Control method and device of robot system, electronic equipment and storage medium

Abstract

The application provides a control method and device of a robot system, electronic equipment and a storage medium, wherein the method comprises: obtaining an actual state at a current time and a desired state within a preset time period after the current time; calculating a nominal control law corresponding to the desired state within the preset time period and a nominal state based on a nominal model predictive control strategy, the nominal model predictive control strategy being constructed by introducing a sliding mode based on a nominal model of the robot system; determining an estimation deviation based on the actual state and the nominal state; determining an auxiliary control law corresponding to the estimation deviation based on an auxiliary controller; determining an actual control law within the preset time period based on the auxiliary control law and the nominal control law, and controlling the robot system based on the actual control law. The method, device, electronic equipment and storage medium provided by the application improve the accuracy of subsequent robot system control.

Description

Technical Field

[0001] The present invention relates to the field of robot intelligent control technology, and in particular to a control method, device, electronic equipment and storage medium for a robot system. Background Art

[0002] The robotic system is a complex multi-input and multi-output system that is affected by factors such as unmodeled dynamics, measurement errors, and external disturbances. It has characteristics such as high nonlinearity, time-varying, and strong coupling.

[0003] In the existing technology, the model predictive control method for robot systems cannot guarantee the control performance of the system under the conditions of relaxing the constraints on the terminal state. In addition, there are nonlinear uncertainties in the robot system model. Introducing nonlinear uncertainties into the optimization solution of model predictive control increases the complexity of solving the optimization problem. Summary of the Invention

[0004] The present invention provides a control method, device, electronic device, and storage medium for a robotic system. These methods address the inability of existing model predictive control methods for robotic systems to guarantee system control performance while relaxing constraints on terminal states. These methods also avoid incorporating nonlinear, uncertain models of the robotic system into the optimization solution, overcoming the high complexity of solving optimization problems in model predictive control methods for uncertain robotic systems.

[0005] The present invention provides a control method for a robot system, comprising:

[0006] Obtaining the actual state at the current moment and the expected state within a preset period after the current moment;

[0007] Calculating a nominal control law and a nominal state corresponding to the desired state within the preset time period based on a nominal model predictive control strategy, wherein the nominal model predictive control strategy is constructed by introducing a sliding mode based on a nominal model of the robot system;

[0008] determining an estimated deviation based on the actual state and the nominal state;

[0009] determining an auxiliary control law corresponding to the estimated deviation based on an auxiliary controller, wherein the auxiliary controller is a neural network that adaptively adjusts neuron nodes and / or weight values ​​based on the estimated deviation;

[0010] Based on the auxiliary control law and the nominal control law, an actual control law within the preset time period is determined, and based on the actual control law, the robot system is controlled.

[0011] According to a control method for a robot system provided by the present invention, determining the auxiliary control law corresponding to the estimated deviation based on the auxiliary controller includes:

[0012] Performing nonlinear mapping on the estimated deviation to obtain a mapping deviation;

[0013] Determining an activation rate based on the mapping deviation, and updating neuron nodes of the neural network based on the activation rate to obtain a neural network after node update;

[0014] Based on the neural network after the node update, an auxiliary control law corresponding to the mapping deviation is determined.

[0015] According to a control method for a robot system provided by the present invention, updating the neuron nodes of the neural network based on the activation rate includes:

[0016] The formula for determining the activation rate of the neural network is as follows:

[0017]

[0018] in, is a Gaussian function, representing the activation output of the jth neuron node, X + represents the augmented state quantity, which is determined based on the mapping deviation; l f Represents the number of neuron nodes in the middle layer of the current neural network, π F represents the maximum m of neurons in the middle layer f The average of the activation outputs;

[0019] In π F <Φ f In the case of adding a new intermediate layer neuron node, the formula of the Gaussian function of the newly added neuron node is as follows:

[0020]

[0021] in, Indicates the maximum m f The average of the Gaussian function means of the neurons that activate the output, o f ∈(0,1) represents an adjustable parameter, and the variance of the Gaussian function is σ f , Φ f is the preset threshold, the neural network activation output after the node is increased and the weight values ​​of the neural network The formula is as follows:

[0022]

[0023] in, Represents the connection weight of the newly added node, and the number of neuron nodes in the middle layer of the updated neural network is redefined as l f =l f +1.

[0024] According to a control method for a robot system provided by the present invention, determining the auxiliary control law corresponding to the mapping deviation based on the neural network after the node update includes:

[0025] The auxiliary control law corresponding to the mapping deviation is determined based on the following formula:

[0026]

[0027] in, represents the auxiliary control law at the current moment, represents the coordinate transformation of the mapping deviation, b2 represents the parameter in the nonlinear mapping, K2 represents the control gain, represents the weight value of the neural network, represents the activation function of the neural network, X + represents the augmented state quantity, yes The transposed matrix of .

