Collaborative robot adaptive compliance control method based on impedance incremental learning

Abstract

The invention relates to a cooperative robot adaptive compliance control method based on impedance incremental learning. The method comprises the following steps: establishing a robot dynamic model and defining a disturbance term; designing a fuzzy adaptive superspiral disturbance observer to estimate disturbance in real time; an admittance parameter is optimized online through a finite time admittance filter in combination with an impedance incremental learning mechanism, and a compliant expected trajectory is generated; a layered sliding mode controller is adopted to realize high-precision tracking, and buffeting and singularity are avoided; finally, a control moment is synthesized through a composite control law, and feedforward compensation, disturbance estimation and neural network online compensation are fused. Impedance parameters can be adaptively adjusted according to interaction characteristics of different operators, personalized compliant control is realized, the problem that a traditional method depends on manual parameter adjustment and cannot adapt to individual differences is effectively solved, and safety and naturalness of man-machine cooperation are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of robot control, and particularly relates to a collaborative robot adaptive compliant control method based on impedance incremental learning. BACKGROUND

[0002] Collaborative robots are increasingly widely used in the fields of industry, medicine, service, etc., and the core requirement is to have high compliance and adaptive ability to adapt to different operators' interactive intentions and behavior characteristics. Traditional human-machine interaction control methods mostly use fixed-parameter admittance control or impedance control, which is difficult to cope with the challenges brought by model uncertainty, external disturbance and operator individual differences. In the prior art, although disturbance observers or sliding mode controls are introduced to enhance robustness, there are still the following deficiencies: 1. The admittance parameters usually need to be set in advance by experience, and cannot be adjusted adaptively online; 2. There is a lack of perception and learning ability for individualized interactive characteristics of operators; 3. Chattering or slow convergence speed problems are prone to occur in the control process. Therefore, an intelligent control method capable of learning operator preferences online, adaptively adjusting impedance characteristics and having strong anti-disturbance ability is urgently needed to improve the flexibility, safety and efficiency of human-machine collaboration. SUMMARY

[0003] The application provides a collaborative robot adaptive compliant control method based on impedance incremental learning, which comprises the following steps: System modeling and signal acquisition: collect robot joint motion information and human-machine interaction torque, establish a robot dynamics model containing uncertainty and define a comprehensive disturbance term; Disturbance observation: adopt a super-spiral observer structure, and introduce a fuzzy adaptive gain adjustment mechanism based on observation error to calculate and output disturbance estimation value in real time; Admittance filtering and impedance learning: according to the disturbance estimation value, generate an expected trajectory through an admittance equation; at the same time, based on historical interaction data, trajectory tracking error and operator identity, calculate the admittance parameter gradient through a multi-objective performance index function, and combine the individualized reference parameters to incrementally adjust the inertia, damping and stiffness of the three admittance parameters online; Hierarchical sliding mode control: calculate the tracking error according to the actual trajectory and the expected trajectory, construct a first-layer sliding mode surface for fast tracking, and further construct a second-layer sliding mode surface containing fractional differential, integral and continuous robust terms for disturbance suppression, and finally generate a nominal control quantity; Composite control law synthesis: synthesize the nominal control quantity, robot nominal dynamics feedforward compensation, disturbance estimation value and neural network adaptive compensation term to synthesize the final control torque output to the robot; Data storage and update: store the current interaction process data for updating the individualized reference parameters associated with the operator identity.

[0004] Specifically, in the disturbance observation step, the super-twisting observer updates the state variable and the virtual control variable, and the gain is a time-varying parameter, which is dynamically adjusted according to the absolute value of the observation error by fuzzy rules. When the error is large, the gain increases, and when the error is small, the gain decreases.

[0005] Specifically, in the admittance filtering and impedance learning step, the admittance equation is: where γ(t) is a time-varying index, M d ,C d ,K d are admittance parameters corresponding to inertia, damping, and stiffness, respectively.

