A method and system for online identification of load mass and center of gravity in dual-arm closed-chain material handling scenarios

Abstract

This invention belongs to the field of robot dynamics modeling and online parameter identification, and specifically discloses a method and system for online identification of load mass and center of mass in a dual-arm closed-chain handling scenario. The method includes: estimating joint friction torque based on the joint feature vectors under robot load conditions, and removing it from the measured joint torque under load conditions to obtain the friction-free joint torque, thereby obtaining the joint residual torque; introducing an internal force suppression regularization term through an internal force suppression coefficient to construct a contact force screw solution model; adaptively determining the internal force suppression coefficient based on the joint residual torque and solving the contact force screw solution model to obtain the net external force screw of the object; outputting the joint observation reliability based on the current joint state, propagating errors through closed-chain force mapping, and constructing a task space weight matrix; obtaining the load mass and center of mass position based on the task space weight matrix and the net external force screw of the object. This invention can achieve high-precision identification of load mass and center of mass position.

Description

Technical Field

[0001] This invention belongs to the field of robot dynamics modeling and online parameter identification, and more specifically, relates to a method and system for online identification of load mass and center of mass in a dual-arm closed-chain transport scenario. Background Technology

[0002] Dual-arm robots are widely used in heavy-duty handling, medical transport, and disaster relief due to their high load-bearing capacity and maneuverability. In these tasks, the robot often needs to form a closed-loop structure with the load for stable gripping. To ensure motion stability during handling, accurately distribute torque, and maintain the balance of the entire machine (such as a humanoid robot), the system needs to accurately obtain the dynamic parameters of the load, especially its mass and center of gravity.

[0003] Currently, some research has been conducted on parameter identification for robot end-effector loads. The mainstream method is based on the robot's dynamics model, using joint torque sensor data or end-effector force and torque sensor data (F / T) to construct a linear regression equation for the load parameters. Algorithms such as least squares (LS) or weighted least squares (WLS) are then used to estimate the load's mass and centroid online or offline. These methods are well-established in single-arm robots or open-chain operation scenarios and can meet general grasping and movement requirements.

[0004] Furthermore, some studies have proposed reconstructing joint torques under quasi-static conditions of single-arm open-chain carrying, using information such as encoder errors, deriving a closed-loop linear relationship between joint torques and load mass and centroid parameters, and introducing an attention mechanism in least-squares identification to adaptively generate weight matrices for different joints (or different sample-joint pairs), thereby improving the estimation accuracy and stability of weighted least squares under noisy and under-excited configurations. The weights in this type of method typically act directly on the joint space residuals. Although effective in open-chain carrying tasks, its direct transfer to dual-arm closed-chain handling scenarios remains limited because it does not explicitly model the coupling between internal force components and torque distribution under dual-arm closed-chain gripping, and lacks a mechanism to propagate the reliability of joint observations to the uncertainty of the object's net external force spinor (task space).

[0005] In summary, directly applying the identification methods based on single-arm or open-chain models to dual-arm closed-chain handling scenarios has significant limitations: (1) Closed-chain internal force coupling problem: When the two arms rigidly grasp the object to form a closed chain, internal forces will be generated inside the system that do not affect the motion of the object but significantly affect the distribution of joint torques. Traditional single-arm identification methods cannot effectively decouple the internal forces from the gravity and inertial forces generated by the load, resulting in distortion of the identification model and a significant decrease in identification accuracy.

[0006] (2) Lack of robust identification mechanism for task space in closed-loop systems: Although some existing works have introduced attention mechanisms to evaluate the reliability of observation data, these methods are usually limited to direct weighting in joint space. In a two-arm closed-loop system, joint torques are also affected by the coupling of internal forces within the closed loop. If weights are only assigned in joint space, it is impossible to distinguish between load torques and internal force disturbances, and the error propagation characteristics from joint space to the object's task space are ignored. This weighting strategy, which does not consider the geometric constraints of the closed loop, is difficult to accurately reflect the uncertainty of the net external force rotation of the object under complex working conditions with multiple posture changes, resulting in a lack of robustness in the identification results.

[0007] (3) Joint friction and noise interference: The joint torque measurements of actual robots contain significant nonlinear friction torque and noise. Existing methods often lack independent modeling and compensation for joint friction. During heavy-load, low-speed handling, a high proportion of friction torque will be incorrectly attributed to the load, resulting in systematic deviations in mass and centroid estimation. Summary of the Invention

[0008] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method and system for online identification of load mass and center of mass in a dual-arm closed-chain handling scenario. The purpose is to improve the identification accuracy of unknown load mass and center of mass position when the robot is carrying out dual-arm closed-chain handling.

