A robot actuator dynamics coupled force estimation and actuator force compensation method and system

Abstract

The application belongs to the field of machine force control, and particularly relates to a robot executor dynamics coupling force estimation and executor force compensation method, which comprises: a teacher-student network force estimation module, a teacher network can realize real-time estimation of three degrees of freedom force, and a student network is generated through distillation training to realize lightweight real-time estimation of single degree of freedom executor force; a robot coupling dynamics modeling and compensation method is used to establish a frequency domain transfer function of coupling force-executor output force-environment contact force, realize analysis and compensation control of the force, and implement network generation and force compensation steps. The application realizes high-precision coupling force estimation without sensors: the application proposes a teacher-student network force estimation module, uses a teacher network to learn three-dimensional force on a multi-degree of freedom trajectory, and obtains a lightweight student network through knowledge distillation to realize real-time high-precision estimation of single degree of freedom force, avoiding dependence on high-cost force sensors in industrial actual deployment.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of machine force control, and particularly relates to a method and system for estimating dynamic coupling forces of robot actuators and compensating for actuator forces. Background Technology

[0002] With the widespread application of industrial robots in high-speed precision machining, manipulation, and force control tasks, the coupling relationship between the robot's own dynamics and the actuator's dynamics (microsystem) generates complex forces during the process of the robot (macrosystem) driving the actuator's motion. These forces directly affect the actuator's force control accuracy, operational stability, and machining quality. In scenarios such as precision assembly, grinding and polishing, and force-sensitive operations, how to estimate the dynamic coupling force between the robot and the actuator in a low-cost, real-time, and accurate manner, and achieve effective force compensation, has become an important research topic in the field of robot force control. Existing technologies for robot actuator force control suffer from the following main problems:

[0003] Macro-micro coupling forces are difficult to measure directly: The macro-micro coupling forces between robots and actuators typically rely on high-precision force sensors or complex measuring devices. These sensors are not only expensive and complex to install, but also difficult to deploy on a large scale in industrial settings, limiting the system's versatility and cost-effectiveness.

[0004] The impact of model noise and uncertainty on compensation effectiveness: The dynamic model of the robot and actuators contains uncertainties, and sensor measurements are subject to noise interference. Relying solely on simple modeling or direct measurement cannot achieve high-precision estimation of macro-micro coupling forces (estimation noise can reach tens or hundreds of Newtons), thus reducing the force compensation effect.

[0005] Highly complex models are difficult to deploy on edge computing: To obtain high-precision coupling forces between a robot and its actuators, it is usually necessary to establish complex dynamic models or perform large-scale calculations. The computing power of actuator controllers or edge devices is limited, and running these models in real time results in high load and large response latency, which limits the application of force compensation methods in practical industrial scenarios.

[0006] The lack of compensation strategies for coupling forces: Existing force control methods mostly focus on controlling the end effector force or local degree-of-freedom forces, lacking a dedicated compensation mechanism for the macro-micro coupling forces exerted by the robot on the actuator. In multi-degree-of-freedom operations or highly dynamic tasks, uncompensated coupling forces can cause actuator forces to deviate from target values, reducing operational accuracy and system stability. Therefore, a dedicated coupling force compensation model is needed to counteract the coupling forces exerted by the robot's body dynamics on the actuator, achieving high-precision force control.

[0007] To address the aforementioned issues, this paper proposes a method for real-time estimation of the macro-micro coupling force between a robot and its actuators without the need for expensive sensors. This method, combined with a coupling force compensation model, enables high-precision force control. This approach is of significant practical importance and application value for improving the force control performance of industrially deployed robots, ensuring stable execution of high-speed dynamic tasks, and reducing system deployment costs. Summary of the Invention

[0008] This invention provides a method for estimating and compensating the dynamic coupling forces of robot actuators, overcoming the shortcomings of existing technologies. The core of this invention lies in proposing a method for estimating and compensating the dynamic coupling forces of robot actuators. Its essential technical features include: a teacher-student network force estimation module, where the teacher network enables real-time estimation of forces in three degrees of freedom, and a student network is generated through distillation training to achieve lightweight real-time estimation of single-degree-of-freedom actuator forces; and a dynamic modeling and compensation method considering robot coupling, used to establish a frequency domain transfer function of the coupling force, actuator output force, and environmental contact force, enabling force analysis and compensation control. This comprehensive technical feature makes this method innovative in the field of robot force control technology, possessing high precision, strong real-time performance, and edge deployment capability.

[0009] This invention is implemented as follows: a method for estimating the dynamic coupling force of a robot actuator and compensating for the actuator force, the method comprising:

[0010] S1: Real-time estimation of three-degree-of-freedom forces is achieved through a teacher network, and a student network is generated by distillation training to achieve lightweight real-time estimation of single-degree-of-freedom actuator forces;

[0011] S2: Establish the dynamic transfer function between the coupling force, the actuator output force, and the environmental contact force, and perform coupling force compensation based on the transfer function;

[0012] S3: Utilize the output of the trained network to predict the force and generate actuator compensation signals to achieve closed-loop force control.

