level label is formed. A hierarchical compensation architecture is constructed using the aforementioned urgency level identifiers, and different weight values ​​are assigned to establish an initial allocation architecture that includes joint identifiers, urgency levels, and compensation weights.

[0005] Furthermore, the step of filtering the joint data according to the initial allocation architecture to generate denoised torque and rotation angle data includes: The adaptive filtering parameters are calculated based on the urgency level and compensation weight. The filter window length and cutoff frequency of joints with different urgency levels are set to form a filter parameter configuration table. The filtering parameter configuration table is used to perform filtering on the real-time torque signal and rotation angle signal. Through state prediction and measurement update iteration calculation, the noise covariance is adjusted for joints with different urgency levels to obtain the preliminary filtered data sequence. By performing a sliding window averaging operation on the pre-filtered data sequence, setting the window size according to the urgency level, collecting data points at time intervals, and generating denoised and smoothed torque and rotation angle data arranged in time series.

[0006] Furthermore, the step of determining the torque distribution set of each joint using the denoised torque and rotation angle data, and analyzing the complex quantitative evaluation value of whole-body joint compensation, includes: A joint torque distribution matrix is ​​constructed based on the denoised torque and rotation angle. The denoised torque is then grouped using a clustering algorithm to form torque distribution sets of different levels. Principal component analysis is performed using the aforementioned torque distribution set to extract the main torque distribution directions, calculate the projection values ​​of each joint torque on the main torque distribution directions, and classify them into dominant joint groups; The degree of dispersion of the distribution is calculated by the standard deviation of the torque values ​​within the dominant joint group. The coupling coefficient is determined based on the ratio of torque to angle change rate of adjacent joints. The weighted sum of the degree of dispersion of the distribution and the coupling coefficient is used to calculate the complex quantitative evaluation value of whole-body joint compensation.

[0007] Furthermore, the step of dynamically adjusting the torque distribution weights of each joint based on the complex quantified evaluation value to generate an optimized torque distribution scheme includes: The weight adjustment parameters are determined based on the difference between the complex quantitative evaluation value and the preset boundary. The basic ratio of torque distribution for each joint is calculated by combining the weight adjustment parameters and the centroid offset. Calculate the expected torque value based on the basic ratio value, compare the difference between the expected torque value and the actual torque value, and output the optimized allocation scheme.

[0008] Furthermore, the step of calculating the basic ratio of the torque distribution of each joint by combining the weight adjustment parameters and the centroid offset includes: using the weight adjustment parameters, constructing a system of equations containing the torque variables of each joint and the centroid position constraints by combining the centroid offset, solving for the extreme value of the weight distribution by the Lagrange multiplier method under the constraint that the centroid offset remains unchanged, and obtaining the basic ratio of the torque distribution of each joint.

[0009] Furthermore, the step of identifying the actual deviation based on the optimized torque distribution scheme, redistributing torque to the weakly coordinated joint combinations, and determining the corrected torque distribution compensation scheme includes: The actual torque value is calculated based on the theoretical torque value and the smooth rotation angle in the optimized scheme, and the actual deviation is determined by the ratio of the difference between the theoretical value and the actual value. The actual deviation is compared with a preset threshold to identify weak joint combinations based on the joint kinematic chain topology. Based on the deviation characteristics of the weak joint combination, the deviations are prioritized according to their magnitude. An adjustment coefficient is determined according to the ratio of the magnitude of the deviation to the average deviation. The torque distribution ratio is then adjusted, and the corrected compensation scheme is output.

[0010] Furthermore, the method also includes: using the corrected torque distribution compensation scheme and real-time data to generate a coordination quantification score, adaptively adjusting the data acquisition parameters, and cyclically updating the input to the initial allocation architecture to achieve whole-body joint coordination assessment.

[0011] Furthermore, the step of using the corrected torque distribution compensation scheme and real-time data to generate a coordination quantification score, and adaptively adjusting the data acquisition parameters, includes: The compensation scheme after integration and correction is combined with real-time torque data. The ratio of the deviation between theoretical torque and actual torque is calculated as a coordination index. The index is then weighted and averaged according to the importance of the joints to generate a quantitative assessment report on coordination. The weighted average value is extracted from the coordination quantitative assessment report as the coordination quantitative score, and the deviation rate is calculated by comparing the coordination quantitative score with the preset boundary. The data acquisition frequency adjustment amount is determined based on the deviation rate, and the frequency is updated to obtain real-time data.

[0012] Furthermore, the cyclical update input to the initial allocation architecture to achieve whole-body joint coordination assessment includes: The real-time data obtained from the data acquisition frequency updated according to the adjustment coefficient is used as the new input and sent to the initial allocation architecture; The torque allocation and deviation identification process is re-executed using the initial allocation architecture to generate a new round of coordination quantitative assessment report; By comparing the coordination score in the new round of evaluation report with the preset boundary, the data acquisition parameters are continuously adjusted, and dynamic evaluation of the coordination of the humanoid robot's whole-body joints is achieved through continuous iteration.