[0009] To achieve the above objectives, according to one aspect of the present invention, a method for online identification of load mass and center of mass in a dual-arm closed-chain transport scenario is proposed, comprising the following steps: Based on the joint feature vectors under robot load conditions, estimate the joint friction torque; remove the joint friction torque from the measured joint torque under load conditions to obtain the friction-free joint torque; Based on the friction-reducing joint torque, the equivalent joint torque component caused by the load is obtained, namely the joint residual torque; In the quadratic programming problem concerning joint residual torque and contact force spinor, the internal force suppression coefficient is used... An internal force suppression regularization term is introduced to construct a solution model for the contact force screw. Based on the joint residual torque, the internal force suppression coefficient is adaptively determined and the contact force screw solution model is solved to obtain the contact force screw, and then the net external force screw of the object is obtained. The joint observation confidence is output based on the current joint feature vector, and then error propagation is performed through closed-chain force mapping to construct the task space weight matrix; the load mass and center of mass position are obtained based on the task space weight matrix and the net external force spinor of the object.

[0010] As a further preferred embodiment, the contact force spinor solution model is expressed as follows:

[0011] in, For joint residual torque, For contact force spinor, To solve for the obtained contact force spinor, Represents the L2 norm; The internal force projection matrix, For closed-loop internal forces of the system; , , These are the Jacobian matrices of the left and right arms in the base system, respectively.

[0012] As a further preferred method, the adaptive method for determining the internal force suppression coefficient is as follows: within the range of internal force suppression coefficients that satisfy the fitting constraints, select the internal force suppression coefficient with the smallest physical consistency residual. The fitting constraint is expressed as follows:

[0013]

[0014] in, This represents the joint fitting error. This is the empirical tolerance coefficient. This is the set of candidate internal force suppression coefficients. for The internal force suppression coefficient in; This represents the total number of pose samples within the sliding window. , , The first Joint residual moments and matrices corresponding to each sample pose Contact force spinor; For positive numbers that are numerically stable.

[0015] As a further preferred embodiment, the formula for calculating the physical consistency residual is:

[0016] in, For physical consistency residuals, It is a diagonal weight matrix. and They are respectively to The vector and matrix formed by stacking the net external force spin of the object and the static gravity regression matrix of each sample in columns. For candidate internal force suppression coefficient The estimated load parameter vector is obtained below.

[0017] As a further preferred method, the task space weight matrix is ​​constructed as follows: Based on the reliability of joint observation The joint residual noise covariance was obtained. This leads to the net external force spin covariance. , This is the closed-chain mapping matrix between the net external force spin of the object and the joint residual torque; Construct the task space weight matrix , For positive numbers that are numerically stable, It is a 6th order identity matrix.

[0018] As a further preferred option, based on a pre-trained attention network, the confidence level of joint observation is output according to the current joint feature vector; The training method for the attention network includes: Obtain the net external force spin of the object under different real loads, and construct the corresponding spin based on the robot's closed-chain kinematics and geometric contact information. , This is the static gravity regression matrix; The attention network is used for forward inference with joint feature vectors as input and joint observation confidence as output; the mapping matrix between joint observation confidence and closed chain is based on the output of the attention network. Calculate the task space weight matrix, and then combine it with the net external force spinor of the object and the static gravity regression matrix. Obtain load parameter estimates and with known actual load parameters As supervision, the weighted loss of relative mass error and centroid error is minimized to complete the training of the attention network.

[0019] As a further preferred method, the load mass and center of mass position are obtained based on the task space weight matrix and the net external force spinor of the object, including: Within a multi-pose sampling sliding window, the following linear identification model of the object's net external force spinor and load parameter vector is constructed:

[0020] in, For load parameter vectors, , For load quality, The location of the center of mass; This represents the total number of pose samples within the sliding window. Represents the L2 norm; For the first Net external force rotation of an object in a given orientation For the first Task space weight matrix for each pose; Based on the task space weight matrix and the net external force spinor of the object, the load parameter estimates are obtained by solving the linear identification model using the weighted least squares method. .

[0021] As a further preferred embodiment, the joint feature vector Specifically:

[0022]

[0023]

[0024] in, , , The first Standardized joint position, joint tracking error, and joint velocity for each joint. , For the first One-hot encoding of the velocity symbols of each joint. For the first Joint velocity of each joint.

[0025] As a further preferred embodiment, the joint friction torque is estimated using a pre-trained friction torque model based on the joint feature vectors under robot load conditions. The training method for the friction torque model includes: Obtain joint feature vectors and measure joint torques in the robot under no-load conditions. ; To measure joint torque Inverse dynamic torque calculated with robot dynamics model The difference is the joint friction torque. ; to reduce joint friction torque Standardization yields joint feature vectors. Corresponding friction label This allows us to construct a training set. The friction torque model based on the multilayer perceptron is trained using the training set to obtain a trained friction torque model.