[0013] Furthermore, the teacher network adopts a long short-term memory layer and a fully connected layer structure, with multi-degree-of-freedom trajectory data as input and three-dimensional force as output. The student network adopts a long short-term memory layer and a fully connected layer structure, with k-dimensional features after feature selection processing as input and one-dimensional force as output.

[0014] Furthermore, when modeling the dynamic transfer function, the stator and mover of the actuator are connected by an equivalent spring damping, and the mover is connected to the environment by an equivalent stiffness. The transfer function is used to describe the influence of stator mass, mover mass, connection stiffness, connection damping, and environmental stiffness on the coupling force of the actuator.

[0015] Another objective of this invention is to provide a method for standardizing multi-degree-of-freedom trajectory data of robot actuators, comprising:

[0016] Acquire actuator current, joint position, velocity, acceleration, and three-dimensional force sensor data;

[0017] The above data is subjected to global standardization, which uses mean and variance normalization to form input samples that can be directly trained by the neural network.

[0018] Furthermore, the standardization process includes: constructing the i-th trajectory into a data matrix with a time step of T and a feature dimension of 27, subtracting the mean from each feature dimension and dividing by the variance, while keeping the force sensor data and the dynamic model estimate as supervision signals.

[0019] Another objective of this invention is to provide a method for predicting the force of a robot actuator based on distillation training, comprising:

[0020] Three-dimensional force prediction results were generated using a teacher network.

[0021] Filter the input feature subset through correlation analysis;

[0022] The feature subset is input into the student network, and the weighted loss is calculated by combining the teacher network output and the sensor measured values ​​through distillation training, thereby optimizing the accuracy of the student network in predicting one-dimensional forces.

[0023] Furthermore, the teacher network loss function is composed of a weighted sum of the Euclidean distance between the sensor data and the prediction result, and the Euclidean distance between the dynamic model output and the prediction result, wherein the weighting factor is adaptively adjusted according to the piecewise Bayesian weights.

[0024] Furthermore, the calculation of the piecewise Bayesian weights includes: when the difference between the dynamic model estimate and the sensor measurement is greater than a threshold, an exponentially decaying weight is used; when the difference is less than a threshold, a student t-distribution probability density function is used to assign weights.

[0025] Furthermore, the student network loss function is composed of the teacher network prediction value, the student network prediction value, and the sensor measured value. The loss function includes distillation of training coefficients to balance the contributions of soft labels and hard labels.

[0026] Another objective of this invention is to provide a system for estimating and compensating for dynamic coupling forces of a robot actuator, comprising:

[0027] The data processing module is used to collect and standardize the multi-degree-of-freedom trajectory data of the actuator;

[0028] The teacher-student network module is used for force prediction and distillation training.

[0029] The dynamic modeling and compensation module is used to establish the transfer function and generate compensation force signals;

[0030] The control and execution module is used to achieve closed-loop force control of the actuator based on the compensation signal.

[0031] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0032] First, (1) Achieving high-precision coupling force estimation without sensors: This invention proposes a teacher-student network force estimation module, which uses the teacher network to perform three-dimensional force learning on multi-degree-of-freedom trajectories and obtains a lightweight student network through knowledge distillation, thereby achieving real-time high-precision estimation of single-degree-of-freedom forces and avoiding dependence on high-cost force sensors in actual industrial deployment.

[0033] (2) Significantly suppress the influence of model noise and measurement uncertainty: The momentum observer model is embedded in the teacher network as a physical information model embedding, and the loss weighting unit of the teacher network is constructed by combining the segmented Bayesian weighting mechanism to dynamically assign weights to abnormal data, which effectively reduces the interference of sensor noise and dynamic modeling error on prediction accuracy.

[0034] (3) Lightweight design that can be deployed in edge controllers: By screening the optimal feature subset through Pearson correlation analysis, only a small number of features most relevant to the target force are retained as input to the student network, which greatly reduces the number of network parameters and enables millisecond-level real-time inference in the actuator edge controller, improving the system deployment flexibility.

[0035] (4) Construct an equivalent dynamic model of robot-actuator stator-mover-environment and derive the frequency domain transfer function to realize the quantitative relationship analysis between coupling force-actuator output force-environment contact force, and provide a theoretical basis for force compensation control.

[0036] (5) Active compensation for coupling forces of macro system: Based on the compensation expression derived from the transfer function, the coupling forces estimated by the student network are injected into the control loop in real time to achieve active cancellation of disturbance forces of macro system, significantly improving the stability and accuracy of actuator force control, and is particularly suitable for high-speed and high-dynamic force control tasks.

