Robot force sensor external force estimation and admittance control method, system, and medium

Abstract

The application provides a robot force sensor external force estimation and mobility control method, system and medium, the method comprising: driving the robot to move along a Fourier series excitation trajectory and collecting joint state and torque data, and realizing robust identification of rigid body dynamics base parameters based on iterative weighted least squares; constructing a double-flow physical information neural network to fit the dynamics residual, training under a physical constraint loss function and solidifying the network weight; feeding forward the double-flow physical information neural network with solidified weight into a generalized momentum observer to realize joint external torque estimation; obtaining end Cartesian contact force through a damped least squares pseudo-inverse mapping; outputting a smooth contact mixing factor from a hysteresis anti-shake contact state machine to drive a three-stage variable damping mobility controller to realize adaptive compliant contact control. The application can effectively solve the technical problems of poor sensorless force estimation accuracy and difficult control and stabilization in the prior art.

Description

Technical Field

[0001] This application relates to the field of defect detection technology, specifically to a method, system, and medium for estimating the external force and controlling the admittance of a robot's powerless sensor. Background Technology

[0002] In contact-based machining scenarios such as grinding, polishing, deburring, and contour-following assembly, the contact force between the robot's end effector and the workpiece directly determines the machining quality and control stability. Existing force control solutions mostly rely on six-dimensional force / torque sensors to collect contact forces. However, these sensors occupy valuable space in the end effector tool, are difficult to deploy under complex tooling, and are costly and require periodic calibration. Furthermore, the cable layout is cumbersome, and the sensors have poor impact resistance and are prone to failure under high-speed conditions, which seriously restricts the large-scale engineering promotion of force-controlled robots.

[0003] To circumvent force sensors, the industry has proposed external force observation schemes based on motor current or joint torque. The core of these schemes is to rely on the robot's dynamics model, eliminating rigid body dynamics terms such as inertia, gravity, Coriolis force, and friction from the driving torque, and equating the residual torque with the external torque. Among these, the generalized momentum observer has become the mainstream sensorless estimation framework due to its simple structure and direct differentiation without acceleration. However, this scheme has significant drawbacks: the Stribeck friction characteristics of the harmonic reducer in the low-speed region cannot be accurately fitted by traditional linear models; coupled with unmodeled dynamics such as reducer hysteresis, assembly deviations, and joint flexibility, the external force estimation suffers from steady-state bias; near zero velocity crossing, the uncertainty of the friction direction, combined with measurement noise, causes fluctuations in the estimated value, interfering with constant force control.

[0004] Therefore, there is an urgent need for a sensorless closed-loop control scheme that balances physical interpretability and modeling accuracy, and achieves organic integration of dynamic estimation and contact control. Summary of the Invention

[0005] In view of this, this application provides a method, system and medium for sensorless force estimation and admittance control of robots, in order to solve the technical problems of poor accuracy and difficulty in stabilizing sensorless force estimation in the prior art.

[0006] In a first aspect, this application provides a method for estimating the external force and admittance control of a robot's powerless sensor, including: The robot is driven to move along a Fourier series excitation trajectory and joint state and torque data are collected. Robust identification of rigid body dynamics basis parameters is achieved based on iterative weighted least squares. A two-stream physical information neural network is constructed to fit the dynamic residuals, and the network weights are fixed under the physical constraint loss function. By feeding forward a dual-stream physical information neural network with fixed weights into a generalized momentum observer, joint external torque estimation can be achieved. The Cartesian contact force at the end is obtained through damped least-squares pseudo-inverse mapping. The smooth contact mixing factor is output by the hysteresis anti-jitter contact state machine, which drives the three-stage variable damping admittance controller to achieve adaptive compliant contact control.

[0007] Secondly, this application provides a robot force sensor external force estimation and admittance control system, comprising: The module includes a data acquisition module, a robust identification module, a dual-stream physical information neural network modeling module, an external force estimation module, a force mapping and contact identification module, a variable damping admittance control module, and an execution module. The system is used to perform the robot force estimation and admittance control method of the first aspect, which can achieve high-precision force estimation and compliant contact control without the need for a six-dimensional force / torque sensor.

[0008] Thirdly, this application provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, realize the robot force estimation and admittance control method of the first aspect.

[0009] The robot force estimation and admittance control scheme for the powerless sensor provided in this application has at least the following beneficial effects: By providing a framework around the mechanism and using hybrid modeling to compensate for data biases, the four core modules of robust rigid body dynamics parameter identification, offline-trained dual-stream physical information neural network residual compensation, enhanced generalized momentum observer (GMO), and three-stage variable damping admittance control are organically connected. This enables high-precision external force estimation and low-impact, high-steady-state contact force control without relying on any additional six-dimensional force / torque sensors. Its core technical logic is as follows: The rigid body dynamics mechanism model is based on the Newton-Euler equations, providing physical structural constraints for external force estimation; the dual-flow PINN specifically compensates for the unmodeled position-related dynamics caused by Stribeck friction nonlinearity, assembly deviation, and joint flexibility in the low-speed region of the harmonic reducer from the data side, and ensures the physical rationality of the network output across the entire speed range through training with a joint loss function that includes multiple physical constraints such as odd symmetry and dissipation; the hybrid dynamic torque compensation model formed by the superposition of the two enables the generalized momentum observer to maintain high-precision joint external torque estimation with low bias even under conditions with prominent model uncertainties, such as the low-speed region of the harmonic reducer, which can significantly reduce the steady-state external force estimation error compared to the observer scheme of the pure rigid body model; the high-precision external force estimation further supports the low-impact contact establishment and steady-state constant force high-precision tracking of the variable damping admittance controller. Attached Figure Description