[0013] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses a real-time evaluation method for the coordination of all joints in a humanoid robot. Addressing the coordination deficiencies caused by centroid shift and uneven joint torque distribution during robot operation, the method acquires data from torque sensing units and inertial measurement units, identifies the centroid shift, and verifies its accuracy by combining real-time torque and rotation angles, constructing a hierarchical compensation initial allocation architecture. The invention adaptively adjusts filtering parameters based on urgency weights to denoise and smooth the data, thereby determining the joint torque distribution set, dynamically optimizing the torque allocation scheme, and correcting the allocation scheme through deviation filtering and priority reordering. Finally, a coordination evaluation report is generated, and the sampling frequency is adaptively adjusted for continuous optimization. Through iterative optimization of centroid shift constraints and torque redistribution, this invention ensures improved joint coordination and stability, significantly enhancing the humanoid robot's motion performance in complex environments. Attached Figure Description

[0014] Figure 1 This is a flowchart of a method for real-time evaluation of the coordination of the whole-body joints of a humanoid robot according to the present invention. Detailed Implementation

[0015] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] like Figure 1 This embodiment of a method for real-time evaluation of the coordination of the whole-body joints of a humanoid robot may specifically include: Step S101: Obtain real-time data of each joint of the humanoid robot, determine the angle relationship between the center of mass offset and the forward direction, and construct the initial allocation architecture.

[0017] Real-time torque values ​​are acquired from six-axis torque sensors deployed at the hip, knee, ankle, shoulder, elbow, and wrist joints of the humanoid robot. Simultaneously, the rotation angle of the object's inertial tensor principal axis relative to the robot's coordinate system is read from an independent inertial measurement unit mounted on an externally mounted object handling a non-homogeneous object. The centroid offset direction vector is projected onto the robot's forward direction vector to obtain the included angle θ. The centroid offset is obtained by multiplying the cosine of the included angle by the offset distance d. The centroid offset is cross-validated with the real-time torque values ​​of each joint. The ratio of the sum of the components of each joint's torque value along the centroid offset direction to the theoretical torque value is calculated as a consistency index. If the consistency index exceeds a first preset threshold, the current offset is retained; if the consistency index is below the first preset threshold, the offset is recalculated based on the correspondence between the rotation angle of the inertial tensor principal axis and the centroid position to obtain the verified centroid offset. The verified centroid offset is compared with the preset compensation trigger boundary. If the offset exceeds the second preset threshold of the boundary value, it is marked as high urgency; if it is between the third and second preset thresholds, it is marked as medium urgency; if it is between the fourth and third preset thresholds, it is marked as low urgency, forming an urgency level identifier. An initial allocation architecture for hierarchical compensation is constructed using these urgency level identifiers. High-urgency joint groups receive a fifth preset weight value, medium-urgency joint groups receive a sixth preset weight value, and low-urgency joint groups receive a seventh preset weight value, establishing an initial allocation architecture that includes joint identifiers, urgency levels, and compensation weights.

[0018] Specifically, in one implementation, when a humanoid robot is carrying a non-homogeneous object, it achieves real-time monitoring of torque and posture by deploying a high-precision sensor array on key joints throughout its body.

[0019] Specifically, six-axis torque sensors are installed in the joint bearing housings of the hip, knee, ankle, shoulder, elbow, and wrist joints. Each sensor contains three orthogonal force sensing elements and three orthogonal torque sensing elements, with a sampling frequency set to 1000Hz to ensure the capture of transient torque changes during handling. An inertial measurement unit (IMU) is installed at the center of the robot's torso, integrating a three-axis gyroscope, a three-axis accelerometer, and a three-axis magnetometer for real-time measurement of the object's attitude changes relative to the robot's body coordinate system. The calculation of the center-of-mass offset involves transformations between multiple coordinate systems and vector operations.

[0020] For example, when a robot moves a container filled with liquid, the sloshing of the liquid causes a dynamic change in the position of the object's center of mass. First, the rotation matrix of the object's inertial tensor principal axes in the robot's body coordinate system is obtained using an inertial measurement unit. This matrix describes the directional relationship between the object's principal inertial axes and the robot's coordinate axes. Then, based on the readings from the torque sensors at each joint, the position vector of the object's center of mass in the robot coordinate system is calculated using the static equilibrium equations. Robot's forward direction vector Provided by the motion control system, this represents the robot's current direction of movement. Centroid offset direction vector. Defined as the difference between the centroid position vector and the center point of the robot's two hands gripping the object. This is calculated... exist The projection on the surface yields the included angle. The final calculation of the centroid offset is as follows: This value reflects the effective component of the centroid offset in the forward direction.

[0021] It should be noted that the consistency index is used to verify the accuracy of the center of mass offset. By analyzing the component distribution of the torque of each joint in the direction of the center of mass offset, the degree of matching between the measured torque and the theoretical torque is evaluated. In specific calculations, the torque vector Fi of each joint is projected onto the unit vector uoffset in the direction of the center of mass offset to obtain the offset direction torque component Fiproj of each joint.

[0022] In one possible implementation, the theoretical torque value is calculated based on the object's mass m, gravitational acceleration g, and the distance of the center of mass offset. Consistency Indicators Defined as This ratio reflects the degree of agreement between the measured torque distribution and the theoretical expectation.

[0023] Preferably, when the consistency index is lower than a first preset threshold, the offset correction procedure is initiated. The correction process utilizes the mapping relationship between the rotation angle of the inertial tensor principal axis and the position of the center of mass. By querying a pre-established table of rotation angles and center of mass positions, the corrected coordinates of the center of mass are obtained. This table is obtained through offline calibration of the object in different postures and stores the correspondence between the rotation angle increment and the center of mass displacement. The urgency grading mechanism determines the compensation priority of each joint based on the relative relationship between the center of mass offset and the preset compensation trigger boundary.