[0037] Secondly, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0038] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0039] As a key unit of intelligent manufacturing, "robots + end effectors" have been widely applied in high-end equipment manufacturing, precision assembly, complex surface processing, intelligent services, and human-machine interaction. For example, in aerospace and automotive manufacturing, robot actuators need to achieve stable normal force control in processes such as surface drilling, riveting, welding, and polishing to ensure the machining accuracy and connection strength of structural components. In 3C electronics and semiconductor manufacturing, robot actuators undertake the assembly, bonding, and packaging of high-precision components, requiring extremely high force control response speed and micro-Newton level accuracy. In medical assistive devices and rehabilitation robots, actuators need to apply compliant and stable assistive forces to the patient's body to ensure the safety and comfort of the interaction. However, traditional methods heavily rely on expensive high-precision force sensors or complex analytical dynamic models, which increases deployment costs and makes it difficult to meet the real-time and robustness requirements of industrial settings. This invention overcomes the technical bottlenecks of sensor dependence and complex modeling by achieving real-time estimation and active compensation of macro-micro coupling forces at the actuator level. It significantly improves robot force control accuracy and dynamic stability without the need for external high-precision force sensors, thereby substantially reducing system hardware costs and long-term maintenance costs. Furthermore, the knowledge distillation and lightweight design proposed in this invention enable the student network to run on the edge controller at millisecond levels, ensuring the deployability and efficiency of complex force control strategies in resource-constrained environments. This advantage not only enhances the feasibility of robot applications in complex tasks such as high-speed, high-precision machining, assembly, polishing, and human-robot interaction, but also provides conditions for large-scale promotion. It can be implemented on a large scale in multiple industries such as industrial robots, aerospace, automotive manufacturing, precision electronics, and medical assistive devices. This will not only help improve the autonomous force control capabilities and industrial competitiveness of domestic high-end robot equipment, but also bring significant economic and social benefits by reducing costs and improving operational reliability and safety.

[0040] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0041] When robots perform complex precision machining, assembly, and surface treatment tasks, the accuracy of normal force control and dynamic response capability at the actuator end effector are crucial to machining quality and operational safety. However, there is currently no mature solution that can achieve an integrated force control method at the robot actuator level that combines "multi-degree-of-freedom trajectory data driving + dynamic coupling force estimation + active coupling force compensation." Existing technologies mostly focus on direct measurement by end effector force sensors or single-degree-of-freedom local force control, lacking a mechanism for unified modeling and real-time compensation of macro-micro coupling forces caused by robot body motion. This leads to problems such as delayed force control response, insufficient estimation accuracy, and untimely compensation in high-speed, high-precision tasks, severely limiting the system's dynamic force control performance and adaptability to complex operations.

[0042] This invention is the first to introduce a momentum observer into the training process of a deep teacher network, achieving high-precision modeling of the coupling forces in robot body dynamics. It also innovatively proposes a piecewise Bayesian residual weighting strategy, effectively suppressing abnormal prediction biases in the teacher network under high-noise trajectory data. Subsequently, knowledge distillation enables the migration of the teacher network to a lightweight student network, significantly reducing the computational load of online inference, thus enabling the actuator to possess embedded real-time force control inference capabilities. This invention overcomes the bottlenecks of computational complexity and difficulty in real-time deployment inherent in traditional analytical dynamic models. It is the first of its kind proposed and systematically implemented both domestically and internationally, filling a technological gap in the estimation and compensation control of hierarchical coupling forces in robot actuators for precision operations.

[0043] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully:

[0044] Low-cost measurement of macro-micro system coupling forces: By using a teacher network to learn three-dimensional coupling forces end-to-end based on trajectory data, the high cost and complex deployment of high-precision force sensors are avoided.

[0045] The problem of large estimation errors caused by model noise and uncertainty: By embedding a momentum observer model and a piecewise Bayesian weighting mechanism, the interference of dynamic model residuals and sensor noise on the prediction results is effectively suppressed.

[0046] The problem of complex dynamic models being difficult to run in real time on edge controllers: By generating lightweight student networks through feature selection and knowledge distillation, the computational complexity of the model is greatly reduced, enabling millisecond-level real-time inference in edge deployments.

[0047] The problem of lacking an active compensation mechanism for coupling forces: By constructing a macro-micro system dynamics transfer model and deriving the compensation expression, the predicted coupling forces are used in real time for control compensation, which significantly improves the accuracy of actuator output force and system dynamic stability.

[0048] (4) The technical solution of this invention overcomes technical bias: This invention effectively breaks the above-mentioned bias through core technological innovation: On the one hand, in response to the traditional perception that "high-precision force control decoupling must rely on high-precision force sensors", a force estimation and compensation method based on teacher-student networks and simplified parameter identification is proposed. Model calibration and force estimation can be completed by relying only on basic motion data collected by conventional sensors, thereby significantly reducing the dependence on expensive force sensors and breaking the technical stereotype that "high precision is inevitably accompanied by high cost". On the other hand, in response to the technical bias that "complex coupled dynamic models cannot run in real time", a lightweight dynamic coupling algorithm and knowledge distillation strategy are designed. Under the condition of not relying on ultra-high computing power hardware, the calculation dimension of multi-joint coupling parameters is significantly compressed, and millisecond-level real-time compensation is achieved, proving that complex coupling effects can be dynamically corrected through algorithm optimization. Thus, this invention not only breaks through the inherent cognitive limitations that have existed in the industry for a long time and expands the technical boundaries of robot dynamic control, but also promotes the transformation of high-precision force compensation technology from "high-end niche" to "universal and practical". Attached Figure Description

[0049] Figure 1 This is a flowchart of a method for estimating dynamic coupling forces and compensating actuator forces in a robot actuator according to an embodiment of the present invention;

[0050] Figure 2 This is a block diagram of a robot actuator dynamic coupling force estimation and compensation system provided in an embodiment of the present invention;

[0051] Figure 3 This is a schematic diagram of a robot actuator system modeling provided in an embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram of the control architecture of the compensation method provided in an embodiment of the present invention;

[0053] Figure 5 This is a schematic diagram illustrating the network generation and force compensation implementation provided in an embodiment of the present invention;

[0054] Figure 6 This is a robot motion trajectory diagram generated by an algorithm, as provided in an embodiment of the present invention;

[0055] Figure 7 This refers to the actuator used in the embodiments of the present invention;

[0056] Figure 8 This is a set of comparative simulation / experimental results of the coupling force in the X, Y, and Z directions changing over time, provided by an embodiment of the present invention.