[0010] Figure 1 A schematic diagram of the control flow of a robot force estimation and admittance control system based on a robot's powerless sensor is shown. Figure 2 This diagram illustrates the hardware connection of a robot's powerless sensor external force estimation and admittance control system. Figure 3 A flowchart illustrating a method for estimating external force and admittance control using a robot's powerless sensor is shown. Figure 4 A schematic diagram of a two-stream PINN residual modeling network topology is shown. Figure 5 A schematic diagram of an enhanced generalized momentum observer is shown. Figure 6 A schematic diagram of a three-stage variable damping adaptive admittance controller is shown. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] One embodiment of this application provides a robot force sensor external force estimation and admittance control system, including: The system comprises a data acquisition module, a robust identification module, a dual-stream physical information neural network modeling module, an external force estimation module, a force mapping and contact recognition module, a variable damping admittance control module, and an execution module; used in the following method for estimating and controlling the external force and admittance of a robot without a force sensor. Related details are provided in the following method embodiments, as they belong to the same concept. The specific implementation process is detailed in the following method implementation methods and will not be repeated here.

[0013] It should be noted that the robot force sensor external force estimation and admittance control system provided in this application is only illustrated by the division of functional modules / functional units when performing robot force sensor external force estimation and admittance control. In practical applications, the functions can be assigned to different functional modules / functional units as needed, that is, the internal structure of the robot force sensor external force estimation and admittance control system can be divided into different functional modules / functional units to complete all or part of the functions.

[0014] For example, such as Figure 1 As shown, the robot force estimation and admittance control system based on the robot's powerless sensor consists of two main parts: an offline identification and training subsystem and an online control subsystem. The two subsystems utilize fixed rigid body parameters. Connect to the PINN network weights. Figure 1 It demonstrates a complete technical route from data acquisition, rigid body parameter identification, dual-stream PINN offline training to online external force estimation, contact identification and variable damping admittance control.

[0015] The offline subsystem completes the following steps before the robot is put into actual operation: First, it collects excitation trajectory data and performs robust parameter identification to obtain rigid body dynamic basis parameters; then, it calculates the dynamic residual dataset based on the identification results and trains the dual-stream PINN residual compensation network, and solidifies the trained network weights for use by the online module.

[0016] The online subsystem operates in real time during robot operation. Its control loop includes the following sequentially executed calculation steps: acquiring the current joint state and joint torque, calling the fixed weight PINN for forward inference and injecting an enhanced generalized momentum observer to estimate the joint external torque, then mapping it to the end effector Cartesian space to obtain the contact force, outputting the contact mixing factor through the contact state machine, and finally completing the variable damping admittance calculation based on this, and outputting the corrected pose command to the servo controller.

[0017] like Figure 2 As shown, the system hardware architecture of this application includes: a real-time control computer, a joint servo driver, an industrial robot body (a six-joint serial robotic arm), and an encoder acquisition module (for obtaining joint positions). With speed ) and current sampling module (estimate joint driving torque) The real-time control computer runs modules such as PINN inference, generalized momentum observer, force mapping and admittance controller, and communicates with the servo driver through a real-time bus. The entire loop does not require the installation of a six-dimensional force / torque sensor.

[0018] The following section will elaborate on the specific implementation methods of the robot's powerless sensor external force estimation and admittance control method.

[0019] See Figure 3 As shown, this application provides a method for estimating the external force and admittance control of a robot's powerless sensor, which includes the following steps.

[0020] S301. Drive the robot to move along the Fourier series excitation trajectory and collect joint state and torque data. Robust identification of rigid body dynamics basis parameters is achieved based on iterative weighted least squares. S302. Construct a dual-stream physical information neural network to fit the dynamic residuals, complete the training under the physical constraint loss function, and solidify the network weights. S303. Feedforward the dual-stream physical information neural network with fixed weights into the generalized momentum observer to achieve joint external torque estimation. S304, the end Cartesian contact force is obtained through damped least squares pseudo-inverse mapping; S305, the hysteresis anti-shake contact state machine outputs a smooth contact mixing factor, which drives a three-stage variable damping admittance controller to achieve adaptive compliant contact control.

[0021] Specifically, to address the technical pain points of insufficient external force estimation accuracy, poor contact control compliance, and susceptibility to jitter or impact in robot scenarios without force sensors, an integrated solution is proposed that combines rigid body dynamics identification, physical information neural network, and variable damping admittance control. This solution can achieve high-precision external force estimation and adaptive compliant contact control without the need for additional force sensors.

[0022] First, the robot is driven to move along a preset Fourier series excitation trajectory. This trajectory has rich frequency components, which can fully stimulate the dynamic characteristics of each joint of the robot, ensuring the comprehensiveness and effectiveness of the collected data. During the movement, the position, velocity, acceleration, and other state data of each joint of the robot, as well as the joint driving torque data, are collected in real time, providing basic data support for the identification of dynamic parameters.

[0023] Based on the collected dataset, an iterative weighted least squares algorithm is used to robustly identify the rigid body dynamics basis parameters. This algorithm iteratively updates the weight coefficients to suppress outliers in the data, reducing the impact of measurement noise, trajectory disturbances, and other factors on the identification results. Ultimately, it yields high-precision robot rigid body dynamics basis parameters, providing a reliable dynamic model foundation for subsequent external force estimation.

[0024] Because non-rigid body effects such as friction, flexible deformation, and joint clearances exist during actual robot motion, rigid body dynamics models alone cannot fully describe their dynamic characteristics, resulting in non-negligible dynamic residuals that affect the accuracy of external force estimation. Therefore, a two-stream physical information neural network is constructed to fit these dynamic residuals.