[0024] For example, when a robot is carrying a toolbox whose center of gravity is severely skewed to one side, the offset may exceed twice the boundary value. In this case, the joint group bearing the main eccentric torque is marked as high stress; these joints typically include the hip and ankle joints on the skewed side. A second preset threshold is usually set at 1.5 times the boundary value; when the offset exceeds this threshold, the system determines that the current state requires immediate and significant compensation. The third and fourth preset thresholds correspond to 1.0 and 0.5 times the boundary value, respectively, to distinguish between moderate and low stress conditions. Through this hierarchical mechanism, the system can identify which joints are under high load and which joints still have compensation margin. The weight allocation of the hierarchical compensation architecture reflects the idea of ​​optimal resource allocation.

[0025] In one embodiment, high-pressure joint groups are assigned a fifth preset weight value, which typically accounts for about 50% of the total compensation resources, ensuring that these joints receive priority and sufficient torque compensation. Medium-pressure and low-pressure joint groups are assigned sixth and seventh preset weight values, respectively, which are dynamically adjusted according to the robot's load capacity and stability requirements. The weight values ​​not only determine the distribution ratio of compensation torque but also affect the timing priority of the compensation response.

[0026] Understandably, the initial allocation architecture uses a tree-like data structure to store joint compensation information. The root node contains the overall compensation requirements and allocation strategy parameters, while each joint node stores the joint identifier, current torque value, target compensation value, urgency level, and allocation weight. Each node also contains pointers to adjacent joints, forming a joint link topology, which facilitates consideration of coupling effects between joints during the compensation process. The architecture also reserves a dynamic adjustment interface, allowing for updates to weight allocation and compensation strategies based on real-time feedback. This hierarchical architecture design enables the robot to adaptively adjust the coordination mode of its joints when handling non-homogeneous objects with different characteristics.

[0027] Step S102: Filter the joint data according to the initial allocation architecture to generate denoised torque and rotation angle data.

[0028] Based on the urgency level and compensation weight of each joint in the initial allocation architecture, adaptive filtering parameters are calculated. The filtering window length for high-urgency joints is set to the base window length divided by the weight value, and the cutoff frequency is set to the weight value multiplied by the base frequency. For medium-urgency and low-urgency joints, the window length decreases and the cutoff frequency increases proportionally according to the weight, forming a filtering parameter configuration table containing joint identifiers, window lengths, and cutoff frequencies. Using the window lengths and cutoff frequencies in the filtering parameter configuration table, Kalman filtering is performed on the real-time torque and rotation angle signals of each joint. Through state prediction and measurement update iteration calculations, a larger process noise covariance Q is used for high-urgency joints, where Q represents the process noise covariance that preserves the rapid change characteristics of the signal, while a smaller process noise covariance Q is used for low-urgency joints to suppress noise interference, obtaining the pre-filtered torque and angle sequences. By performing a sliding window averaging operation on the pre-filtered torque and angle sequences, the window size is set to 1, 2, and 3 times the preset baseline value according to the stress level of each joint, and data points are collected at time intervals that are integer multiples of the sensor sampling period, generating denoised smoothed torques and smoothed rotation angles arranged in time series.

[0029] Specifically, in one implementation, the adaptive filtering parameters are determined based on the urgency level and compensation weight values ​​in the initial allocation architecture.

[0030] Specifically, the reference window length is set to 100 sampling points, and the reference cutoff frequency is set to 50Hz. For high-stress joints, the filter window length is calculated by multiplying the reference window length by the compensation weight value of the joint; when the weight value is 0.5, the window length is 50 sampling points. The cutoff frequency is obtained by dividing the reference frequency by the weight value, i.e., 100Hz. This setting allows high-stress joints to respond quickly to torque changes and retain more dynamic information. The process noise covariance matrix Q and measurement noise covariance matrix R in the Kalman filtering process are differentiated according to the joint stress level. The process noise covariance of high-stress joints is set to a smaller value, such as 0.001, making the filter more confident in the system model prediction, thus preserving the rapid change characteristics of the signal; the process noise covariance of low-stress joints is set to a larger value, such as 0.01, to enhance noise suppression. Kalman filtering uses the state prediction equation... ,in Let A be the prior state estimate at the current moment, and let A be the state transition matrix. The state estimate and measurement update equation for the previous time step. , where xk is the posterior state estimate at the current time, K is the Kalman gain, zk is the current measurement value, H is the measurement matrix, and the process is executed iteratively, where K is the Kalman gain, which is dynamically calculated based on the covariance of process noise and measurement noise.

[0031] Preferably, the sliding window averaging operation uses a weighted moving average method, where data points within the window are assigned different weights based on their time proximity.

[0032] For example, when the robot shakes while handling a liquid container, the five newest sampling points have a total weight of 0.3 and are evenly distributed, the five next newest sampling points have a total weight of 0.2 and are evenly distributed, and the remaining sampling points are evenly distributed with the remaining total weight of 0.5, resulting in a total weight sum of 1. The window size is set according to the urgency level: high urgency joint windows contain 20 sampling points, medium urgency windows contain 50 sampling points, and low urgency joint windows contain 100 sampling points.

[0033] In one possible implementation, the time interval is directly related to the sensor sampling period. When the sensor sampling frequency is 1000Hz, the sampling period is 1ms, and the time interval is set to 10 times the sampling period, meaning one data point is collected every 10ms for subsequent processing. Through this hierarchical filtering process, the output is a denoised smoothing torque and a smoothed rotation angle arranged in a time series.

[0034] Step S103: Determine the set of torque distributions for each joint using the denoised torque and rotation angle data, and analyze the complex quantitative evaluation value of whole-body joint compensation.