[0057] Figure 9 This is a contact force state diagram after compensation and control provided in an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0059] like Figure 1 As shown, the method for estimating the dynamic coupling force of a robot actuator and compensating for the actuator force provided in this embodiment of the invention includes:

[0060] S1: Real-time estimation of three-degree-of-freedom forces is achieved through a teacher network, and a student network is generated by distillation training to achieve lightweight real-time estimation of single-degree-of-freedom actuator forces;

[0061] S2: Establish the dynamic transfer function between the coupling force, the actuator output force, and the environmental contact force, and perform coupling force compensation based on the transfer function;

[0062] S3: Utilize the output of the trained network to predict the force and generate actuator compensation signals to achieve closed-loop force control.

[0063] The teacher-student network force estimation module provided in this embodiment of the invention includes a data loading and normalization unit, a multi-step sequence construction unit, a teacher network unit, a feature selection unit, a student network unit, a teacher network loss weighting unit, and a student network loss weighting unit.

[0064] The data loading and normalization unit provided in this embodiment of the invention is used to load and collect multi-degree-of-freedom trajectory data of the robot actuator, including robot current. Joint position ,speed acceleration 3D force sensor data A total of 27 dimensions of data were collected, and the data underwent global standardization processing.

[0065] The data loading and normalization unit standardization processing method provided in this embodiment of the invention is as follows:

[0066] Let the first The data matrix of the trajectories is ,in For time steps, For the feature dimension (joint current) Joint position ,speed acceleration ,common ); normalized to , , ,in For network feature data, For force sensor data used to calculate the loss function, Force data for the dynamic model calculated by the momentum observer. The mean of the network feature data , where is the variance of the network feature data. The average value of force sensor data .

[0067] The multi-step sequence construction unit provided in this embodiment of the invention divides the normalized multi-degree-of-freedom trajectory data into steps with a fixed step size. Constructing the input sequence of the network ;

[0068] The teacher network unit includes a sequence input layer with an input dimension of 24, an LSTM layer, a fully connected layer, and an activation layer.

[0069] The feature selection unit selects the k features with the highest correlation to a specified single-dimensional force from the original 24-dimensional features through Pearson correlation analysis for lightweight training of the student network.

[0070] The student network unit includes a sequence input layer with an input dimension of k, an LSTM layer, a fully connected layer, and an activation layer.

[0071] The teacher network loss weighting unit provided in this embodiment of the invention is constructed by combining the Euclidean norm of GMO data weighted by sensor data, and its expression is: ,in For the number of trajectories, Representing sensor data Teacher network prediction results Euclidean distance, Represents GMO data Teacher network prediction results Euclidean distance, For the total loss function, The segmented Bayesian weighting is used to assign weights to the noise.

[0072] The piecewise Bayesian weighted weights provided in this embodiment of the invention are based on the Euclidean norm of the difference between the GMO force estimate and the sensor's measured value, and are applied through a threshold. Determine the segmented calculation weight When the value is greater than the threshold, weights are assigned through exponential decay (exp); when the value is less than the threshold, weights are assigned through the probability density function (t-pdf) of the student's t-distribution, the expression of which is:

[0073] , For exponential coefficients;

[0074] Furthermore, the student network loss weighted unit is used to distill the training process, making the model lightweight as a one-dimensional force prediction model, with the loss function... Defined as:

[0075] ,in The distillation training coefficient, For the student network prediction of unidimensional forces, For one-dimensional force teacher network prediction, This is data from a single-dimensional force sensor.

[0076] In S2 provided by this embodiment of the invention, considering the dynamic modeling and compensation method for robot coupling, the system is modeled as equivalent: the robot coupling force acts on the actuator stator, the actuator stator is connected to the actuator mover through a spring-damped actuator, and the actuator mover is equivalently connected to the environment by a spring and generates a force. There is a driving force between the mover and the stator;

[0077] Its transfer function can be expressed as: , ,in For the actuator stator mass, Actuator mover mass, For the connection stiffness between the stator and mover of the actuator, For the connection damping between the stator and mover of the actuator, For the interaction stiffness between the actuator mover and the environment, Applying coupling force to the actuator of the robot system, The driving force between the actuator mover and stator. For the displacement of the actuator stator and the robot end effector, Actuator mover displacement, This represents the contact force exerted by the actuator mover on the environment;

[0078] Define the intermediate polynomial , , We can obtain an explicit solution for the transfer function of the dynamic modeling and compensation method considering robot coupling: ;

[0079] In order to maintain the contact force at the target value The compensation transfer function considering robot coupling can be calculated as follows: The first item Used to achieve target contact force Feedforward force; second term Used to compensate for the coupling force generated by the robot on the actuator. Disturbance of contact force;

[0080] In actual deployment of a single-degree-of-freedom actuator system, the coupling force exerted by the robot system on the actuator is equal to the predicted force value of the student model. When deploying a multi-degree-of-freedom actuator system, the coupling force exerted by the robot system on the actuators is equal to the predicted force value of the teacher model. .