[0006] Specifically, the incremental adjustment law of the admittance parameter is: where η is the learning rate, and β is the personalized weight used to adjust the operator preference, M d ref (u id ), C d ref (u id ), K d ref (u id ) are personalized reference admittance parameters, u id is the identifier, and J is the multi-objective optimization function.

[0007] Specifically, the multi-objective optimization function is: where T is the time window length, and α1, α2, and α3 are weight coefficients.

[0008] Specifically, in the hierarchical sliding mode control step, the first layer sliding mode surface is composed of the derivative of the tracking error and the nonlinear function of the tracking error.

[0009] Specifically, the second layer sliding mode surface is composed of the fractional order derivative of the first layer sliding mode surface, its time integral, and a continuous robust term processed by a smoothing factor.

[0010] Specifically, the calculation of the nominal control variable is determined by the expected acceleration, the nonlinear feedback term corresponding to the first layer sliding mode surface, and the weighted term of the second layer sliding mode surface.

[0011] Specifically, the weight matrix of the neural network compensation term is adaptively updated according to the product of the second layer sliding mode surface and the radial basis function network basis function vector with a negative definite learning rate matrix.

[0012] Beneficial technical effects: It can autonomously learn the force preferences and movement habits of different operators online, dynamically adjust the admittance parameters to generate a smooth and efficient desired trajectory, significantly improving the naturalness and safety of human-robot collaboration; by adopting a strategy combining a fuzzy adaptive superspiral observer and a fractional integral sliding mode surface, it achieves rapid estimation and compensation of complex uncertainties without the need for precise upper bounds of perturbations, greatly enhancing the robustness and response speed of the system; by fusing feedforward compensation, perturbation observation and online neural network learning through a composite control law, it effectively reduces the dependence on precise dynamic models and avoids control chattering. Ultimately, while ensuring high-precision trajectory tracking, it enables the robot to actively adapt to different operators and changing task scenarios, significantly improving the intelligence level of collaborative robots and the human-robot interaction experience. Attached Figure Description

[0013] Appendix Figure 1 The flowchart shows the adaptive compliant control method for cooperative robots based on impedance increment learning according to the present invention.

[0014] Appendix Figure 2 This is a schematic diagram of a mass-spring-damping system. Detailed Implementation

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

[0016] like Figure 1 As shown, this invention provides an adaptive compliant control method for collaborative robots based on impedance increment learning, specifically including: 1. System Modeling and Signal Acquisition Collect robot joint angle q, joint angular velocity, and joint angular acceleration; obtain human-machine interaction torque τ through torque sensors. h Establish robot dynamics equations that include uncertainties: The disturbance term is defined to include model parameter errors (such as mass estimation errors) and unmodeled dynamics (such as external shocks), specifically: Where M is the inertia matrix, C is the Coriolis force matrix, G is the gravity term, F is the friction term, ξ(t) is the unmodeled dynamics, and τ c To control the torque.

[0017] 2. Disturbance observer: calculate disturbance estimation. Specifically: initialize observer state, state variables include torque tracking state z0, disturbance estimation state z1, virtual control v0, real-time calculation of torque tracking error e0, update virtual control v0, update z0 and update z1. Where the observer update rate is specifically: .

[0018] The final output disturbance estimation value is equal to z1. Observer gains Λ0, Λ1, Λ2 are fuzzy gain adjustment, which satisfies: Where, Λ i0 is the base gain of gain Λ i , ΔΛ i is the gain adjustment amplitude, determined by the output of the fuzzy logic controller, μ(|e0|) is the fuzzy membership function, so that the gain Λ0, Λ1, Λ2 dynamically adjusts with |e0|, the fuzzy rule is that |e0| is large, then the gain increases, |e0| is small, then the gain decreases.

[0019] 3. Finite time admittance filter: calculate performance index according to current interaction force and tracking error, query or update the operator's personalized parameter benchmark, calculate admittance parameter increment and update admittance parameter, and generate new expected trajectory.

[0020] Input the estimated interaction torque, convert human force into compliant motion trajectory through the admittance equation, and adaptively adjust the impedance characteristics. Referring to the model shown in Figure 2 The admittance equation is:

[0021] Where the time-varying index γ(t) is: .