[0026] According to another aspect of the present invention, an online load mass and centroid identification system for dual-arm closed-chain transport scenarios is provided, including a processor, the processor being used to execute the above-described online load mass and centroid identification method for dual-arm closed-chain transport scenarios.

[0027] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages: 1. This invention determines the joint residual torque after removing noise interference from joint friction torque in the measurement of joint torque; then, it introduces an internal force suppression regularization term and an adaptive internal force suppression coefficient to effectively decouple the closed-chain internal forces formed when the two arms rigidly grasp an object in the joint torque, and calculates a more accurate net external force spin; then, it combines the task space weight matrix based on the joint observation credibility to achieve high-precision identification of load mass and centroid position.

[0028] 2. By introducing an internal force suppression regularization term and combining it with an adaptive selection mechanism for the internal force suppression coefficient that prioritizes physical consistency and imposes fitting constraints, the contamination of joint torque by closed-chain internal forces can be effectively suppressed. At the same time, the distortion of net force spinor caused by excessive suppression can be avoided, making the reconstructed net external force spinor of the object more stable, interpretable, and easier to identify in the future.

[0029] 3. Joint reliability is evaluated through an attention mechanism and mapped to net external force spinor uncertainty based on error propagation, thereby adaptively generating a task space weight matrix. This can automatically emphasize the contribution of attitude or spinor components with high information content and low noise in multi-pose identification, thereby improving robustness and generalization ability under unknown loads, different attitudes and different friction states.

[0030] 4. Data-driven joint friction compensation can significantly reduce the systematic bias in joint residual torque and improve the accuracy of subsequent external torque reconstruction.

[0031] 5. This invention has a low overall computational load. The core operations can be reduced to solving low-dimensional linear equations and small-scale network forward inference, making it suitable for online deployment in embedded or real-time control systems. During the robot's handling tasks, there is no need to retrain the model for the current unknown load. Relying only on the joint friction compensation model and task space weight generation model obtained through offline training, parameter updates can be completed within each control cycle or a preset update cycle. Attached Figure Description

[0032] Figure 1 Flowchart of the online identification method for load mass and center of mass in a dual-arm closed-chain handling scenario provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of joint friction compensation provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of adaptive internal force suppression and net external force rotation reconstruction of an object provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the task space weighted least squares identification process based on error propagation provided in an embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0034] This invention provides an online method for identifying load mass and center of gravity in a dual-arm closed-chain material handling scenario, such as... Figure 1 As shown, it includes the following steps: (1) Data acquisition and preprocessing: Two types of data, no-load and with-load, are collected in advance. The collected raw data are subjected to validity screening (e.g., removing abnormal postures based on the geometric relationship between the hands and the object), time alignment and standardization (e.g., zero-mean unit variance standardization for each joint), and network input features (joint position, error, velocity and velocity sign, etc.) are constructed.

[0035] (2) Joint friction compensation (Module A): The friction torque model is obtained by offline training using no-load data. At this stage, since there is no load and no constraint is formed on the object by the double-arm closed chain, the joint torque does not include the internal force term generated by the contact between the double-arm and the object closed chain, which is suitable for learning the friction term. After training, the friction torque model is applied to the loaded data, that is, based on the joint motion state data under the load state, the joint friction torque is output; then the friction component is removed from the measured joint torque to obtain the friction-compensated joint torque (friction-free joint torque).

[0036] (3) Adaptive internal force suppression and net external force rotation reconstruction (Module B): Based on the friction-reducing joint torque and the residual of the robot inverse dynamics model (joint residual torque) A contact force screw solution model with closed-chain constraints is constructed. Under the bi-arm closed chain, the joint residuals simultaneously contain external force components that generate the net external force (screw) of the object and internal force components that circulate within the closed chain. Therefore, based on the joint residual fitting, an internal force suppression regularization term based on the null space of the grasping matrix is ​​introduced to reconstruct the contact force screw and obtain the net external force screw of the object. Furthermore, the consistency and interpretability of the static gravity model are used as evaluation criteria and physical priors to adaptively select the internal force suppression coefficient. This allows for the suppression of the influence of closed-chain internal forces without significantly sacrificing joint space fitting, resulting in a more physically consistent net external force spinor of the object.

[0037] (4) Task Space Weighted WLS Identification Based on Error Propagation (Module C): Within the multi-pose sampling window, a linear identification model is constructed for the net external force spinor and the parameters of the load mass and centroid. The joint observation confidence is output through an attention mechanism, and the joint observation confidence is propagated into the uncertainty of the net external force spinor (net external force spinor covariance) using closed-chain force mapping. Its inverse is used as the task space weight matrix. Weighted least squares (WLS) is performed within the multi-pose window to solve for the load mass and centroid, thereby improving the estimation robustness and generalization ability.