[0081] S3 provided in this embodiment of the invention specifically includes:

[0082] S31: Data Acquisition and Standardization: The data loading and normalization unit acquires multi-degree-of-freedom trajectory data of the robot actuator, including actuator current, joint position, velocity and acceleration, and force sensor data installed between the robot and the actuator, and performs global standardization processing on the data.

[0083] S32: Data Construction: The multi-step sequence construction unit constructs the input sequence from the normalized multi-degree-of-freedom trajectory data according to a fixed step size, so as to form the time series samples required for teacher network training;

[0084] S33: Teacher Network Training: The teacher network unit is trained based on the teacher network loss weighted unit to obtain the teacher network;

[0085] S34: Student Network Feature Selection: The feature selection unit selects the most important input dimensions based on the correlation between the teacher network's prediction of the target single-degree-of-freedom force and each input feature, in order to reduce the computational complexity of the student network.

[0086] S35: Student Network Training: The student network unit is trained based on the student network loss weighted unit, which only predicts the target single-degree-of-freedom force. The teacher network soft label and actual sensor measurement are combined and weighted through loss weighting and distillation training units to optimize the prediction accuracy of the student network.

[0087] S36: Compensation and Control: The trained student network uses the estimated single-degree-of-freedom force to perform coupled dynamic compensation calculations through the prediction and compensation output unit, generating actuator force compensation signals to achieve closed-loop force control, thereby completing the training and application process of the entire teacher-student network force estimation module.

[0088] The teacher network provided in this embodiment of the invention adopts a long short-term memory layer and a fully connected layer structure. The input is multi-degree-of-freedom trajectory data, and the output is a three-dimensional force. The student network adopts a long short-term memory layer and a fully connected layer structure. The input is k-dimensional features after feature selection processing, and the output is a one-dimensional force.

[0089] When modeling the dynamic transfer function provided in this embodiment of the invention, the stator and mover of the actuator are connected by an equivalent spring damping, and the mover is connected to the environment by an equivalent stiffness. The transfer function is used to describe the influence of stator mass, mover mass, connection stiffness, connection damping and environmental stiffness on the coupling force of the actuator.

[0090] This invention provides a method for standardizing multi-degree-of-freedom trajectory data of a robot actuator, characterized by comprising:

[0091] Acquire actuator current, joint position, velocity, acceleration, and three-dimensional force sensor data;

[0092] The above data is subjected to global standardization, which uses mean and variance normalization to form input samples that can be directly trained by the neural network.

[0093] The standardization process provided in this embodiment of the invention includes: constructing the i-th trajectory into a data matrix with a time step of T and a feature dimension of 27, subtracting the mean and dividing by the variance for each feature dimension, while keeping the force sensor data and the dynamic model estimate as supervision signals.

[0094] This invention provides a method for predicting the force of a robot actuator based on distillation training, characterized by comprising:

[0095] Three-dimensional force prediction results were generated using a teacher network.

[0096] Filter the input feature subset through correlation analysis;

[0097] The feature subset is input into the student network, and the weighted loss is calculated by combining the teacher network output and the sensor measured values ​​through distillation training, thereby optimizing the accuracy of the student network in predicting one-dimensional forces.

[0098] The teacher network loss function provided in this embodiment of the invention is composed of a weighted sum of the Euclidean distance between sensor data and prediction results, and the Euclidean distance between the dynamic model output and prediction results, wherein the weighting factor is adaptively adjusted according to the piecewise Bayesian weights.

[0099] The calculation of the piecewise Bayesian weights provided in this embodiment of the invention includes: when the difference between the dynamic model estimate and the sensor measurement is greater than a threshold, an exponentially decaying weight is used; when the difference is less than a threshold, a student t-distribution probability density function is used to assign weights.

[0100] The student network loss function provided in this embodiment of the invention is composed of teacher network predictions, student network predictions, and sensor measured values. The loss function includes distillation of training coefficients to balance the contributions of soft labels and hard labels.

[0101] like Figure 2 As shown in the figure, an embodiment of the present invention provides a system for estimating and compensating dynamic coupling forces of a robot actuator, comprising:

[0102] The data processing module is used to collect and standardize the multi-degree-of-freedom trajectory data of the actuator;

[0103] The teacher-student network module is used for force prediction and distillation training.

[0104] The dynamic modeling and compensation module is used to establish the transfer function and generate compensation force signals;

[0105] The control and execution module is used to achieve closed-loop force control of the actuator based on the compensation signal.