[0022] Where, τ h * is the interaction force threshold, which is a preset constant. Through this setting, γ(t) →1 (smooth) at high interaction force, and →0.5 (fast response) at low interaction force.

[0023] Where, M d , C d , K d are admittance parameters, corresponding to inertia, damping and stiffness respectively, where the inertia coefficient M d controls the start / stop sensitivity, the damping coefficient C d determines the motion fluency, and the stiffness coefficient K d affects the position tracking stiffness, which satisfies:

[0024] where k is the update amplitude, used to prevent parameter mutation, sat is a saturation function, used to limit the update amplitude of the parameter to prevent system instability due to a single update being too large. The increment is: where η is the learning rate, used to control the update speed of the increment, β is the individualization weight, used to adjust the operator preference, M d ref (u id ), C d ref (u id ), K d ref (u id ) is the individualized reference admittance parameter, which stores different operator preference parameters, such as a doctor requiring high compliance and a worker requiring high stiffness, u id is an identifier that can be used to retrieve a pre-stored individualized parameter set. The system can provide default parameters to new users and learn and associate their u id stores the optimal parameters, and J is a multi-objective optimization function aimed at minimizing the weighted integral of the interaction torque (compliance), trajectory tracking error (accuracy), and desired acceleration (smoothness), which is specifically: where T is the time window length, a1, a2, a3 are weight coefficients used to balance the torque / error / acceleration, so that J(t) optimizes the interaction force, tracking error, and motion smoothness at the same time, e is the tracking error, e = q - q d , i.e., the difference between the current joint angle and the desired joint angle.

[0025] At first use, the admittance parameters can use default parameters, and thereafter, through online learning, individualized parameters are stored; when used again, historical parameters can be called, and incremental optimization is continued.

[0026] 4. Hierarchical sliding mode controller: calculate trajectory error, build double-layer sliding mode surface, generate nominal control quantity. Realize finite time high precision tracking of trajectory, avoid chattering and singularity of traditional sliding mode.

[0027] where the first layer sliding mode surface is used for rapid convergence of position error, and is specifically: ; The second layer sliding mode surface is used for disturbance suppression and dynamic compensation, and is specifically:

[0028] A hierarchical structure is designed, s1 focuses on trajectory tracking, and s2 handles disturbances; fractional order integration is used to enhance anti-interference, an integral term is introduced to suppress steady-state error, and a continuous robust term is introduced to avoid chattering.

[0029] The control law is a nominal control quantity u nom Specifically,

[0030] Wherein, the position error convergence index γ1 is used for balancing the convergence speed, and γ1 can be set as 0.7;D t ν Caputo fractional derivative is represented, wherein the fractional order ν can be set as 0.8, λ, ρ, δ are integral gain, robust gain, smoothing factor respectively, and k1, k2 are control gains 5. Compound control law: calculating model compensation term, neural network online compensating model uncertainty, synthesizing final control torque.

[0031] The final control torque includes a feedforward term, a model compensation term, a disturbance compensation term and a neural network compensation term, and is specifically:

[0032] The neural network compensation term is:

[0033] Wherein, W is a neural network weight matrix, Φ(·) represents an RBF basis function, which can adopt a Gaussian function, and Γ is an adaptive learning rate.

[0034] 6. Executing the control command, updating the system state, and storing the current interaction data for next learning.

[0035] The application realizes incremental online optimization through personalized impedance learning, adjusts parameters according to real-time interaction data, simultaneously associates operator characteristics, binds a personalized parameter library, and simultaneously optimizes operating force, tracking accuracy and motion smoothness; Through hierarchical anti-interference control, the first layer focuses on trajectory tracking, the second layer processes disturbances, and through fractional order + integral compound design, low-frequency disturbance suppression is enhanced; Impedance parameter personalization is performed at the front end, sliding mode gain adaptation is performed at the middle end, and neural network online compensation is performed at the back end, forming a "interaction-learning-control" closed-loop optimization, realizing autonomous evolution of impedance characteristics in human-machine collaboration, and solving the problem that traditional methods need manual parameter adjustment and cannot adapt to individual differences.