[0038] Specifically, the following explanation uses a two-arm system with 7 degrees of freedom each, totaling 14 degrees of freedom, as an example; and to avoid ambiguity, the following conventions are given in advance: (a) Both arms One degree of freedom, in this example, is taken as... Let the joint vector be:

[0039] (b) The object coordinate system is denoted as The robot base coordinate system is denoted as . Let represent the rotation matrix from the system of objects to the base system. Then, the spinor rotation transformation from the base system to the system of objects is:

[0040] (c) The contact force spinor is stacked in the order of "force-torque" and is defined as the contact force spinor applied by the robot to the object:

[0041] Its concatenation vector is:

[0042] (d) The net external force spin at the object's reference point (the origin of the object's coordinate system) is denoted as:

[0043] The relationship between the two is as follows:

[0044] (e) The gravitational acceleration vector in the base frame is denoted as:

[0045] Considering that the contact force spinor is in the direction of "robot supporting the object", for ease of static modeling, the upward support direction is defined as follows:

[0046] (f) Covariance and Weights: The joint residual noise covariance is denoted as... The net external force rotation covariance is denoted as The task space weight matrix is ​​denoted as .

[0047] Furthermore, regarding the joint friction compensation in step (2), such as Figure 2 As shown, the goal of module A is to learn the joint friction torque. And from measuring joint torque After removing the friction-free joint torques, we obtain the friction-free joint torques used for subsequent closed-loop mechanics solutions. .

[0048] (a) No-load data and feature construction. For each time-series sample obtained from the no-load trajectory acquisition, the positions of the biarm joints are extracted. Desired joint position Joint velocity And measuring joint torque wait.

[0049] Define joint tracking error:

[0050] And on , , Standardize according to joints to obtain , , Simultaneously, construct a one-hot encoding for the velocity symbol, for the... One joint:

[0051] The input feature vector for each joint is defined as:

[0052] (b) Joint torque decomposition and friction label. Measurement of joint torque. It can be represented as:

[0053] in The inverse dynamic torque (including gravity, inertia, and Coriolis terms, etc.) is calculated from the robot's dynamics model. The equivalent joint torque caused by external load, For joint friction torque, This is for measuring noise and unmodeled terms.

[0054] In unloaded data, it can be considered that The friction label would then be:

[0055] After standardization, friction labels are .

[0056] (c) Friction torque model structure. Establish the friction torque model. Its features are based on the characteristics of each joint. As input, output a standardized prediction of the joint friction torque. As one embodiment, the friction torque model employs a multilayer perceptron structure of "independent encoder per joint + independent output layer per joint": the encoder of each joint will... The data is mapped to a low-dimensional embedding and then passed through a linear layer to obtain friction predictions. This structure can either share some parameters (improving generalization) or remain completely independent (facilitating the decoupling of joint differences).

[0057] In some embodiments, the friction torque model can be replaced by a physical parameterization model (Coulomb friction + viscous friction + Stribeck term) or other machine learning models (such as Gaussian process, recurrent neural network, etc.); the input features can also include information such as motor current, temperature or load direction to improve generalization across operating conditions.

[0058] (d) Training and application of the friction torque model. The friction torque model is trained on an unloaded dataset by minimizing the mean square error loss:

[0059] in This represents the number of training samples.

[0060] Standardized friction predictions output by the friction torque model Obtained by inverse standardization of the mean and variance during training. After training, the friction torque model is applied to loaded data, allowing estimation of joint friction torque based on joint motion data under load. This leads to the friction-compensated joint torque, i.e., the friction-reduced joint torque. :

[0061] Furthermore, regarding the adaptive internal force suppression and net external force spinor reconstruction in step (3), such as Figure 3 As shown, under closed-chain gripping conditions, the joint residual torque contains both external force components that generate the net external force spin of the object and internal force components circulating within the closed chain. By introducing an internal force suppression regularization term based on the inverse dynamics residual fitting, the decoupling of the contact force spin and the reconstruction of the net external force spin of the object are achieved. The consistency of the static gravity model and the joint fitting error are used as evaluation criteria to adaptively select the internal force suppression coefficient. .

[0062] (a) Relationship between closed-chain forces and joint residuals. Define the spinor of the left and right hand contact forces and the splicing vector as follows:

[0063]

[0064] The equivalent joint torque component caused by external load is defined as the joint residual torque. ,but:

[0065] Under the assumptions of rigid gripping or small deformation, the joint residuals and contact force spin satisfy the following:

[0066] in , Let be the Jacobian matrix of the ends of the left and right arms in the basis system.