[0035] A joint torque distribution matrix is ​​constructed based on smoothed torque and smoothed rotation angle. Each matrix element corresponds to the torque value of a joint at a specific rotation angle. K-means clustering is used to group all torque values ​​in the matrix, with three cluster centers, forming a torque distribution set containing high, medium, and low levels. Subsequently, a row of torque values ​​is extracted from each joint in the matrix as a torque vector, and assigned to the corresponding set based on the clustering results. Principal component analysis is performed on the torque vectors of each joint in the torque distribution set. The first principal component is extracted as the main torque distribution direction, and the projection value of each joint torque on the main distribution direction is calculated. If the projection value exceeds a preset multiple of the mean of all projection values, the joint is assigned to the dominant joint group. The dispersion of the torque distribution is calculated using the standard deviation of the torque values ​​of each joint within the dominant joint group. The coupling coefficient is determined based on the ratio of the torque change rate to the angle change rate between adjacent joints and directly connected joints in the robot topology. A weighted sum of the dispersion and coupling coefficient is used to calculate the complex quantitative evaluation value of whole-body joint compensation. The weighting coefficients are preset according to the robot's load state.

[0036] Specifically, in one implementation, the joint torque distribution matrix is ​​constructed in the form of a two-dimensional matrix, with rows corresponding to the joint numbers of the robot and columns corresponding to discrete sampling points of the rotation angle.

[0037] Specifically, the rotation range from 0 to 360 degrees is discretized at 10-degree intervals to form 36 angle sampling points. Combined with the robot's 18 main joints, an 18×36 torque distribution matrix is ​​constructed. Each matrix element records the smooth torque value of a specific joint at a specific angle, reflecting the distribution characteristics of torque with attitude changes during the handling process.

[0038] It should be noted that the K-means clustering algorithm first expands the moment values ​​in the matrix into a one-dimensional vector, and sets three initial cluster centers corresponding to the 25%, 50%, and 75% quantiles of the moment value range, respectively. Iteratively, the Euclidean distance from each moment value to each cluster center is calculated, and the value is assigned to the cluster of the nearest center. After multiple iterations until the cluster centers no longer change significantly, three sets of moment levels—high, medium, and low—are ultimately formed. The high level corresponds to joints bearing the main load, and the low level corresponds to joints providing auxiliary support.

[0039] For example, the principal component analysis process uses the torque vector of each joint as input data to construct a covariance matrix and calculate its eigenvalues ​​and eigenvectors. The first principal component corresponds to the eigenvector of the largest eigenvalue, representing the main direction of torque change. When the robot handles an eccentric load, the first principal component typically points in the direction of the center of mass shift. The projection values ​​of the torque of each joint in the principal distribution direction are obtained through vector dot product operations. If the projection value of a joint exceeds 1.5 times the mean of all projection values, then that joint is classified into the dominant joint group, and these joints undertake the main task of compensating for eccentric loads.

[0040] Preferably, the calculation of the coupling coefficient takes into account the kinematic correlation between joints. The rate of change of torque between adjacent joints is defined as the difference in torque between the two joints divided by the sampling time interval, and the rate of change of angle is the difference in angle between the two joints divided by the sampling time interval. The ratio of these two values ​​is the coupling coefficient. This coefficient reflects the degree of coordination between adjacent joints during the compensation process; a larger coupling coefficient indicates a stronger mutual influence between joints.

[0041] Understandably, the complexity quantification evaluation value is calculated using a normalized weighted sum method. The dispersion of the torque distribution is quantified using the standard deviation and normalized to the 0-1 range; the coupling coefficient is also normalized. The weighting coefficients are dynamically adjusted according to the robot's current load rate: at high load, the dispersion weight is 0.7 and the coupling coefficient weight is 0.3; at low load, both weights are 0.5. The complexity quantification evaluation value reflects the overall difficulty of the current compensation task.

[0042] Step S104: Dynamically adjust the torque distribution weights of each joint based on the complex quantified evaluation value to generate an optimized torque distribution scheme.

[0043] The degree of deviation is calculated based on the difference between the quantitative evaluation value of complexity and the preset complexity boundary. The degree of deviation is equal to the evaluation value minus the boundary value, divided by the boundary value. If the degree of deviation is positive and exceeds the preset upper limit threshold, the weight coefficient of the joint bearing the main load is increased proportionally to the degree of deviation. If the degree of deviation is negative and below the preset lower limit threshold, the weight coefficient of the joint bearing the auxiliary load is decreased proportionally to the degree of deviation, thus obtaining preliminary weight adjustment parameters. Using the preliminary weight adjustment parameters, a system of equations containing the torque variables of each joint and the centroid position constraints is constructed in conjunction with the centroid offset. The extreme values ​​of the weight allocation are solved using the Lagrange multiplier method under the constraint that the centroid offset remains unchanged, obtaining the basic ratio value of the torque allocation proportion of each joint. The expected torque value borne by each joint is calculated based on the basic ratio value and compared with the current actual torque value. If the difference exceeds the preset torque deviation threshold, torque transfer is performed between adjacent joints proportionally according to the magnitude of the difference, updating the numerical ratio relationship. The torque is redistributed using the updated numerical ratio relationship. The gradient descent method is used to iteratively adjust the allocation weight of each joint. The overall torque deviation after each iteration is calculated until the deviation is less than the convergence threshold or the maximum number of iterations is reached. The optimized torque distribution scheme for all joints is then output.