[0028] According to a control method for a robot system provided by the present invention, the auxiliary control law corresponding to the mapping deviation is determined based on the neural network after the node update, and then further includes:

[0029] Based on the mapping deviation, the weight value of the neural network is updated, and the updated weight value is shown in the following formula:

[0030]

[0031] Among them, α f ,α B Represents the learning rate preset by the neural network, Indicates state-related parameters, The coordinate transformation representing the mapping deviation, yes The transposed matrix, η f >0,η B >0,k f >0,k B >0 indicates learning parameters.

[0032] According to a control method for a robot system provided by the present invention, the nominal model predictive control strategy is specifically:

[0033]

[0034] in, represents the state of the nominal model, represents the state-space form of the nominal model, represents the coefficient matrix, 0 n×n ,I n represent the zero matrix and identity matrix of order n, respectively. represents the augmented control quantity, Ω τ represents the nominal input constraint, L(s,τ mpc ,l|t k )=s T s represents the stage cost, Ψ(s)=s T Ps represents the terminal cost, P represents a positive definite symmetric matrix, s represents the sliding mode defined according to the nominal error, Represents a virtual state quantity.

[0035] According to a control method for a robot system provided by the present invention, the nominal model is specifically:

[0036]

[0037] in, Respectively represent t k The nominal joint angle, angular velocity, and angular acceleration at the moment, It is a designable matrix. τ mpc Represents the control quantity of the nominal model, and the initial value of the nominal model is selected as

[0038] The present invention also provides a control device for a robot system, comprising:

[0039] an acquisition unit, configured to acquire an actual state at a current moment and an expected state within a preset period after the current moment;

[0040] a calculation unit, configured to calculate a nominal control law and a nominal state corresponding to the desired state within the preset time period based on a nominal model predictive control strategy, wherein the nominal model predictive control strategy is constructed by introducing a sliding mode based on a nominal model of the robot system;

[0041] an estimation deviation determining unit, configured to determine an estimation deviation based on the actual state and the nominal state;

[0042] an auxiliary control law determination unit, configured to determine an auxiliary control law corresponding to the estimated deviation based on an auxiliary controller, wherein the auxiliary controller is a neural network that adaptively adjusts neuron nodes and / or weight values ​​based on the estimated deviation;

[0043] A control unit is configured to determine an actual control law in the preset time period based on the auxiliary control law and the nominal control law, and control the robot system based on the actual control law.

[0044] The application further provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the control method of the robot system according to any one of the above.

[0045] The application further provides a non-transitory computer readable storage medium, which stores a computer program, and the computer program is executable on a processor to implement the control method of the robot system according to any one of the above.

[0046] The application further provides a computer program product, which includes a computer program, and the computer program is executable on a processor to implement the control method of the robot system according to any one of the above.

[0047] The application provides the control method and device of the robot system, the electronic device and the storage medium, the actual control law in the preset time period is determined based on the auxiliary control law and the nominal control law, and the robot system is controlled based on the actual control law, the auxiliary controller is a neural network based on adaptive adjustment of estimated deviation of neuron nodes and / or weight values, therefore, the auxiliary control law output by the auxiliary controller can dynamically compensate for the non-limited uncertain items of the robot system, and the accuracy of subsequent robot system control is improved; and the nominal model predictive control strategy is constructed based on the sliding mode of the nominal model of the robot system, and the control performance of the system can be ensured under the relaxed constraint condition of the terminal state. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0049] Figure 1 is one of the flowcharts of the control method of the robot system provided by the present application;

[0050] Figure 2 is the second flowchart of the control method of the robot system provided by the present application;

[0051] Figure 3 is the structural schematic diagram of the control device of the robot system provided by the present application;

[0052] Figure 4 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0053] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0054] In related technologies, sliding mode control (SMC) is an efficient robust control method that can achieve high control accuracy. Furthermore, under SMC, the system state can converge along a defined sliding surface, which can be considered a state constraint.

[0055] In related technologies, research has been conducted on the combination of model predictive control and sliding mode control, but most of the work focuses on the following two aspects: (1) using model predictive control to solve the auxiliary control law of sliding mode control, so that the output trajectory of the system can converge to the sliding mode surface in an optimal way, improving the vibration problem in sliding mode control; (2) using model predictive control to solve the nominal control law of the nominal model, and using sliding mode control to estimate and compensate for the uncertainties or disturbances of the system.

[0056] In view of the above problems, the present invention provides a control method for a robot system. Figure 1 FIG. 1 is a flow chart of a control method for a robot system provided by the present invention, such as Figure 1 As shown, the method includes:

[0057] Step 110: Acquire the actual state at the current moment and the expected state within a preset period after the current moment.