[0067] (b) Grasp the matrix and the net external force spinor of the object. Let the position of the object's reference point in the base frame be... The left and right contact points are located in the base system. , Then the relative position:

[0068] Net external force rotation Contact force spinor satisfy:

[0069] The capture matrix It can be written as:

[0070] In the formula for Unit array, An antisymmetric matrix of cross product (satisfying) Object pose , and the location of the contact point , It can be obtained by visual localization (such as marker point or point cloud registration), solution of known object geometric model combined with end pose, or by tactile or contact estimation algorithm; in this embodiment, it can be regarded as known input.

[0071] (c) Internal force subspace. A closed-loop internal force is defined as a force component that does not change the net external force spin of the object but circulates within the closed loop. Let:

[0072] in for Moore-Penrose pseudo-reverse, for Unit array. Then the internal force components. It can be represented as:

[0073] (d) Optimization solution with internal force suppression regularization term. To simultaneously fit the joint residuals and suppress internal forces, solve the following quadratic programming problem:

[0074] This problem can be transformed into solving a linear equation:

[0075] in Small positive numbers used for numerical stability. Calculated. Then, further calculations can be performed:

[0076] Further converting the net external force spin to the material system:

[0077] (e) Interpretability based on static gravity model Adaptive selection. In the quasi-static phase of dual-arm closed-chain transport, the net external force rotation of the object is mainly determined by the gravity term; therefore, a set of net external force rotations sampled under multiple attitudes for the same unknown load should be consistently interpretable by the same set of mass and center-of-mass parameters in the static gravity model. This invention utilizes this "interpretability" as... Selection criteria.

[0078] Specifically, data is collected within a sliding window. From pose samples, we obtain .in Indicates the use of candidates Solve for the net external force rotation of the object. This is the corresponding static gravity regression matrix (the specific calculation method will be described in the next section). To account for the dimensional differences between the equilibrium force and torque components, a diagonal weight matrix can be introduced:

[0079] in , This is a preset constant used for balancing the dimensions of force and torque. Then, for each candidate... First, solve the following:

[0080] Redefining physical consistency residuals (interpretable error):

[0081] in and They are respectively to A vector or matrix formed by stacking samples column by column Small positive numbers used for numerical stability.

[0082] Simultaneously define the joint fitting error (used for constraints). (To avoid being too large and causing underfitting)

[0083] Finally, the optimal solution is selected using the criterion of "prioritizing physical consistency and constraining fit". That is, from the candidate set that satisfies the following fitting constraints, priority is given to selecting... The smallest (if necessary, further additions can be made such as "a smaller proportion of internal force") (Refine the selection based on parallel criteria such as "smaller")

[0084] in as a candidate Sets, which can take logarithmically uniform networks; This is the empirical tolerance coefficient.

[0085] In some embodiments, the optimization problem of module B can be further constrained by contact constraints (friction cone, non-negativity of normal force, end moment limiting, etc.) or solved using constrained quadratic programming; internal force suppression can also be achieved using an equivalent form of explicit decomposition. And on Regularization is used to suppress internal forces; The selection can also be updated adaptively in different windows or different task stages.

[0086] Furthermore, for step (4) of the task space weighted least squares identification based on error propagation, such as Figure 4 As shown, the net external force spinor obtained by reconstructing module B is used to identify the load mass and centroid within the multi-pose window; and the joint credibility is learned through an attention mechanism, combined with closed-chain force mapping for error propagation, and the task space weight matrix is ​​adaptively constructed to improve robustness.

[0087] (a) Static Gravity Model and Parameterization. In the quasi-static stage, the net external force spin of the object is mainly determined by the gravity support term. Define the support direction vector under the object system:

[0088] in, Let be the gravitational acceleration vector in the base frame. For the first k The rotation matrix from the object system to the base system under each attitude; The net force on the object can then be approximated as:

[0089] If the pose of the load's center of mass relative to the object's reference point (the object's origin) is defined as follows: Then the torque produced by gravity is:

[0090] To maintain linearity, a parameter vector can be defined. This indicates the load quality. And the product of mass and the position of the center of mass:

[0091] (b) Linear regression form and WLS solution. There exists a relationship regarding... linear mapping matrix Make:

[0092] in It can be written as:

[0093] In certain symmetrical load scenarios, further assumptions can be made. Then the parameters can be simplified to , Correspondingly degenerate into matrix.

[0094] In a containing Within a sliding window of pose samples, and Stacked as:

[0095] And construct the block diagonal weight matrix:

[0096] The weighted least squares problem (linear identification model) is then:

[0097] Its closed-form solution is:

[0098] Depend on Further results can be obtained:

[0099] in To prevent Positive decimals that are too small, resulting in unstable values.

[0100] (c) Attention mechanism outputs joint reliability. To characterize the differences in observation reliability across different joints and poses, an attention network is constructed. For each pose Each joint Output credibility As one embodiment, the attention network uses joint feature vectors The input (features consistent with or shared with module A) is obtained through encoding and scoring. Then, after Sigmoid mapping:

[0101] In other embodiments, softmax, normalized sigmoid, or other methods can be used to map the scores to... The function implements the credibility output.