[0025] This neural network employs a two-stream structure, receiving robot joint state data and dynamic basis parameter identification results as inputs. It is trained by fusing a physical constraint loss function (constructed in conjunction with fundamental laws of robot dynamics), ensuring the network output accurately matches the actual dynamic residuals while maintaining physical consistency and avoiding predictions that contradict dynamic principles. After training, the network weights are solidified, forming a directly usable residual fitting model that supplements and corrects external torque estimation.

[0026] A two-stream physical information neural network with fixed weights is used as a feedforward module and injected into the generalized momentum observer. The generalized momentum observer calculates the theoretical generalized momentum of the robot joints by combining real-time acquired joint state data with the identified rigid body dynamics basis parameters; at the same time, it uses the dynamic residuals output by the two-stream physical information neural network to correct the theoretical generalized momentum and improve the observation accuracy.

[0027] By using closed-loop regulation of a generalized momentum observer and feedforward compensation of a neural network, accurate estimation of external torques at each joint of the robot is achieved. This solves the core problem of the difficulty in directly measuring external torques in scenarios without sensors, and requires no additional hardware investment, thus reducing system costs.

[0028] There is a clear kinematic mapping relationship between the joint external torques and the end effector Cartesian contact force. Based on the robot's kinematic model, a damped least squares pseudo-inverse algorithm is used to map and transform the estimated joint external torques to obtain the robot's end effector contact force in the Cartesian coordinate system.

[0029] The damped least squares pseudo-inverse algorithm can effectively avoid the mapping failure problem caused by kinematic singularities. At the same time, by introducing a damping term, it can improve the stability and robustness of the mapping process, ensure the smoothness and accuracy of the end contact force calculation results, and provide direct force feedback basis for subsequent contact control.

[0030] A hysteresis-resistant contact state machine is designed to monitor the contact state (no contact, just making contact, stable contact, disengagement) between the robot's end effector and the environment in real time. Based on changes in the contact state, a smooth contact blending factor is output. This blending factor effectively suppresses jitter during the contact process, avoids control oscillations caused by sudden changes in contact state, and improves the stability of the contact process.

[0031] Driven by a contact mixing factor, a three-stage variable damping admittance controller is employed. This controller adaptively adjusts damping parameters based on different contact states, dividing the process into three stages: a non-contact stage (high damping to ensure motion stability), a contact transition stage (gradual damping to avoid impact), and a stable contact stage (optimal damping to achieve compliant control). Ultimately, this achieves adaptive and compliant contact control between the robot and its environment, meeting the control requirements of precision operations and human-robot collaboration scenarios.

[0032] This application effectively solves the technical problems of low external force estimation accuracy, poor contact control compliance, and easy jitter in powerless sensor robots by integrating multiple technologies. Compared with the prior art, it has the following significant advantages: Without the need for additional force sensors, high-precision external force estimation and contact control can be achieved solely through existing joint state and torque acquisition modules, combined with algorithm and model optimization, significantly reducing the hardware cost and deployment complexity of the robot system. Simultaneously, it avoids the space limitations, cumbersome wiring, and susceptibility to damage associated with force sensor installation, enhancing the method's applicability in complex scenarios such as confined spaces and harsh environments.

[0033] An iterative weighted least squares algorithm is employed to robustly identify the rigid body dynamics basis parameters, effectively suppressing the influence of data noise and outliers and ensuring the accuracy of the dynamic model. Combining this with a two-stream physical information neural network to fit the dynamic residuals compensates for the shortcomings of the rigid body dynamics model, achieving a precise match between the theoretical model and the actual system. The fusion of these two methods with a generalized momentum observer further improves the accuracy and robustness of joint external torque estimation, significantly reducing the calculation error of the end-effector Cartesian contact force and meeting the force sensing requirements of precision operation.

[0034] The hysteresis-resistant contact state machine can accurately identify changes in contact state and output a smooth contact mixing factor, effectively suppressing jitter and impact during the contact process and avoiding control oscillations. The three-stage variable damping admittance controller adaptively adjusts the damping parameters according to the contact state, realizing a smooth transition from non-contact to stable contact. This ensures the stability of the robot's motion process and achieves compliant contact with the environment, reducing damage to the environment or the object being operated. It is suitable for scenarios with high requirements for contact control, such as human-robot collaboration, precision assembly, and flexible grinding.

[0035] Each step of the algorithm exhibits good stability and convergence: the iterative weighted least squares algorithm converges quickly and provides reliable identification results; the dual-stream physical information neural network, combined with physical constraint training, avoids model overfitting and demonstrates strong generalization ability; the damped least squares pseudo-inverse algorithm solves the kinematic singularity problem and ensures stable mapping; the coordinated operation of the hysteresis-stabilized contact state machine and the variable damping admittance controller further enhances the stability of the entire control system. Furthermore, all algorithms can be implemented using conventional robot control hardware, requiring no special hardware support, resulting in high engineering feasibility and facilitating large-scale applications.

[0036] This method breaks through the technical bottleneck of force perception and compliant control in robots without external force sensors, enabling robots to achieve high-precision external force estimation and compliant contact control without external force sensors. This greatly expands the application scope of robots and can be widely used in many fields such as industrial assembly, human-robot collaboration, medical rehabilitation, and precision machining. It has high engineering application value and promotion prospects.

[0037] In this embodiment, preferably, the Fourier series excitation trajectory is generated by maximizing the excitation index of the dynamic regression matrix. The trajectory covers the target workspace and satisfies the joint motion constraints. The beginning and end of the trajectory are spliced ​​together with static segments to ensure that the start and stop speeds and accelerations are both zero.

[0038] In this application, the Fourier series excitation trajectory is as follows: in, , , These are the discrete Fourier coefficients of each joint. For the fundamental frequency, It represents the harmonic order.