[0058] Specifically, the actual state at the current moment and the expected state within a preset period after the current moment can be obtained. The actual state here refers to the actual position and actual speed of the robot arm controlled by the robot system. The preset period after the current moment here refers to a period of time preset with the current moment as the starting point. For example, t k is the current moment, then the preset time period after the current moment can be [t k ,t k +T). The desired state here refers to the desired position and desired velocity of the robot's manipulator controlled by the robotic system.

[0059] At step 120, a nominal model predictive control strategy is used to calculate a nominal control law and a nominal state corresponding to the desired state in the preset time period, and the nominal model predictive control strategy is constructed based on a nominal model of the robot system.

[0060] Specifically, after the desired state in the preset time period at the current time is obtained, the nominal model predictive control strategy is used to calculate the nominal control law and the nominal state corresponding to the desired state in the preset time period. The nominal control law here refers to a control instruction formed by the nominal model, and the nominal state here refers to a nominal joint position and a nominal joint speed.

[0061] The nominal model predictive control strategy here is constructed based on the nominal model of the robot system, so that the control performance of the system can be ensured under the condition that the constraint condition of the terminal state is relaxed.

[0062] At step 130, an estimated deviation is determined based on the actual state and the nominal state.

[0063] Specifically, after the actual state at the current time is obtained, the estimated deviation can be determined based on the actual state and the nominal state. The estimated deviation here refers to the deviation between the actual state and the nominal state, for example, may represent the actual position and the actual speed in the actual state at the current time, respectively, may represent the nominal joint position and the nominal joint speed at the current time, respectively, and then the estimated deviation x e is as follows:

[0064]

[0065] wherein, is the transpose matrix of x e1 , and is the transpose matrix of x e2 .

[0066] At step 140, an auxiliary control law corresponding to the estimated deviation is determined based on an auxiliary controller, and the auxiliary controller is a neural network that adaptively adjusts the neuron nodes and / or the weight values based on the estimated deviation.

[0067] Specifically, after the estimated deviation is determined, the auxiliary control law corresponding to the estimated deviation can be determined based on the auxiliary controller. The auxiliary controller here can be a neural network that adaptively adjusts the neuron nodes based on the estimated deviation, can be a neural network that adaptively adjusts the weight values based on the estimated deviation, or can be a neural network that adaptively adjusts the neuron nodes and the weight values based on the estimated deviation, and the present embodiment does not make a specific limitation in this regard.

[0068] The auxiliary control law here is the control instruction output by the auxiliary controller. The auxiliary controller is a neural network that adaptively adjusts neuron nodes and / or weight values ​​based on estimated deviations. Therefore, the auxiliary control law output by the auxiliary controller can dynamically compensate for the unrestricted uncertainties of the robot system, thereby improving the accuracy of subsequent robot system control.

[0069] Step 150 : determining an actual control law within the preset time period based on the auxiliary control law and the nominal control law, and controlling the robot system based on the actual control law.

[0070] Specifically, after obtaining the auxiliary control law and the nominal control law, the actual control law within a preset time period can be determined based on the auxiliary control law and the nominal control law, and the robot system can be controlled based on the actual control law.

[0071] Among them, the formula of the actual control law τ is as follows:

[0072]

[0073] in, represents the auxiliary control law, represents the nominal control law, represents the weight value of the auxiliary controller, represents the activation function of the auxiliary controller, Indicates the actual position in the actual state at the current moment.

[0074] For example, the actual control law can be input into the robot system, and the robot system outputs the actual position and actual speed of the robot's mechanical arm to realize the control of the robot system.

[0075] The method provided by an embodiment of the present invention determines the actual control law within a preset time period based on the auxiliary control law and the nominal control law, and controls the robot system based on the actual control law. The auxiliary controller is a neural network that adaptively adjusts neuron nodes and / or weight values ​​based on estimated deviations. Therefore, the auxiliary control law output by the auxiliary controller can dynamically compensate for the unrestricted uncertainties of the robot system, thereby improving the accuracy of subsequent robot system control; and the nominal model predictive control strategy is constructed by introducing a sliding mode based on the nominal model of the robot system, which can ensure the control performance of the system under the condition of relaxing the constraints on the terminal state.

[0076] Based on the above embodiment, step 140 includes:

[0077] Step 141 : Perform nonlinear mapping on the estimated deviation to obtain a mapping deviation.

[0078] Specifically, after determining the estimated deviation, nonlinear mapping can be performed on the estimated deviation to obtain a mapping deviation. The mapping deviation here reflects the degree of deviation at the mapping level. The nonlinear mapping here can use a sigmoid function, a tanh function, or a ReLU function, which is not specifically limited in this embodiment of the present invention.