[0044] Specifically, in one implementation, the degree of deviation is calculated based on the relative relationship between the complexity quantification evaluation value and the preset complexity boundary.

[0045] Specifically, when a humanoid robot is handling goods with a severely offset center of gravity, the quantitative assessment value of complexity reflects the difficulty level of the current compensation task. The preset complexity boundary is pre-calibrated based on the robot's structural parameters and load capacity, typically set as the standard complexity value under the robot's rated load. The deviation degree is calculated as: Deviation degree = (Assessment value - Boundary value) / Boundary value × 100%. When the deviation degree is +20%, it indicates that the current task complexity exceeds the standard state by 20%, and the system determines that the compensation capability of the main load-bearing joints needs to be enhanced. The adjustment of the weight coefficients adopts a graded response mechanism, with an upper limit threshold of 15% and a lower limit threshold of -10%. When the deviation degree exceeds the upper limit threshold, the weight coefficient of the joint group bearing the main load is increased by 0.8 times the deviation degree, i.e., new weight = original weight × (1 + 0.8 × deviation degree). The joint group bearing the main load typically includes large lower limb joints such as the hip and knee joints, which bear the main weight of the object during handling. Correspondingly, the weighting coefficients of joint groups bearing auxiliary loads, such as wrist and finger joints, remain unchanged or are slightly reduced, thereby achieving a redistribution of compensation resources.

[0046] For example, the process of constructing the constraint equations comprehensively considers two key factors: mechanical equilibrium and center of mass stability. The equations contain n joint moment variables. And the centroid position constraint. The centroid offset constraint is expressed as... Where Ti is the torque of the i-th joint. This is the lever arm length from the joint to the center of mass. For the mass of the object, It is the acceleration due to gravity. This represents the centroid offset distance. The Lagrange function is constructed as follows: ,in These are the weighting coefficients for each joint. Let L be the Lagrange multiplier. By taking the partial derivative with respect to L and setting it to zero, the optimal torque distribution value satisfying the center-of-mass constraint is obtained.

[0047] In one possible implementation, after the basic ratio values ​​are determined, the system enters the torque verification and adjustment phase. The expected torque value for each joint is calculated by multiplying the basic ratio value by the total compensation torque. The actual torque value is acquired in real time by torque sensors. When the difference between the expected and actual values ​​exceeds the torque deviation threshold, the torque transfer mechanism is activated. The torque deviation threshold is set differently according to the joint type: 8% of the rated torque for large joints and 5% of the rated torque for small joints.

[0048] Preferably, torque transfer occurs between adjacent joints according to the joint coupling strength.

[0049] For example, when the actual torque of the shoulder joint exceeds the expected value by 12%, the system calculates that 60% of the excess torque needs to be transferred. This transfer amount is allocated to the elbow and wrist joints respectively, according to the shoulder-elbow joint coupling coefficient of 0.7 and the shoulder-wrist joint coupling coefficient of 0.3. The new ratios after the transfer are: Elbow joint ratio = Original ratio + Transfer amount × 0.7 / Elbow joint rated torque; Wrist joint ratio = Original ratio + Transfer amount × 0.3 / Wrist joint rated torque.

[0050] Specifically, the iterative optimization process of the gradient descent method aims to minimize the overall torque deviation. The objective function is defined as follows: ,in The actual torque of the i-th joint. The desired torque is given. Gradient calculation uses a numerical differentiation method, with the learning rate initially set to 0.01 and dynamically adjusted according to the iteration progress. In each iteration, the weight update rule is as follows: ,in For learning rate, For the new weights, For the old weights, the loss function Regarding weight The gradient is calculated. The algorithm is considered convergent when the objective function changes by less than 0.001 over five consecutive iterations or when the number of iterations reaches 100. In real-world handling scenarios, when a robot handles boxes containing irregularly shaped goods, the center of gravity offset may dynamically change during the handling process. A complete optimization process is executed every 50ms, updating the torque distribution scheme in real time. When a sudden change in the center of gravity offset exceeding 10cm is detected, the system automatically increases the optimization frequency to once every 20ms to ensure rapid adaptation to load changes.

[0051] For example, when a robot starts from a standstill, carrying a heavy object with its center of gravity slightly to the left and moving forward, the left hip and knee joints initially bear a large torque. Through dynamic optimization, some of the torque is gradually transferred to the right joints and upper limb joints, eventually achieving a stable state of coordinated operation of all joints. The optimized scheme maintains the torque utilization rate of each joint within 40%-80% of the rated value, avoiding local overload.

[0052] In one embodiment, a stability metric is also introduced into the convergence determination. In addition to the overall torque deviation, the system simultaneously monitors the oscillation amplitude of the torque at each joint. If the torque fluctuation of a joint exceeds 15% of the mean within 10 consecutive sampling periods, the weight adjustment step size of that joint is reduced to prevent oscillations caused by the optimization process.

[0053] Step S105: Identify the actual deviation based on the optimized torque distribution scheme, redistribute the torque for the weakly coordinated joint combination, and determine the corrected torque distribution compensation scheme.