[0039] The trajectory coefficients are optimized by maximizing the inverse of the condition number of the dynamic regression matrix (or maximizing the excitation index) to ensure sufficient identification information while constraining the trajectory within the range of motion of the robot joints.

[0040] By splicing a static trajectory at the beginning and end of the excitation trajectory, the robot can smoothly enter and exit the excitation motion from a stationary state, avoiding transient transitions that could affect the quality of the identification data.

[0041] Collect the location of each joint ,speed acceleration and joint driving torque The data is used to construct a dynamic regression matrix model as follows: in The measured torque vectors for each joint. For the dynamic regression matrix, The vector of rigid body dynamics basis parameters to be identified (including the minimum linearly independent combination of inertia parameters, center of mass parameters, and friction parameters).

[0042] This application optimizes the generation of Fourier series excitation trajectories by maximizing the excitation index of the dynamic regression matrix. This ensures that the trajectory fully covers the robot's target workspace, fully stimulates the dynamic characteristics of each joint, and strictly meets motion constraints such as joint position, velocity, and acceleration, preventing joint movements from exceeding safe limits or causing mechanical damage. The design of stitching the beginning and end of the trajectory with static segments allows the robot's velocity and acceleration to smoothly transition to zero during start-up and stop, effectively eliminating shocks and vibrations during the start-up and stop process. This ensures the stability of the joint state and torque data acquisition process, thereby improving the accuracy of subsequent dynamic parameter identification. This lays a high-quality data foundation for the entire external force estimation and control process, while also extending the service life of the robot's joint mechanical structure and reducing equipment maintenance costs.

[0043] In this embodiment, preferably, the iterative weighted least squares identification uses the Tukey double-weight function to remove abnormal samples and superimposes a velocity-related suppression factor to reduce the weight of samples in the low-speed zero-crossing interval and restore the full weight of samples at stationary points.

[0044] Specifically, traditional least squares (OLS) is sensitive to measurement noise and low-speed friction anomalies, easily leading to biased identification results towards outliers. In this application, iterative weighted least squares (IRLS) is employed, which automatically reduces the contribution of outliers through a robust weighting function, thereby obtaining more reliable parameter estimates. The iteratively updated parameter estimation formula is as follows: In each iteration, the residuals of each sample under the current parameters are calculated first, and then the robust weights are calculated from the residuals for the next round of weighted solution.

[0045] The weighting function uses the Tukey double-weight function to remove outliers. The Tukey double-weight function is as follows: in, For the first The weights of each sample, For the first Standardized residuals of each sample For the first The joint torque residuals of each sample; The robust scaling parameters are estimated using the median absolute deviation. This is the adjustment constant; The threshold for outlier detection is set when the residual amplitude exceeds [a certain threshold]. Abnormal samples have their weights reset to zero.

[0046] Considering the significant uncertainty inherent in samples within the low-velocity zero-crossing region, this application further incorporates velocity-related suppression factors onto the robust weights of each sample: in, For the first The joint angular velocity of each sample The velocity scale parameter is used to reduce the weight of high-uncertainty samples in the low-speed zero-crossing interval, and to restore the weight contribution of samples when both joint velocity and acceleration are below the static determination threshold.

[0047] Inhibition factors are used in combination with static judgment rules: when joint velocity With acceleration At the same time, it is below the threshold for determining stillness. , At that time, skip the velocity suppression factor and directly set the sample weights to 0. .

[0048] To fully preserve the contribution of stationary point samples to the identification of gravity parameters, and to avoid the loss of gravity identification information due to the velocity suppression factor approaching zero at zero velocity, the above combined strategy effectively reduces the weight of outlier data, and the rigid body dynamics basis parameter vector is obtained after identification convergence. .

[0049] This application employs an iterative weighted least squares identification method combining Tukey's dual-weight function and a velocity-related suppression factor. This method accurately identifies and removes outliers from the data, effectively avoiding interference from outliers in the identification results of dynamic basis parameters and significantly improving the robustness of the identification algorithm. By reasonably reducing the weights of samples in the low-speed zero-crossing interval, the identification error caused by uncertainties in the data (such as frictional nonlinearity and measurement noise superposition) in this interval can be suppressed. At the same time, the full weights are restored for samples at stationary points, ensuring the identification accuracy of dynamic parameters in static states. This successfully solves the industry pain point of insufficient accuracy of traditional identification methods under low-speed and static conditions. Overall, this optimized scheme significantly improves the identification accuracy and reliability of rigid body dynamic basis parameters, providing more accurate dynamic model support for subsequent joint external torque estimation, and indirectly improving the overall performance of the entire control system.

[0050] In this embodiment, preferably, the Dual-Stream PINN includes a main friction network and a compensation network, wherein the main friction network fits the Stribeck friction with the joint angular velocity as input; in, For the first The main friction network of each joint outputs torque. Main friction neural network, For the first Angular velocity of each joint; The compensation network fits the unmodeled dynamics using joint position, velocity, and acceleration as inputs; in, For the first The joint compensation network outputs torque. Compensation neural network, For the first Angular position and angular acceleration of each joint; The dynamic residual compensation torque is obtained by superimposing the two outputs; in, For the first The total dynamic residual compensation torque of each joint.

[0051] Specifically, rigid body models generate significant systematic residuals in the low-speed region of harmonic reducers, and observers relying solely on rigid body models exhibit poor estimation accuracy under these conditions. Therefore, this application constructs a dual-stream physical information neural network for structured offline learning of the dynamic residuals.