[0036] In several embodiments provided by the application, it should be understood that the disclosed method can be implemented in other ways. For example, the above-described application embodiments are merely illustrative, for example, the division of modules is only a logical function division, and actual implementation can have another division manner.

[0037] The modules described as separate components may or may not be physically separate, and the components displayed as modules may or may not be physical modules, and may be located in one place or distributed to multiple network modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0038] In addition, each functional module in each embodiment of the present application can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module. The integrated module can be realized in the form of hardware or in the form of hardware plus software functional module.

[0039] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the essential characteristics of the present application.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for adaptive compliant control of a collaborative robot based on impedance increment learning, characterized in that, The method comprises the following steps: System modeling and signal acquisition: collect the joint motion information and human-robot interaction torque of the robot, establish a robot dynamics model containing uncertainties, and define a comprehensive disturbance term; Disturbance observation: adopt a super-spiral observer structure, introduce a fuzzy adaptive gain adjustment mechanism based on observation error, and calculate and output the disturbance estimate value in real time; Admittance filtering and impedance learning: according to the disturbance estimate value, generate the expected trajectory through the admittance equation; at the same time, based on the historical interaction data, the trajectory tracking error and the operator identity, calculate the admittance parameter gradient through the multi-objective performance index function, and combine the personalized reference parameters to incrementally adjust the inertia, damping and stiffness three admittance parameters online; Hierarchical sliding mode control: calculate the tracking error according to the actual trajectory and the expected trajectory, construct the first layer sliding mode surface for fast tracking, and further construct the second layer sliding mode surface containing fractional order differential, integral and continuous robust term for disturbance suppression, and finally generate the nominal control quantity; Composite control law synthesis: synthesize the nominal control quantity, the robot nominal dynamics feedforward compensation, the disturbance estimate value and the neural network adaptive compensation term to synthesize the final control torque output to the robot; Data storage and update: store the current interaction process data for updating the personalized reference parameters associated with the operator identity.

2. The method of claim 1, wherein, In the disturbance observation step, the super-spiral observer updates the state variable and the virtual control quantity, and the gain is a time-varying parameter, which is dynamically adjusted by fuzzy rules according to the absolute value of the observation error. When the error is large, the gain increases, and when the error is small, the gain decreases.

3. The method of claim 1, wherein, In the admittance filtering and impedance learning step, the admittance equation is: where γ(t) is a time-varying exponent, M d ,C d ,K d are admittance parameters corresponding to inertia, damping, stiffness, respectively, q is the joint angle, is the estimated interaction torque.

4. The method of claim 3, wherein, The incremental adjustment law of the admittance parameter is: where η is a learning rate, β is a personalization weight for adjusting the operator preference, M d ref (u id ), C d ref (u id ), K d ref (u id ) is a personalization reference admittance parameter, u id is an identifier, and J is a multi-objective optimization function.

5. The method of claim 4, wherein, The multi-objective optimization function is: where T is the time window length, a1, a2, a3 are weight coefficients, e is the tracking error, e = q - q d i.e. the difference between the current joint angle and the desired joint angle.

6. The method of claim 1, wherein, In the hierarchical sliding mode control step, the first layer sliding mode surface is composed of the derivative of the tracking error and the nonlinear function of the tracking error.

7. The method of claim 1, wherein, The second layer sliding mode surface is composed of the fractional order differential of the first layer sliding mode surface, the time integral thereof, and a continuous robust term processed by a smoothing factor.

8. The method of claim 1, wherein, The calculation of the nominal control quantity is determined by the expected acceleration, the nonlinear feedback term corresponding to the first layer sliding mode surface, and the weighted term of the second layer sliding mode surface.

9. The method of claim 1, wherein, The weight matrix of the neural network compensation term is adaptively updated according to the product of the second layer sliding mode surface and the radial basis function network basis function vector, and the negative definite learning rate matrix.

Images