[0054] Based on the theoretical torque values ​​of each joint in the optimized full-joint torque distribution scheme, and combined with the smoothed rotation angles acquired over time, the actual torque value that each joint should generate at the current angle is calculated using the product relationship between torque and angle. The difference between the theoretical and actual values ​​is divided by the theoretical value to obtain the actual deviation of each joint. The actual deviation is compared with a preset deviation threshold. Based on the topology of the joint kinematic chain, if the deviation of a certain joint and its adjacent joints exceeds the threshold, the joint chain is identified as a weak joint combination, and the joint number and deviation characteristics of each joint in the combination are recorded. Based on the deviation characteristics of the weak joint combination, the joints are prioritized and reordered according to the deviation from largest to smallest. A torque adjustment coefficient is determined based on the ratio of the deviation to the average deviation. Joints with deviations exceeding the average value have their torque distribution reduced proportionally according to the adjustment coefficient, while joints with deviations below the average value have their torque distribution increased accordingly. The corrected full-joint torque distribution compensation scheme is then output.

[0055] Specifically, in one implementation, the actual deviation is calculated based on the dynamic relationship between torque and joint angle.

[0056] Specifically, the theoretical torque values ​​for each joint are derived from the optimized torque distribution scheme across all joints. These values ​​represent the torque that each joint should bear under ideal conditions. The actual torque values ​​are calculated by substituting the smoothed rotation angle into the joint dynamics equations. The calculation shows that, among which For torque, For the moment of inertia of the joint, Angular acceleration, This is an angle-related gravity compensation term. The actual deviation is defined as (actual value - theoretical value) / theoretical value × 100%, and this percentage directly reflects the degree of deviation from the expected performance.

[0057] It should be noted that the identification of weak joint combinations relies on the topology of the joint kinematic chains. The joint chains of a humanoid robot are distributed in a tree-like structure, extending from the torso to the extremities. Deviation thresholds are set differently depending on the joint type: 8% for large joints such as the hip, 6% for medium joints such as the elbow, and 4% for small joints such as the finger joints. When the deviation of a joint and its adjacent joints exceeds their respective thresholds, it indicates a coordination problem in that local kinematic chain, and these joints are marked as weak combinations.

[0058] For example, the torque adjustment coefficient is determined using a proportional allocation principle. First, the average deviation of all joints within the weak joint combination is calculated. Then, the adjustment coefficient is determined based on the ratio of each joint's deviation to the average value. If a joint's deviation is 12% and the average deviation is 8%, then the adjustment coefficient for that joint is 12 / 8 = 1.5. During torque redistribution, the new allocation value for that joint = the original allocation value / 1.5, and the reduced torque is transferred proportionally to the load-bearing capacity of adjacent joints.

[0059] Preferably, the priority reordering follows the principle of decreasing deviation, with the joint with the largest deviation receiving the highest adjustment priority. In actual handling, if the left knee joint deviation is 15%, the left hip joint deviation is 10%, and the left ankle joint deviation is 7%, then the adjustment order is left knee, left hip, and left ankle. After each adjustment, the overall torque balance is recalculated to ensure the stability of the compensation scheme. The corrected full-joint torque distribution compensation scheme achieves dynamic equilibrium through iterative adjustments.

[0060] For example, when a robot turns while carrying eccentric cargo, the outer leg joints often exhibit weak coordination. By identifying and redistributing the torque, some of the torque is transferred to the inner leg and upper limb joints, achieving whole-body coordination compensation and improving handling stability.

[0061] Step S106: Using the corrected torque distribution compensation scheme and real-time data to generate a coordination quantification score, adaptively adjust the data acquisition parameters, and cyclically update the input to the initial distribution architecture to achieve whole-body joint coordination assessment.

[0062] The integrated and corrected full-joint torque distribution compensation scheme is combined with real-time torque data for each joint. The ratio of the theoretical torque to the actual torque deviation for each joint is calculated as a coordination index. The coordination indexes of all joints are weighted and averaged according to their importance, resulting in a quantitative coordination assessment report containing joint number, coordination index, weighted average, and timestamp. The weighted average is extracted from this report as the quantitative coordination score, compared with a preset score boundary, and the score deviation rate is calculated. If the deviation rate exceeds a preset upper threshold, the adjustment coefficient is set to the deviation rate multiplied by the gain factor plus one; if the deviation rate is below a preset lower threshold, the adjustment coefficient is set to one minus the deviation rate multiplied by the attenuation factor. Based on the adjustment coefficient, the sensor sampling frequency is adjusted. The new sampling frequency is equal to the current frequency multiplied by the adjustment coefficient. Updated real-time data for each joint, including torque, angle, and angular velocity values, are obtained according to the new sampling frequency. The updated real-time data is fed into the initial allocation architecture initially established by the system as a new input loop, and the torque allocation, deviation identification and correction process is re-executed to form a new round of coordination quantitative assessment report. Through continuous looping, the coordination of the whole body joints is continuously assessed.

[0063] Specifically, in one implementation, the process of constructing a coordination quantitative assessment report involves the integration and processing of multi-dimensional data.

[0064] Specifically, the theoretical torque values ​​of each joint are first extracted from the corrected full-joint torque distribution compensation scheme. These theoretical values ​​represent the load that each joint should bear under ideal coordination conditions. Simultaneously, the actual torque values ​​of each joint are collected in real time from torque sensors. The coordination index is calculated using the relative deviation method, i.e. ,in Let be the coordination index of the i-th joint. and These are the actual and theoretical torque values, respectively. Joint importance weights are pre-set based on the joint's position and load-bearing capacity in the kinetic chain: hip joint weight is 0.3, knee joint weight is 0.25, ankle joint weight is 0.2, shoulder joint weight is 0.15, and other small joints share the remaining weights. The weighted average is calculated as follows: This value serves as a quantitative score for overall coordination. The evaluation report uses a structured data format, including five fields: timestamp, joint number array, coordination index array for each joint, weighted average, and evaluation period identifier. The adjustment coefficient is determined following a segmented control principle to achieve fine-tuning of the sampling frequency. Deviation rate The calculation formula is ,in To quantify the coordination score, This sets the preset scoring boundaries. The preset upper threshold is typically set to 20%, and the lower threshold to -15%. When the deviation rate exceeds the upper threshold, it indicates that the system's coordination has seriously deviated from expectations, requiring faster data acquisition to capture more detailed changes. At this point, the adjustment coefficient... ,in This is the gain factor, ranging from 0.5 to 1.0. When the deviation rate is below the lower threshold, it indicates that the system's coordination is better than expected, and the sampling frequency can be appropriately reduced to save computational resources. (Adjustment coefficient) ,in This is the attenuation factor, with a value ranging from 0.2 to 0.4. When the deviation rate is between the upper and lower thresholds, the adjustment coefficient remains at 1, maintaining the current sampling frequency unchanged.