[0106] In current robot actuator control, a single force sensor or a dynamic model-based observer is commonly used to estimate the force. However, due to the significant dynamic coupling effect between the actuator and the robot body, a systematic deviation occurs between the sensor feedback signal and the actual interaction force. Simultaneously, sensors are affected by noise, temperature drift, and hysteresis, while model observers are prone to estimation errors due to model simplification. The combination of these factors results in force estimation lacking the real-time performance and accuracy required for industrial applications, severely limiting the robot's application in precision manipulation and force control scenarios.

[0107] To address the aforementioned issues, this method introduces a teacher-student network architecture. The teacher network learns the mapping relationship between force sensor data and dynamic model estimation data under multi-degree-of-freedom conditions, achieving high-precision three-degree-of-freedom force estimation. Furthermore, through distillation training, the predictive power of the teacher network is compressed into the student network, enabling lightweight real-time prediction at the single-degree-of-freedom actuator level. This approach overcomes the contradiction between excessive model complexity and unstable sensor data in traditional methods, ensuring millisecond-level inference speeds even in high-frequency control loops.

[0108] At the dynamic level, this method explicitly considers the energy coupling effect brought about by the spring-damped structure by equivalently modeling the relationship between the actuator stator and mover. The contact process between the actuator mover and the environment is equivalent to an elastic connection, and the coupling force is formally described as the frequency domain relationship between the input and output through a transfer function. With the help of this transfer function, external environmental disturbances, internal interactions of the actuator, and coupling forces of the robot body can be unified into a single computational framework, thus providing a theoretical basis for compensation control. This modeling approach transforms traditional compensation methods that rely on empirical parameter adjustments into a systematic design based on frequency domain analysis.

[0109] In the loss function design, the teacher network's learning process relies not only on sensor measurements but also on auxiliary estimates based on a dynamic observer. By constructing a weighted Euclidean distance and combining it with a piecewise Bayesian weighting mechanism, the weights can be adaptively adjusted under different noise levels and anomaly conditions. For example, when there is a significant difference between the sensor data and the model estimate, the system automatically reduces the contribution of that data to the loss to avoid the model getting overfitted to anomalous noise. This mechanism, combining statistics and control theory, significantly improves the robustness of the prediction.

[0110] During the training phase, the student network uses a feature selection unit to select the input dimension most relevant to the target single-degree-of-freedom force, reducing redundant information interference. Simultaneously, during the distillation process, it combines the soft labels from the teacher network with the hard labels from the sensors for joint training. The soft labels provide a smoother function approximation, while the hard labels ensure the consistency of physical measurements. The combination of these two allows the student network to maintain low computational complexity while achieving prediction accuracy close to that of the teacher network. This design adapts to the dual constraints of computational speed and resource consumption in industrial robot controllers.

[0111] In the actual actuator control process, the predicted force output in real time by the student network is fed into the compensation controller, and the compensation signal is obtained by inversely solving the transfer function. This signal contains both a feedforward component to maintain the target contact force and a feedback component to counteract the disturbance of the coupling force. In closed-loop operation, the actuator can stably output the desired force, and the environmental contact process exhibits higher compliance and stability. Therefore, this method not only solves the shortcomings of existing technologies in accurate force estimation and coupling compensation, but also significantly improves the operational reliability and industrial application value of robots in complex task scenarios.

[0112] This invention provides a method for estimating the dynamic coupling force of a robot actuator and compensating for the actuator force. The method is characterized by a teacher-student network force estimation module, a dynamic modeling and compensation method considering robot coupling, and network generation and force compensation implementation steps.

[0113] The teacher-student network force estimation module includes a data loading and normalization unit, a multi-step sequence construction unit, a teacher network unit, a feature selection unit, a student network unit, a teacher network loss weighting unit, and a student network loss weighting unit.

[0114] The data loading and normalization unit is used to load and collect multi-degree-of-freedom trajectory data from the robot actuator, including robot current. Joint position ,speed acceleration 3D force sensor data A total of 27 dimensions of data were collected, and the data underwent global standardization processing.

[0115] The data loading and normalization unit standardization processing method is as follows:

[0116] Let the first The data matrix of the trajectories is ,in For time steps, For the feature dimension (joint current) Joint position ,speed acceleration ,common ); normalized to , , ,in For network feature data, For force sensor data used to calculate the loss function, Force data for the dynamic model calculated by the momentum observer. The mean of the network feature data , where is the variance of the network feature data. The average value of force sensor data ;

[0117] Multi-step sequence construction unit, which divides the normalized multi-degree-of-freedom trajectory data into units with a fixed step size. Constructing the input sequence of the network ;

[0118] The teacher network unit includes a sequence input layer with an input dimension of 24, an LSTM layer, a fully connected layer, and an activation layer.

[0119] The feature selection unit selects the k features with the highest correlation to a specified single-dimensional force from the original 24-dimensional features through Pearson correlation analysis, which are then used for lightweight training of the student network.

[0120] The student network unit consists of a sequence input layer with an input dimension of k, an LSTM layer, a fully connected layer, and an activation layer.