[0102] (d) Construct the task space weight matrix based on error propagation. Characterize the accuracy of joint residual noise and define the joint noise variance:

[0103] in Empirical constants are used to align the dimensions of the equations. The joint residual noise covariance is then constructed. :

[0104] On the other hand, from the linear solution process of module B, we can obtain: given When module B is adaptively selected, there is a linear relationship between the net external force rotation of the object and the joint residuals:

[0105] Where the closed-chain mapping matrix for:

[0106] Therefore, the net external force spinor covariance can be obtained from error propagation. :

[0107] And take the task space weight matrix for:

[0108] in , , Small positive numbers used for numerical stability. for A symmetric positive definite matrix can simultaneously characterize the uncertainties and correlations of force and torque components; a block diagonal weight matrix can be constructed within a sliding window. Then substitute them into the weighted least squares to solve for the load parameters.

[0109] (e) Attention Network Training. The attention network is pre-trained offline on several known load samples: For each training window, the net external force spinor of the object is first obtained through reconstruction. and construct the corresponding Static gravity regression matrix Its function is to connect the load parameters with the net external force spinor of the object, describing the object's own physical properties. Its value and calculation directly depend on the object's spatial orientation and gravity direction in the base coordinate system; closed-chain mapping matrix. Its function is to connect the joint residual torque and the net external force spinor of the object, describing the coupling properties of the entire system. Its value is determined by the robot joint position and the position of the object grasped by the robot, which is calculated by the Jacobian matrix A (related to joint position), the grasping matrix G and the internal force projection matrix N.

[0110] Attention networks use joint feature vectors Input: Joint observation confidence level , and then combine Calculate the task space weight matrix Furthermore, combining the net external force rotation of the object with Load parameter estimates are obtained by solving a linear identification model. And based on the actual load parameters. As supervision, the weighted loss of relative quality error and centroid error is minimized, enabling the attention network to learn joint confidence assignments that reduce the final recognition error.

[0111] After training is complete, the online phase only requires forward inference based on the real-time joint states to obtain the reliability of joint observations. And combined with current closed-loop geometric calculations ,structure Finally, by combining the net external force spinor of the object to solve the linear identification model, robust online identification under multi-pose conditions can be achieved.

[0112] In some embodiments, the parameter vector of module C can be expanded from mass + center of mass to complete inertial parameters (mass, center of mass, moment of inertia) or consider linear and angular acceleration terms of the object to adapt to more dynamic handling processes; correspondingly It can be constructed from the equations of rigid body dynamics. Furthermore, the task space weight matrix... It can be obtained based on error propagation, and is generally symmetric positive definite. Dense matrices; in scenarios with limited computational resources or where only independent noise components are of concern, their diagonal or block diagonal forms can be approximated. Attention networks can also be jointly trained end-to-end with WLS solvers, or retrained offline in different tasks to adapt to new noise distributions and grasping methods.

[0113] The following are specific examples: This embodiment uses a humanoid robot model with two arms, each with 7 degrees of freedom (14 degrees of freedom in total), to perform a closed-chain grasping and handling task in a simulation platform. The robot's left and right arms form rigid grasping constraints on the same object being handled, creating a closed-chain system between the robot and the object. The object's reference point is taken as the origin of the object's coordinate system, denoted as . The robot base coordinate system is denoted as During the transport process, the robot provides quasi-static support to the load in multiple different postures (i.e., the linear and angular velocities of the object are relatively small, and the inertia term is relatively negligible) to meet the applicable conditions of the static gravity model for mass and center of mass identification.

[0114] A no-load dataset and a with-load dataset containing multiple weights and centroids of varying masses were collected. The no-load data was used in module A to learn joint friction torques; the with-load dataset was used in module B to reconstruct the net external force rotation of the object, and also for parameter identification training and testing in module C. The sampled data included: bi-arm joint positions. Desired joint position Joint velocity Measuring joint torque and the inverse dynamic torque calculated from the robot dynamics model In the loaded data, the contact point position vectors required for the grasping matrix are further calculated based on the geometric relationship between the hands and the object. , And calculate the terminal Jacobian matrix based on the current configuration. , To ensure sufficient observational information, this example uses 16 excitation poses (this example employs a fixed-length window, which can be considered a special case of a sliding window). It should be noted that designing more effective excitation poses on top of this will yield more observational information and further improve estimation accuracy.

[0115] To avoid data leakage, data is numbered by load. The loaded dataset is divided into training, validation, and test sets. The training set contains five types of loads used for model learning, the validation set contains two types of loads used to monitor training and avoid overfitting, and the test set contains three types of loads used to evaluate the model's generalization ability on unseen loads. The evaluation metric is the average quality error estimated from the three unseen loads. and average centroid error .