[0065] For example, the adaptive adjustment of the sensor sampling frequency directly affects the system's response speed and data accuracy. The baseline sampling frequency is set to 1000Hz, with the adjustment range limited to between 500Hz and 2000Hz. When the adjustment coefficient K=1.5, the new sampling frequency is 1500Hz, and data is collected every 0.67ms; when K=0.7, the new sampling frequency drops to 700Hz, and the system collects data every 1.43ms. Increasing the frequency captures more transient changes, helping to identify rapid deterioration in coordination; decreasing the frequency reduces data redundancy and lowers the processing burden. The updated real-time data includes not only torque values ​​but also joint angle and angular velocity values, which together constitute the input for the next round of evaluation.

[0066] Preferably, the sampling frequency is adjusted gradually to avoid the impact of sudden frequency changes on system stability. Each adjustment is limited to within 30% of the current frequency. If the calculated target frequency exceeds this range, the adjustment is performed gradually in multiple stages to reach the target frequency.

[0067] In one possible implementation, the cyclical evaluation mechanism combines event-driven and timed triggering. Under normal conditions, a complete evaluation cycle is executed every 100ms. When a sudden change in the coordination score exceeding 30% is detected, an emergency evaluation is triggered immediately, without waiting for the timed cycle. The cyclical process begins with updating real-time data input, sequentially going through the initial allocation architecture, torque allocation optimization, deviation identification, and torque reallocation, ultimately generating a new evaluation report. The evaluation report for each cycle is stored in a cyclic buffer, retaining the historical data from the last 20 cycles for trend analysis.

[0068] Understandably, the convergence judgment of continuous evaluation is based on the stability of scores over multiple rounds. When the consistency score fluctuation range of 5 consecutive evaluations is less than 2%, it is considered to have reached a stable state, at which point the evaluation period can be appropriately extended to 200ms. If the scores of 3 consecutive rounds show a monotonically increasing or decreasing trend, it is identified as a divergent state, and the sampling frequency is immediately increased to the maximum value, while the evaluation period is shortened to 50ms to strengthen the monitoring of the system status.

[0069] For example, when a robot is performing a turning maneuver while handling a liquid container, the sloshing of the liquid causes a rapid change in the center of gravity, and the coordination score may drop from 85 to 60 points in a short period of time. Upon detecting a sharp drop of 25 points, the sampling frequency is immediately increased from 1000Hz to 1800Hz, and the evaluation cycle is shortened to 50ms. After approximately 10 evaluation cycles of rapid adjustment, the torques of each joint regain balance, and the coordination score gradually recovers to above 80 points. Furthermore, the evaluation report includes an anomaly marker field to identify states requiring special attention. When the coordination index of a joint exceeds 50% for three consecutive rounds, or when multiple adjacent joints simultaneously show a deviation of more than 30%, an anomaly marker is added to the report. These markers trigger additional safety checks, limiting the robot's movement speed or load capacity if necessary to prevent mechanical damage.

[0070] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for real-time evaluation of the coordination of all joints in a humanoid robot, characterized in that, include: Acquire real-time data of each joint of the humanoid robot, determine the angle relationship between the center of mass offset and the forward direction, and construct the initial allocation architecture; Based on the initial allocation architecture, the joint data is filtered to generate denoised torque and rotation angle data. The torque distribution set of each joint is determined by the denoised torque and rotation angle data, and the complex quantitative evaluation value of whole-body joint compensation is analyzed. Based on the complex quantitative evaluation value, the torque distribution weight of each joint is dynamically adjusted to generate an optimized torque distribution scheme. Based on the optimized torque distribution scheme, the actual deviation is identified, the torque is redistributed to the weakly coordinated joint combination, and the corrected torque distribution compensation scheme is determined.

2. The real-time evaluation method for the coordination of the whole-body joints of a humanoid robot as described in claim 1, characterized in that, The process of acquiring real-time data of each joint of the humanoid robot, determining the angular relationship between the center of mass offset and the forward direction, and constructing an initial allocation architecture includes: Real-time torque values ​​are obtained from sensors deployed at each joint, rotation angle data are obtained from the inertial measurement unit, the angle between the center of mass offset direction and the forward direction is calculated, and the center of mass offset is determined by multiplying the cosine of the angle by the offset distance. Cross-validation was performed using the centroid offset and the torque value; Based on the comparison between the verified centroid offset and the preset boundary, the urgency level is divided and a urgency level label is formed. A hierarchical compensation architecture is constructed using the aforementioned urgency level identifiers, and different weight values ​​are assigned to establish an initial allocation architecture that includes joint identifiers, urgency levels, and compensation weights.