[0052] The dynamic residual is defined by subtracting the predicted torque from the measured torque in the training data: The residuals contain a mixture of velocity-dependent Stribeck friction contributions and position-dependent unmodeled dynamic contributions. The two-stream architecture approach involves fitting these two types of residual sources to different networks and then summing the results for the output. Main friction network Each joint is designed independently as a single-input fully connected network, using only the joint angular velocity. As input, it specifically fits velocity-related frictional characteristics (including the Stribeck effect). For example... Figure 4 As shown, the main friction network (single-input stream at the top) and the compensation network (three-input stream at the bottom) are connected in parallel. The residual compensation torque is obtained by superimposing the two outputs through an adder node. The right side shows the five components of the joint loss function: data consistency, Stribeck prior consistency, odd symmetry constraint, dissipative constraint, and compensation zeroing penalty.

[0053] Compensation network Using joint position, velocity, and acceleration as inputs, the network fits position-dependent unmodeled dynamics, such as gearbox backlash and disturbances caused by assembly deviations. During training, the network is subject to a nullification penalty to ensure that its output amplitude under typical operating conditions does not overfit terms already accurately described by the rigid body model.

[0054] The neural network training employs a joint physical constraint loss function, which includes a data consistency term, a Stribeck prior consistency term, an odd symmetry constraint term, a dissipative constraint term, and a compensation nullification penalty term, to ensure that the output satisfies the frictional physical properties. The joint physical constraint loss function is: in, For the joint loss letter, This is a data consistency item to ensure that the network output is consistent with the residual data; This represents the total residual moment vector predicted by the network. This is the measured residual torque vector; As an odd symmetry constraint term, it requires that the principal friction network satisfy odd symmetry with respect to the velocity sign, which is consistent with the physical nature of friction; The main friction network outputs torque to the forward and reverse velocities. The term represents a dissipative constraint to ensure that friction does not have positive power at any sample point, preventing the network from generating physics-incorrect outputs; ⊙ represents the Hadamard product. The power is the element-wise product of frictional torque and velocity. For Stribeck's prior consistency terms, The derivative (slope) of the main friction network with respect to velocity. For rectified linear units, only positive slopes are retained, and negative slopes are set to 0. For the Stribeck characteristic velocity, the main friction network is in the Stribeck velocity range. Applying monotonically decreasing soft constraints helps the network maintain physically reasonable outreach behavior in sparse data regions. To compensate for the nulling penalty term, L2 regularization is applied to the magnitude of the output of the compensation network to avoid redundant fitting that cancels out with the main friction network. To compensate for the output torque of the network; , , , , These are the weighting coefficients for each loss term.

[0055] After training on an offline dataset, the network weights of the dual-stream PINN are fixed and exported in a form that can be called by the real-time inference module. During the online control phase, the module calculates the residual compensation torque in real time via forward propagation, and the network weights remain unchanged throughout the entire online operation.

[0056] This application employs a dual-input structure of a two-stream physical information neural network, which can efficiently process robot joint state data and dynamic basis parameter identification results separately, achieving deep fusion of the two types of data. Compared with single-input networks, it can more comprehensively capture robot dynamic characteristics and improve the fitting accuracy of dynamic residuals. The training method incorporating a physical constraint loss function forces the network output to conform to the basic laws of robot dynamics, effectively avoiding residual prediction results that do not conform to physical logic, reducing the risk of model overfitting, and improving the network's generalization ability and reliability. After training, the network weights are fixed and can be directly used as a feedforward module without repeated training, reducing computational overhead. Simultaneously, it ensures the consistency of residual compensation, providing accurate and stable supplementary corrections for joint external torque estimation, further reducing external force estimation errors.

[0057] In this embodiment, preferably, the generalized momentum observer (GMO) uses low-speed dead zone linear smoothing to process the explicit friction torque and suppress zero-crossing chattering.

[0058] Specifically, the generalized momentum observer utilizes generalized momentum. As an integral state, the external torque is estimated by monitoring the deviation between its actual evolution and the model's predicted evolution, thus avoiding the noise problem caused by direct numerical differentiation calculation of acceleration. Figure 5 As shown, the enhanced generalized momentum observer integral loop is driven by both the rigid body dynamics term and the PINN residual feedforward, and is equipped with an integral limiting module to prevent integral drift under abnormal operating conditions.

[0059] In this application, the observer's integral driving term The calculation formula is: in, For the first Real-time measured joint driving torque vector The Coriolis moment vector, The gravitational torque vector For the first The explicit friction torque vector at any given time. For the first The PINN residual compensation torque vector at time step. Compared to the traditional method, this application additionally eliminates two items: explicit friction torque. And PINN residual compensation moment estimation The explicit feedforward injection of these two items makes the unexplained torque observed by the observer closer to the true external torque, thereby reducing the bias of the external force estimation.

[0060] In this application, implicit discretization is used to update the integral state, and amplitude limiting is applied to the internal state to prevent integral drift. The estimated external torque of the joint is calculated from the deviation between the generalized momentum and the internal state of the observer: in, For the first The internal state vector of the time-matter observer. It is the identity matrix. For control period (sampling period). For the first Generalized momentum vector at any moment This is the estimated vector of external joint torques; The observer gain matrix can be configured as a scalar, vector, or full matrix. To prevent unbounded drift of numerical integration under abnormal operating conditions, this application specifies the integration state... Amplitude limiting was applied.