[0116] Three comparative experiments were set up around modules A, B, and C, with each comparison changing only the corresponding part of the target module while keeping the other parts the same. The method of this invention (Ours, A+B+C): friction compensation (A) + closed-loop internal force suppression and object net external force spinor reconstruction (B) + task space weighted WLS based on error propagation (C).

[0117] Experiment 1: Comparison of mass and centroid estimation results between the method of this invention and the baseline Unweighted method under no-load conditions. The baseline Unweighted method, as a traditional least-squares baseline method, follows the same procedure as the method of this invention (still using friction compensation in A and net external force spinor reconstruction in B), but does not employ the error propagation weighting of module C, allowing... This is used to verify the contribution of module C.

[0118] Experiment 2: Ablation experiments were conducted to verify the contribution of module A. The ablation experiments were configured to disable module A, directly using measured joint torques for subsequent identification (the contribution of module C has been verified in Experiment 1 (Unweighted comparison); the contributions of module B and its synergistic contribution with module C will be further explained in Experiment 3 in the comparison with existing joint-space weighted methods).

[0119] Experiment 3: Comparison of the mass and centroid estimation results of the method of this invention with existing joint-space weighted least squares methods under unseen loads. To further demonstrate that the present invention is more applicable to closed-loop scenarios with two arms, an existing joint-space weighted least squares method based on an attention mechanism is set up for comparison. For a fair comparison, this comparison method also uses the joint torque (A) after friction compensation, but does not perform net external force spinor reconstruction and internal force processing for module B, nor does it perform task space error propagation weighting for module C. Instead, it directly constructs regression in joint space and solves it based on the attention mechanism. Note: Since the weight construction of module C depends on the net external force spinor output by module B and its mapping relationship, module B and module C have a cooperative relationship in closed-loop scenarios. Therefore, this experiment is used to compare with existing joint space methods to reflect the role of module B (and its cooperation with C).

[0120] Table 1 presents the results of Experiment 1. As can be seen, compared to the unweighted baseline, the task space weighting method of this invention reduces the average mass error from 0.3380 kg to 0.0746 kg (approximately a reduction of 77.9%) and the average centroid error from 0.0443 m to 0.0086 m (approximately a reduction of 80.6%). The results indicate that in multi-pose sampling of bi-arm closed-chain transport, there are significant differences in the reliability of joint observations and the noise level of the net external force spinor components under different poses. This invention evaluates joint reliability through an attention mechanism and constructs task space weights through error propagation, making WLS emphasize observations with higher information content and lower noise, thereby improving robustness and generalization ability on unseen loads.

[0121] Table 1: No results were found for mass and centroid estimation on the load.

[0122] Table 2 presents the results of Experiment 2. As can be seen, after disabling module A (B+C), the average mass error and centroid error increased, indicating that the joint friction torque accounts for a high proportion of the joint residuals and has a systematic bias. If friction compensation is not performed first, it will be difficult to distinguish between external load and friction, thus causing significant deviations in mass and centroid identification.

[0123] Table 2 Contribution of Ablation Experiment Validation Module A

[0124] Table 3 presents the results of Experiment 3. It can be seen that the proposed method demonstrates superior estimation accuracy when faced with three types of unseen loads. This indicates that in the dual-arm closed-chain scenario, if regression is constructed and weighted directly in the joint space, the joint residual error caused by the coupling of internal forces within the closed chain may still be misattributed to load parameters. In contrast, this invention explicitly suppresses the null-space internal forces of the grasping matrix through module B and performs uncertainty modeling and weighted identification of the net external force spinor in the task space, which better conforms to the physical constraints of the closed chain and thus exhibits better identification accuracy and stability.

[0125] Table 3 Comparison with Joint-space Weighted Least Squares

[0126] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for online identification of load mass and center of mass in a dual-arm closed-chain material handling scenario, characterized in that, Includes the following steps: Based on the joint feature vectors under robot load conditions, estimate the joint friction torque; remove the joint friction torque from the measured joint torque under load conditions to obtain the friction-free joint torque; Based on the friction-reducing joint torque, the equivalent joint torque component caused by the load is obtained, namely the joint residual torque; In the quadratic programming problem concerning joint residual torque and contact force spinor, the internal force suppression coefficient is used... An internal force suppression regularization term is introduced to construct a solution model for the contact force screw. Based on the joint residual torque, the internal force suppression coefficient is adaptively determined and the contact force screw solution model is solved to obtain the contact force screw, and then the net external force screw of the object is obtained. The joint observation confidence is output based on the current joint feature vector, and then error propagation is performed through closed-chain force mapping to construct the task space weight matrix; the load mass and center of mass position are obtained based on the task space weight matrix and the net external force spinor of the object.