3. The real-time evaluation method for the coordination of the whole-body joints of a humanoid robot as described in claim 2, characterized in that, The step of filtering the joint data according to the initial allocation architecture to generate denoised torque and rotation angle data includes: The adaptive filtering parameters are calculated based on the urgency level and compensation weight. The filter window length and cutoff frequency of joints with different urgency levels are set to form a filter parameter configuration table. The filtering parameter configuration table is used to perform filtering on the real-time torque signal and rotation angle signal. Through state prediction and measurement update iteration calculation, the noise covariance is adjusted for joints with different urgency levels to obtain the preliminary filtered data sequence. By performing a sliding window averaging operation on the pre-filtered data sequence, setting the window size according to the urgency level, collecting data points at time intervals, and generating denoised and smoothed torque and rotation angle data arranged in time series.

4. The real-time evaluation method for the coordination of the whole-body joints of a humanoid robot as described in claim 1, characterized in that, The process of determining the torque distribution set of each joint using the denoised torque and rotation angle data, and analyzing the complex quantitative evaluation value of whole-body joint compensation, includes: A joint torque distribution matrix is ​​constructed based on the denoised torque and rotation angle. The denoised torque is then grouped using a clustering algorithm to form torque distribution sets of different levels. Principal component analysis is performed using the aforementioned torque distribution set to extract the main torque distribution directions, calculate the projection values ​​of each joint torque on the main torque distribution directions, and classify them into dominant joint groups; The degree of dispersion of the distribution is calculated by the standard deviation of the torque values ​​within the dominant joint group. The coupling coefficient is determined based on the ratio of torque to angle change rate of adjacent joints. The weighted sum of the degree of dispersion of the distribution and the coupling coefficient is used to calculate the complex quantitative evaluation value of whole-body joint compensation.

5. The real-time evaluation method for the coordination of the whole-body joints of a humanoid robot as described in claim 1, characterized in that, The step of dynamically adjusting the torque distribution weights of each joint based on the complex quantified evaluation value to generate an optimized torque distribution scheme includes: The weight adjustment parameters are determined based on the difference between the complex quantitative evaluation value and the preset boundary. The basic ratio of torque distribution for each joint is calculated by combining the weight adjustment parameters and the centroid offset. Calculate the expected torque value based on the basic ratio value, compare the difference between the expected torque value and the actual torque value, and output the optimized allocation scheme.

6. The real-time evaluation method for the coordination of the whole-body joints of a humanoid robot as described in claim 5, characterized in that, The calculation of the basic ratio of torque distribution for each joint by combining the weight adjustment parameters and the centroid offset includes: using the weight adjustment parameters, constructing a system of equations containing torque variables of each joint and centroid position constraints by combining the centroid offset, solving for the extreme values ​​of weight distribution by using the Lagrange multiplier method under the constraint that the centroid offset remains unchanged, and obtaining the basic ratio of torque distribution for each joint.

7. The real-time evaluation method for the coordination of the whole-body joints of a humanoid robot as described in claim 1, characterized in that, The step of identifying the actual deviation based on the optimized torque distribution scheme, redistributing torque to the weakly coordinated joint combinations, and determining the corrected torque distribution compensation scheme includes: The actual torque value is calculated based on the theoretical torque value and the smooth rotation angle in the optimized scheme, and the actual deviation is determined by the ratio of the difference between the theoretical value and the actual value. The actual deviation is compared with a preset threshold to identify weak joint combinations based on the joint kinematic chain topology. Based on the deviation characteristics of the weak joint combination, the deviations are prioritized according to their magnitude. An adjustment coefficient is determined according to the ratio of the magnitude of the deviation to the average deviation. The torque distribution ratio is adjusted, and the corrected compensation scheme is output.

8. The real-time evaluation method for the coordination of the whole-body joints of a humanoid robot as described in claim 1, characterized in that, The method further includes: using the corrected torque distribution compensation scheme and real-time data to generate a coordination quantification score, adaptively adjusting the data acquisition parameters, and cyclically updating the input to the initial allocation architecture to achieve whole-body joint coordination assessment.

9. The real-time evaluation method for the coordination of the whole-body joints of a humanoid robot as described in claim 8, characterized in that, The process of using the corrected torque distribution compensation scheme and real-time data to generate a coordination quantification score, and adaptively adjusting data acquisition parameters, includes: The compensation scheme after integration and correction is combined with real-time torque data. The ratio of the deviation between theoretical torque and actual torque is calculated as a coordination index. The index is then weighted and averaged according to the importance of the joints to generate a quantitative assessment report on coordination. The weighted average value is extracted from the coordination quantitative assessment report as the coordination quantitative score, and the deviation rate is calculated by comparing the coordination quantitative score with the preset boundary. The data acquisition frequency adjustment amount is determined based on the deviation rate, and the frequency is updated to obtain real-time data.

10. The method for real-time evaluation of the coordination of the whole-body joints of a humanoid robot as described in claim 8, characterized in that, The cyclical update input to the initial allocation architecture enables a full-body joint coordination assessment, including: The real-time data obtained from the data acquisition frequency updated according to the adjustment coefficient is used as the new input and sent to the initial allocation architecture; The torque allocation and deviation identification process is re-executed using the initial allocation architecture to generate a new round of coordination quantitative assessment report; By comparing the coordination score in the new round of evaluation report with the preset boundary, the data acquisition parameters are continuously adjusted, and dynamic evaluation of the coordination of the humanoid robot's whole-body joints is achieved through continuous iteration.

Images