[0121] The teacher network loss weighting unit is constructed by combining the Euclidean norm of GMO data weighted by sensor data, and its expression is: ,in For the number of trajectories, Representing sensor data Teacher network prediction results Euclidean distance, Represents GMO data Teacher network prediction results Euclidean distance, For the total loss function, Piecewise Bayesian weighting for noise;

[0122] Piecewise Bayesian weighted weighting compares the Euclidean norm of the GMO force estimate with the sensor's measured values ​​and uses a threshold. Determine the segmented calculation weight When the value is greater than the threshold, weights are assigned through exponential decay (exp); when the value is less than the threshold, weights are assigned through the probability density function (t-pdf) of the student's t-distribution, the expression of which is:

[0123] , For exponential coefficients;

[0124] The student network loss weighted unit is used for distillation training to lightweight the model into a one-dimensional force prediction model. The loss function... Defined as:

[0125] ,in The distillation training coefficient, For the student network prediction of unidimensional forces, For one-dimensional force teacher network prediction, This is data from a single-dimensional force sensor.

[0126] like Figure 3 As shown, considering the dynamic modeling and compensation method for robot coupling, the system is modeled as follows: the robot coupling force acts on the actuator stator, the actuator stator is connected to the actuator mover through a spring-damped actuator, and the actuator mover is equivalently connected to the environment by a spring and generates a force. There is a driving force between the mover and the stator;

[0127] like Figure 3 As shown, considering the dynamic modeling and compensation method for robot coupling, its transfer function can be expressed as: , ,in For the actuator stator mass, Actuator mover mass, For the connection stiffness between the stator and mover of the actuator, For the connection damping between the stator and mover of the actuator, For the interaction stiffness between the actuator mover and the environment, Applying coupling force to the actuator of the robot system, The driving force between the actuator mover and stator. For the displacement of the actuator stator and the robot end effector, Actuator mover displacement, This represents the contact force exerted by the actuator mover on the environment.

[0128] Define the intermediate polynomial , , We can obtain an explicit solution for the transfer function of the dynamic modeling and compensation method considering robot coupling:

[0129] like Figure 4 As shown, in order to maintain the contact force at the target value The compensation transfer function considering robot coupling can be calculated as follows: The first item Used to achieve target contact force Feedforward force; second term Used to compensate for the coupling force generated by the robot on the actuator. Disturbance of contact force;

[0130] In actual deployment of a single-degree-of-freedom actuator system, the coupling force exerted by the robot system on the actuator is equal to the predicted force value of the student model. When deploying a multi-degree-of-freedom actuator system, the coupling force exerted by the robot system on the actuators is equal to the predicted force value of the teacher model. ;

[0131] like Figure 5 As shown, the steps for generating the inventive network and implementing force compensation are as follows:

[0132] Data acquisition and standardization: The data loading and normalization unit acquires multi-degree-of-freedom trajectory data of the robot actuator, including actuator current, joint position, velocity and acceleration, and force sensor data installed between the robot and the actuator, and performs global standardization processing on the data;

[0133] Data Construction: The multi-step sequence construction unit constructs the input sequence from the normalized multi-degree-of-freedom trajectory data according to a fixed step size, so as to form the time series samples required for teacher network training;

[0134] Teacher network training: Teacher network units are trained based on teacher network loss weighted units to obtain the teacher network;

[0135] Student network feature selection: The feature selection unit selects the most important input dimension based on the correlation between the teacher network's prediction of the target single-degree-of-freedom force and each input feature, so as to reduce the computational complexity of the student network.

[0136] Student network training: The student network unit is trained based on the student network loss weighted unit, which only predicts the target single degree of freedom force. The teacher network soft label and actual sensor measurement are combined for weighted training through loss weighting and distillation training units to optimize the prediction accuracy of the student network.

[0137] Compensation and Control: The trained student network uses the estimated single-degree-of-freedom force to perform coupled dynamic compensation calculations through the prediction and compensation output unit, generating actuator force compensation signals to achieve closed-loop force control, thereby completing the training and application process of the entire teacher-student network force estimation module.

[0138] I. Specific application areas or related products of this invention.

[0139] This invention is applicable to various robot applications with stringent requirements for force accuracy, dynamic response, and system stability. Related applications and products include, but are not limited to:

[0140] In high-end manufacturing sectors such as aerospace, automobile manufacturing, shipbuilding, and energy equipment manufacturing, processes such as drilling, riveting, welding, polishing, and grinding can significantly improve force control stability and processing quality during the processing of complex curved surfaces and high-hardness materials.

[0141] In the precision electronics and semiconductor industry, tasks such as chip packaging, component assembly, and optical device assembly and adjustment can significantly improve assembly accuracy and yield by precisely controlling micro-Newton level forces.

[0142] Intelligent manufacturing and service robots: including human-robot collaborative robots, assembly robots and service robots. With the help of the real-time force compensation capability of this invention, safer and more compliant force control can be achieved in assembly, handling and interaction processes.

[0143] In the medical and rehabilitation fields, such as rehabilitation training robots, assistive robots, and surgical robots, high-precision force compensation control at the actuator level can improve the safety and comfort of interaction with the human body.

[0144] Other robotic systems with force control requirements, such as special-purpose robots, underwater robots, and space robots, can still ensure real-time force control and operational stability in complex environments.

[0145] II. Evidence related to the technical effects obtained by the embodiments of the present invention.