2. The method for online identification of load mass and center of mass in a dual-arm closed-chain transport scenario as described in claim 1, characterized in that, The contact force spinor solution model is expressed as follows: in, For joint residual torque, For contact force spinor, To solve for the obtained contact force spinor, Represents the L2 norm; The internal force projection matrix, For closed-loop internal forces of the system; , , These are the Jacobian matrices of the left and right arms in the base system, respectively.

3. The method for online identification of load mass and center of mass in a dual-arm closed-chain transport scenario as described in claim 2, characterized in that, The adaptive method for determining the internal force suppression coefficient is as follows: within the range of internal force suppression coefficients that satisfy the fitting constraints, select the internal force suppression coefficient with the smallest physical consistency residual. The fitting constraint is expressed as follows: in, This represents the joint fitting error. This is the empirical tolerance coefficient. This is the set of candidate internal force suppression coefficients. for The internal force suppression coefficient in; This represents the total number of pose samples within the sliding window. , , The first Joint residual moments and matrices corresponding to each sample pose Contact force spinor; For positive numbers that are numerically stable.

4. The method for online identification of load mass and center of mass in a dual-arm closed-chain transport scenario as described in claim 3, characterized in that, The formula for calculating the physical consistency residual is: in, For physical consistency residuals, It is a diagonal weight matrix. and They are respectively to The vector and matrix formed by stacking the net external force spin of the object and the static gravity regression matrix of each sample in columns. For candidate internal force suppression coefficient The estimated load parameter vector is obtained below.

5. The method for online identification of load mass and center of mass in a dual-arm closed-chain handling scenario as described in claim 1, characterized in that, The method for constructing the task space weight matrix is ​​as follows: Based on the reliability of joint observation The joint residual noise covariance was obtained. This leads to the net external force spin covariance. , This is the closed-chain mapping matrix between the net external force spin of the object and the joint residual torque; Construct the task space weight matrix , For positive numbers that are numerically stable, It is a 6th order identity matrix.

6. The method for online identification of load mass and center of mass in a dual-arm closed-chain handling scenario as described in claim 5, characterized in that, Based on a pre-trained attention network, the confidence level of joint observation is output according to the current joint feature vector. The training method for the attention network includes: Obtain the net external force spin of the object under different real loads, and construct the corresponding spin based on the robot's closed-chain kinematics and geometric contact information. , This is the static gravity regression matrix; The attention network is used for forward inference with joint feature vectors as input and joint observation confidence as output; the mapping matrix between joint observation confidence and closed chain is based on the output of the attention network. Calculate the task space weight matrix, and then combine it with the net external force spinor of the object and the static gravity regression matrix. Obtain load parameter estimates and with known actual load parameters As supervision, the weighted loss of relative mass error and centroid error is minimized to complete the training of the attention network.

7. The method for online identification of load mass and center of mass in a dual-arm closed-chain handling scenario as described in claim 6, characterized in that, The load mass and center of mass position are obtained based on the task space weight matrix and the net external force spinor of the object, including: Within a multi-pose sampling sliding window, the following linear identification model of the object's net external force spinor and load parameter vector is constructed: in, For load parameter vectors, , For load quality, The location of the center of mass; This represents the total number of pose samples within the sliding window. Represents the L2 norm; For the first Net external force rotation of an object in a given orientation For the first Task space weight matrix for each pose; Based on the task space weight matrix and the net external force spinor of the object, the load parameter estimates are obtained by solving the linear identification model using the weighted least squares method. .

8. The method for online identification of load mass and center of mass in a dual-arm closed-chain transport scenario as described in claim 1, characterized in that, The joint feature vector Specifically: in, , , The first Standardized joint position, joint tracking error, and joint velocity for each joint. , For the first One-hot encoding of the velocity symbols of each joint. For the first Joint velocity of each joint.

9. The method for online identification of load mass and center of mass in a dual-arm closed-chain handling scenario as described in any one of claims 1-8, characterized in that, Based on the joint feature vectors under robot load conditions, the joint friction torque is estimated using a pre-trained friction torque model. ; The training method for the friction torque model includes: Obtain joint feature vectors and measure joint torques in the robot under no-load conditions. ; To measure joint torque Inverse dynamic torque calculated with robot dynamics model The difference is the joint friction torque. ; to reduce joint friction torque Standardization yields joint feature vectors. Corresponding friction label This allows us to construct a training set. The friction torque model based on the multilayer perceptron is trained using the training set to obtain a trained friction torque model.

10. A load mass and center of mass online identification system for dual-arm closed-chain material handling scenarios, characterized in that, Includes a processor, the processor being configured to execute the online load mass and centroid identification method for a dual-arm closed-chain transport scenario as described in any one of claims 1-9.

Images