[0061] In this application, the explicit friction term When the velocity crosses zero, due to directional uncertainty, directly using the tanh-type Coulomb friction approximation may still result in a large slope near zero velocity, inducing chattering. Therefore, this application introduces a low-speed dead-zone linear transition mechanism: when the absolute value of the joint velocity is below the dead-zone threshold... At that time, from the origin to the dead zone boundary point The slope of the secant line is used to make a linear approximation of the frictional force: in, For the first The explicit frictional torque of each joint. , These are the viscous friction coefficient and the Coulomb friction coefficient, respectively. It is the hyperbolic tangent function. Let tanh be the shape parameter. This is the dead zone threshold; When the absolute value of the joint velocity is lower than The original nonlinear friction function is replaced with a linear slope to ensure that the friction force is continuous and the slope is bounded when it crosses zero, so as to suppress the chattering of the external force estimation. This treatment ensures that the frictional force remains continuous and has a bounded slope when crossing zero, effectively suppressing chattering in the low-speed region of the external force estimation.

[0062] This application feeds forward a dual-stream physical information neural network with fixed weights into a generalized momentum observer, achieving a dual optimization of "model prediction + residual compensation." This effectively overcomes the shortcomings of traditional generalized momentum observers, which rely solely on rigid body dynamics models and cannot compensate for non-rigid body effects, significantly improving the accuracy of joint external torque estimation and dynamic response speed. This solution eliminates the need for additional force sensors; accurate external torque estimation can be achieved solely through existing data acquisition modules and algorithm optimization. This significantly reduces the hardware cost and deployment complexity of the robot system, while avoiding the space limitations, cumbersome wiring, and susceptibility to environmental interference associated with force sensor installation. This enhances the applicability of the method in complex industrial scenarios and solves the core technical challenge of directly measuring external torque in sensorless environments.

[0063] In this embodiment, preferably, the contact state machine uses a hysteresis interval where the opening threshold is greater than the closing threshold, combined with anti-shake counting to determine the contact state, and outputs a continuous and smooth contact mixing factor through a first-order low-pass filter.

[0064] Specifically, the estimated joint external torque It needs to be mapped to the end-effector Cartesian space before it can be used for admittance control. This is achieved using the robot's Jacobian matrix. A mapping relationship between joint torque and end-effector rotation can be established: ,right The end-force spinor estimate can be obtained by finding the pseudo-inverse. Considering that the condition number of the Jacobian matrix increases sharply near singular configurations, a direct pseudo-inverse would lead to severe distortion of the mapping results. Therefore, this application employs a damped least squares (DLS) pseudo-inverse: Damping factor The introduction of regularizes the solution near the singular configuration, ensuring the boundedness of the mapping result.

[0065] Mapped end force After passing through dead-zone filtering (setting force components below a set threshold to zero to eliminate measurement and estimation noise when stationary) and low-pass filtering (smoothing high-frequency jitter through first-order inertial filtering), a reliable end-contact force estimate is obtained. .

[0066] The contact state machine uses the projected component of the end contact force in the normal direction as the judgment scalar. Robust contact judgment is achieved through a finite state machine combining hysteresis threshold and anti-jitter counting: when the force scalar continuously exceeds the opening threshold... When the number of consecutive frames reaches the stabilization count threshold, the state machine determines that contact has been established; when the force scalar is continuously below the shutdown threshold... When the continuous frame rate reaches the stabilization count threshold, the contact is determined to have been released. Hysteresis interval The settings ensure that the state machine is insensitive to noise, while the anti-jitter counting further avoids misjudgments caused by brief force fluctuations.

[0067] To make the switching of admittance parameters smoother, the contact mixing factor output by the state machine is... Instead of directly outputting a 0 / 1 step signal, a smooth transition is achieved through a first-order low-pass filter. in, For the first Constantly in contact with the mixed factors, For the first The target value corresponding to each moment is 1 when contact is made and 0 when separation is made; These are the filter coefficients. This is a configurable time constant. This smoothing mechanism prevents abrupt shocks when the admittance parameter switches.

[0068] This application employs a damped least squares pseudo-inverse algorithm to achieve the mapping transformation from joint external torque to end-effector Cartesian contact force. Compared to traditional pseudo-inverse algorithms, by introducing a reasonable damping term, it not only effectively avoids the mapping failure problem caused by robot kinematic singularities, ensuring the continuity and stability of the mapping process, but also suppresses the influence of measurement noise and external torque estimation errors on the contact force calculation results, improving the smoothness and accuracy of the calculation results. This optimized scheme can directly output the contact force data of the robot end-effector in the Cartesian coordinate system, providing a direct and reliable force feedback basis for subsequent contact control without the need for additional coordinate transformation calibration, simplifying the control process, while ensuring the real-time performance of contact force feedback, laying the foundation for adaptive compliant contact control.

[0069] In this embodiment, preferably, admittance control achieves compliant behavior by converting end-contact force information into pose correction amounts, such as... Figure 6 As shown, the three-stage variable damping adaptive admittance controller comprises three cooperating sub-modules: a parameter scheduling module, an adaptive damping law module, and a target force ramp filter. Its core admittance equation is: in This represents the end-effector pose deviation state. , , These are the target inertia matrix, target damping matrix, and target stiffness matrix, respectively. The target contact force (effective during the steady-state processing phase).

[0070] In the free phase, high stiffness and low damping parameters are used to ensure trajectory tracking. Specifically, the robot has not yet made contact with the workpiece, and the contact mixing factor... The admittance parameters are selected from a set of free parameters with high equivalent stiffness and low damping to ensure that the robot can accurately track the planned trajectory and smoothly approach the workpiece.

[0071] During the collision phase, admittance parameters are linearly interpolated using the contact mixing factor to absorb the impact. Specifically, when contact mixing factor During the transition from 0 to 1, , , All of them vary between the free parameter set and the collision parameter set. Linear interpolation: in, The target parameter matrix of the admittance controller. The admittance parameter matrix for the free stage. This is the admittance parameter matrix for the contact phase; The collision parameter set employs a large damping to quickly dissipate the impact kinetic energy at the moment of contact, preventing bouncing and overshoot at the end.