[0146] The invention has been applied to the "robot + end effector grinding and polishing" experimental platform for verification. The robot used in the experiment is the Huashu HSRCo610 six-degree-of-freedom industrial robot, and the end effector is a self-developed force control actuator, which can achieve rigid connection with the end flange of the robot and provide independent force control output capability during the contact process.

[0147] Figure 6 The robot motion trajectory generated by the algorithm is used for the subsequent acquisition of multi-degree-of-freedom trajectory data of the robot actuator.

[0148] Figure 7 The actuator used can be fixed to the end effector of the robot.

[0149] Figure 8 The accuracy of the network prediction power proposed in this invention.

[0150] Figure 9 The contact force state after compensation and control.

[0151] First, a robot motion trajectory was generated based on an algorithm. This trajectory was used to collect operational data such as current, position, velocity, and acceleration of the robot under multi-degree-of-freedom conditions. A self-developed force-controlled actuator was fixed to the robot's end effector, and data acquisition and verification experiments were conducted in a typical grinding and polishing scenario. After data acquisition, the proposed teacher-student network was trained, yielding the force prediction results shown by the proposed network. The results show that the proposed network has high accuracy in force estimation; the error between the predicted and actual measured values ​​is significantly smaller than that of traditional model estimation results, and the results are almost consistent with the actual sensor values, demonstrating the effectiveness and robustness of the method. Furthermore, after introducing the proposed coupled force estimation and compensation method into the compensation control stage, the contact force state between the actuator and the environment was significantly improved, the contact force fluctuation amplitude was significantly reduced, and the output force could be stably maintained near the target value, thus verifying the application effect of this invention in contact operations.

[0152] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0153] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for estimating dynamic coupling forces and compensating actuator forces in robot actuators, characterized in that, include: S1: Real-time estimation of three-degree-of-freedom forces is achieved through a teacher network, and a student network is generated by distillation training to achieve lightweight real-time estimation of single-degree-of-freedom actuator forces; S2: Establish the dynamic transfer function between the coupling force, the actuator output force, and the environmental contact force, and perform coupling force compensation based on the transfer function; S3: Utilize the output of the trained network to predict the force and generate actuator compensation signals to achieve closed-loop force control.

2. The method according to claim 1, characterized in that, The teacher network adopts a long short-term memory layer and a fully connected layer structure. The input is multi-degree-of-freedom trajectory data, and the output is three-dimensional force. The student network adopts a long short-term memory layer and a fully connected layer structure. The input is k-dimensional features after feature selection processing, and the output is one-dimensional force.

3. The method according to claim 1, characterized in that, When modeling the dynamic transfer function, the stator and mover of the actuator are connected by an equivalent spring damping, and the mover is connected to the environment by an equivalent stiffness. The transfer function is used to describe the influence of stator mass, mover mass, connection stiffness, connection damping, and environmental stiffness on the coupling force of the actuator.

4. A method for standardizing multi-degree-of-freedom trajectory data of a robot actuator, implementing the method for estimating the dynamic coupling force and compensating the actuator force as described in any one of claims 1-3, characterized in that, include: Acquire actuator current, joint position, velocity, acceleration, and three-dimensional force sensor data; The above data is subjected to global standardization, which uses mean and variance normalization to form input samples that can be directly trained by the neural network.

5. The method according to claim 4, characterized in that, The standardization process includes: constructing the i-th trajectory into a data matrix with a time step of T and a feature dimension of 27, subtracting the mean from each feature dimension and dividing by the variance, while keeping the force sensor data and the dynamic model estimate as supervision signals.

6. A method for predicting robot actuator forces based on distillation training, implementing the robot actuator dynamic coupling force estimation and actuator force compensation method as described in any one of claims 1-3, characterized in that, include: Three-dimensional force prediction results were generated using a teacher network. Filter the input feature subset through correlation analysis; The feature subset is input into the student network, and the weighted loss is calculated by combining the teacher network output and the sensor measured values ​​through distillation training, thereby optimizing the accuracy of the student network in predicting one-dimensional forces.

7. The method according to claim 6, characterized in that, The teacher network loss function is composed of a weighted sum of the Euclidean distance between the sensor data and the prediction results, and the Euclidean distance between the dynamic model output and the prediction results, wherein the weighting factor is adaptively adjusted according to the piecewise Bayesian weights.

8. The method according to claim 7, characterized in that, The calculation of the piecewise Bayesian weights includes: when the difference between the dynamic model estimate and the sensor measurement is greater than a threshold, an exponentially decaying weight is used; when the difference is less than a threshold, the probability density function of the Student's t-distribution is used to assign weights.

9. The method according to claim 6, characterized in that, The student network loss function is composed of the teacher network prediction value, the student network prediction value, and the sensor measured value. The loss function includes distillation of training coefficients to balance the contributions of soft labels and hard labels.

10. A system for estimating and compensating dynamic coupling forces of a robot actuator, implementing the method for estimating and compensating dynamic coupling forces of a robot actuator as described in any one of claims 1-3, characterized in that, include: The data processing module is used to collect and standardize the multi-degree-of-freedom trajectory data of the actuator; The teacher-student network module is used for force prediction and distillation training. The dynamic modeling and compensation module is used to establish the transfer function and generate compensation force signals; The control and execution module is used to achieve closed-loop force control of the actuator based on the compensation signal.

Images