[0072] In the steady state phase, constant force control is achieved by adaptively adjusting the damping based on the force tracking error. Specifically, after contact stabilization, the admittance parameter is switched to the working parameter group, and then online adaptive adjustment is performed based on the force tracking error: in, This is the current adaptive damping matrix. The steady-state reference damping matrix is... For diagonalization operators, For adaptive gain matrix / vector, For the force tracking error vector, The estimated end Cartesian external force vector; For force / bit decoupling mask vector, For the target contact force vector, , The upper and lower limit matrices of the damping are used to ensure that the damping parameters vary within a safe range. When the force tracking error is large, the adaptive mechanism will increase the damping to suppress response overshoot; when the error approaches zero, the damping will fall back to the basic operating parameters to maintain sufficient response bandwidth.

[0073] In practical contact machining, robots typically need to maintain target contact force in the normal direction while maintaining precise position control along the machining trajectory in the tangential direction. To address this, this application introduces a force / position decoupling mask vector. (For example, for Z-axis normal force control scenarios) Force error feedback is introduced only in the normal force control channel specified by the mask to participate in admittance calculation, while the other channels maintain pure position tracking.

[0074] To prevent the target force from jumping from zero to the desired value and causing an impact excitation on the admittance equation, the target force signal is ramped up through a first-order inertial filter before it takes effect, and its time constant can be configured according to process requirements.

[0075] This application employs a hysteresis-resistant contact state machine design, which can accurately identify the contact state between the robot's end effector and the environment (no contact, just contact, stable contact, and disengagement). By outputting a smooth contact mixing factor, it effectively suppresses control oscillations and jitters caused by abrupt changes in contact state, avoiding impact phenomena during contact and protecting the manipulated object and the robot's end effector from damage. A three-stage variable damping admittance controller adaptively adjusts the damping parameters based on the contact mixing factor. High damping in the no-contact stage ensures robot motion stability, gradual damping in the contact transition stage achieves a smooth transition, and optimal damping in the stable contact stage achieves high-precision compliant control, adapting to the control requirements of different contact conditions. The overall optimized scheme achieves adaptive compliant contact between the robot and the environment, improving the stability and accuracy of contact control. It can be widely adapted to scenarios with high contact control requirements, such as precision assembly, human-robot collaboration, and flexible grinding, expanding the application range of robots.

[0076] Taking constant force grinding on a flat surface as an example, when the robot moves along the planned grinding trajectory, it executes the following process in real time in a loop: reading the current joint angle. Joint velocity and joint driving torque The residual compensation torque estimate is obtained by calling the PINN forward inference with fixed weights. Perform GMO integral update to obtain joint external torque estimate. Calculate the current Jacobian matrix and obtain the end contact force using the DLS pseudo-inverse mapping. The system performs dead-zone and low-pass filtering; updates the contact state machine; and calculates the contact mixing factor. Based on the current stage indicators and hybrid factors, admittance parameters are selected, and adaptive damping updates are performed during the working phase. Force control is applied to the machining normal channel, and position tracking is applied to the other channels. The admittance equation is solved to obtain the end-effector pose correction. The corrected desired pose, velocity, and acceleration commands are sent to the servo controller for execution.

[0077] Through the above cycle, the robot can maintain accurate trajectory tracking when approaching the workpiece, quickly absorb the impact at the moment of contact, and adaptively stabilize the contact force near the target value during the steady-state grinding stage, without the need for a six-dimensional force sensor.

[0078] Taking surface-following grinding and polishing as another embodiment: the normal in the tool coordinate system changes in real time with the end pose. The force / position decoupling mask and the force projection mechanism of the tool coordinate system of the present invention can automatically adapt to this change, ensuring the consistency of normal force control on the entire surface, while the trajectory tracking accuracy along the surface tangential is not affected by the normal force control.

[0079] One embodiment of this application also provides a computer device, including but not limited to a processor and a memory, wherein the memory stores computer instructions, and the processor executes the computer instructions to implement the method described above.

[0080] One embodiment of this application also provides a computer-readable storage medium for storing a computer program that, when executed by a processor, implements the method described in the above-described method embodiments.

Claims

1. A method for estimating external force and controlling admittance of a robot's powerless sensor, characterized in that, include: The robot is driven to move along a Fourier series excitation trajectory and joint state and torque data are collected. Robust identification of rigid body dynamics basis parameters is achieved based on iterative weighted least squares. A two-stream physical information neural network is constructed to fit the dynamic residuals, and the network weights are fixed under the physical constraint loss function. By feeding forward a dual-stream physical information neural network with fixed weights into a generalized momentum observer, joint external torque estimation can be achieved. The Cartesian contact force at the end is obtained through damped least-squares pseudo-inverse mapping. The smooth contact mixing factor is output by the hysteresis anti-jitter contact state machine, which drives the three-stage variable damping admittance controller to achieve adaptive compliant contact control.

2. The method for estimating external force and admittance control of a robot's powerless sensor according to claim 1, characterized in that: The Fourier series excitation trajectory is generated by maximizing the excitation index of the dynamic regression matrix. The trajectory covers the target workspace and satisfies the joint motion constraints. The beginning and end of the trajectory are spliced ​​together with static segments to ensure that the starting and stopping speeds and accelerations are both zero.

3. The method for estimating the external force and admittance control of a robot's powerless sensor according to claim 2, characterized in that: The iterative weighted least squares identification uses the Tukey double-weight function to remove abnormal samples and adds a velocity-related suppression factor to reduce the weight of samples in the low-speed zero-crossing interval and restore the full weight of samples at stationary points. The Tukey double-weight function is: in, For the first The weights of each sample, For the first Standardized residuals of each sample For the first The joint torque residuals of each sample; The robust scaling parameters are estimated using the median absolute deviation. This is the adjustment constant; The threshold for outlier detection is set when the residual amplitude exceeds [a certain threshold]. Abnormal samples are reset to zero; The velocity-related inhibition factor is: in, For the first The joint angular velocity of each sample The velocity scale parameter is used to reduce the weight of high-uncertainty samples in the low-speed zero-crossing interval, and to restore the weight contribution of samples when both joint velocity and acceleration are below the static determination threshold.

4. The method for estimating the external force and admittance control of a robot's powerless sensor according to claim 3, characterized in that: The dual-stream physical information neural network includes a main friction network and a compensation network. The main friction network fits Stribeck friction with joint angular velocity as input. in, For the first The main friction network of each joint outputs torque. Main friction neural network, For the first Angular velocity of each joint; The compensation network fits the unmodeled dynamics using joint position, velocity, and acceleration as inputs; in, For the first The joint compensation network outputs torque. Compensation neural network, For the first Angular position and angular acceleration of each joint; The dynamic residual compensation torque is obtained by superimposing the two outputs; in, For the first The total dynamic residual compensation torque of each joint; The neural network training employs a joint physical constraint loss function, which includes a data consistency term, a Stribeck prior consistency term, an odd symmetry constraint term, a dissipative constraint term, and a compensation nullification penalty term, to ensure that the output satisfies the frictional physical properties. The joint physical constraint loss function is: in, For the joint loss letter, For data consistency items; This represents the total residual moment vector predicted by the network. This is the measured residual torque vector; As an odd-symmetric constraint, the principal friction network is required to satisfy odd symmetry with respect to the velocity sign. The main friction network outputs torque to the forward and reverse velocities. The term represents a dissipative constraint to ensure that friction does not have a positive power at any sample point. ⊙ represents the Hadamard product. The power is the element-wise product of frictional torque and velocity. For Stribeck's prior consistency terms, The derivative (slope) of the main friction network with respect to velocity. For rectified linear units, only positive slopes are retained, and negative slopes are set to 0. For the Stribeck characteristic velocity, the main friction network is in the Stribeck velocity range. Apply a monotonically decreasing soft constraint internally; To compensate for the zeroing penalty, To compensate for the output torque of the network; , , , , These are the weighting coefficients for each loss term.

5. The method for estimating external force and admittance control of a robot's powerless sensor according to claim 4, characterized in that: The generalized momentum observer employs low-speed dead zone linear smoothing to process explicit friction torque and suppress zero-crossing chatter. in, For the first The explicit frictional torque of each joint. , These are the viscous friction coefficient and the Coulomb friction coefficient, respectively. It is the hyperbolic tangent function. For shape parameters, This is the dead zone threshold; When the absolute value of the joint velocity is lower than The original nonlinear friction function is replaced with a linear slope to ensure that the friction force is continuous and the slope is bounded when it crosses zero, so as to suppress the chattering of the external force estimation. Implicit discretization is used to update the integral state, and amplitude limiting is applied to the internal state to prevent integral drift. in, For the first The internal state vector of the time-matter observer. It is the identity matrix. For control period (sampling period). For the first Generalized momentum vector at any moment The observer gain matrix; This is the estimated vector of external joint torques; Integral-driven items The calculation formula is: in, For the first Real-time measured joint driving torque vector The Coriolis moment vector, The gravitational torque vector For the first The explicit friction torque vector at any given time. For the first The residual compensation torque vector of PINN at time step.

6. The method for estimating external force and admittance control of a robot's powerless sensor according to claim 5, characterized in that: The contact state machine uses a hysteresis interval where the opening threshold is greater than the closing threshold, combined with anti-shake counting to determine the contact state, and outputs a continuous and smooth contact mixing factor through a first-order low-pass filter. in, For the first Constantly in contact with the mixed factors, For the first The target value corresponding to each moment is 1 when contact is made and 0 when separation is made; These are the filter coefficients. This is a configurable time constant.

7. The method for estimating the external force and admittance control of a robot's powerless sensor according to claim 6, characterized in that, The three-stage variable damping admittance control includes: High stiffness and low damping parameters are used in the free phase to ensure trajectory tracking; During the collision phase, admittance parameters are linearly interpolated using the contact mixing factor to absorb the impact. in, The target parameter matrix of the admittance controller. For contact mixing factor, The admittance parameter matrix for the free stage. This is the admittance parameter matrix for the contact phase. Let the target inertia matrix be... Let be the target damping matrix. The target stiffness matrix; In the steady-state phase, constant force control is achieved by adaptively adjusting the damping based on the force tracking error. in, This is the current adaptive damping matrix. The steady-state reference damping matrix is... For diagonalization operators, For adaptive gain matrix / vector, For force / bit decoupling mask vector, For the force tracking error vector, For the estimated end Cartesian external force vector, For the target contact force vector, , Let be the upper and lower limit matrices for damping.

8. A robot force estimation and admittance control system based on a powerless sensor, characterized in that, include: The system comprises a data acquisition module, a robust identification module, a dual-stream physical information neural network modeling module, an external force estimation module, a force mapping and contact recognition module, a variable damping admittance control module, and an execution module; the system is used to execute the robot force sensor external force estimation and admittance control method as described in any one of claims 1 to 7.

9. The robot force sensor external force estimation and admittance control system according to claim 8, characterized in that, The system hardware includes: a real-time control computer, joint servo drivers, the robot body, an encoder acquisition unit, and a current sampling unit, which complete command interaction through a real-time bus.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the robot force estimation and admittance control method for unpowered sensors as described in any one of claims 1-7.

Images