[0079] Step 142: determining an activation rate based on the mapping deviation, and updating neuron nodes of the neural network based on the activation rate to obtain a neural network after node update.

[0080] Specifically, after obtaining the mapping deviation, the activation rate can be determined based on the mapping deviation, and the neuron nodes of the neural network can be updated based on the activation rate to obtain the neural network after the nodes are updated.

[0081] First, the formula for determining the activation rate of the neural network is as follows:

[0082]

[0083] in, is a Gaussian function, representing the activation output of the jth neuron node, X + represents the augmented state quantity, which is determined based on the mapping deviation; l f Represents the number of neuron nodes in the middle layer of the current neural network, π F represents the maximum m of neurons in the middle layer f The average of the activation outputs;

[0084] In π F <Φ f In the case of adding a new intermediate layer neuron node, the formula of the Gaussian function of the newly added neuron node is as follows:

[0085]

[0086] in, Indicates the maximum m f The average of the Gaussian function means of the neurons that activate the output, o f ∈(0,1) represents an adjustable parameter, and the variance of the Gaussian function is σ f , Φ f is the preset threshold, the neural network activation output after the node is increased and the weight values ​​of the neural network The formula is as follows:

[0087]

[0088] in, Represents the connection weight of the newly added node, and the number of neuron nodes in the middle layer of the updated neural network is redefined as l f =l f +1.

[0089] Step 143: Determine the auxiliary control law corresponding to the mapping deviation based on the neural network after the node is updated.

[0090] Specifically, after obtaining the neural network after the nodes are updated, the auxiliary control law corresponding to the mapping deviation can be determined based on the neural network after the nodes are updated.

[0091] The auxiliary control law corresponding to the mapping deviation can be determined based on the following formula:

[0092]

[0093] in, represents the auxiliary control law at the current moment, represents the coordinate transformation of the mapping deviation, b2 represents the parameter in the nonlinear mapping, K2 represents the control gain, represents the weight value of the neural network, represents the activation function of the neural network, X + represents the augmented state quantity, yes The transposed matrix of .

[0094] Based on the above embodiment, step 143 may further include:

[0095] Based on the mapping deviation, the weight value of the neural network is updated, and the updated weight value is shown in the following formula:

[0096]

[0097] Among them, α f ,α B Represents the learning rate preset by the neural network, Indicates state-related parameters, The coordinate transformation representing the mapping deviation, yes The transposed matrix, η f >0,η B >0,k f >0,k B >0 indicates learning parameters.

[0098] Based on the above embodiment, the nominal model predictive control strategy can be specifically as follows:

[0099]

[0100] in, represents the state of the nominal model, represents the state-space form of the nominal model, represents the coefficient matrix, 0 n×n ,I n represent the zero matrix and identity matrix of order n, respectively. represents the augmented control quantity, Ω τ represents the nominal input constraint, L(s,τ mpc ,l|t k )=s T s represents the stage cost, Ψ(s)=s T Ps represents the terminal cost, P represents a positive definite symmetric matrix, s represents the sliding mode defined according to the nominal error, Represents a virtual state quantity.

[0101] In the method provided by the embodiment of the present invention, the nominal control strategy introduces a nominal model of a sliding control mode, which can ensure the control performance of the system under the condition of relaxing the constraints on the terminal state.

[0102] Based on the above embodiment, the nominal model can be specifically:

[0103]

[0104] in, Respectively represent t k The nominal joint angle, angular velocity, and angular acceleration at the moment, It is a designable matrix. τ mpc Represents the control quantity of the nominal model, and the initial value of the nominal model is selected as

[0105] Based on any of the above embodiments, Figure 2 This is the second flow chart of the control method of the robot system provided by the present invention, such as Figure 2 As shown in Figure 2, the dynamic model of the manipulator of the robot system is as follows:

[0106]

[0107] in, Respectively represent the actual position, actual velocity and actual acceleration in the actual state at the current moment, represents the symmetric positive definite inertia matrix, and Represent the Coriolis force and gravity term in the joint space, Indicates the input control torque.

[0108] The joint positions and joint velocities of the manipulator in the robotic system must satisfy the following constraints:

[0109]

[0110] Where Γ represents the input constraint threshold, t represents time, i represents the input dimension, and n is a natural number.

[0111] The control goal is to build a suitable controller to enable the robot to quickly and efficiently move to the desired trajectory q while satisfying the above constraints. d (t) = [q d1 (t),q d2 (t),...,q dn (t)] T Tracking. dn (t) represents the component of the nth dimension of the desired trajectory.

[0112] The nominal model can be specified as:

[0113]

[0114] in,

[0115] Respectively represent t k The nominal joint angle, angular velocity, and angular acceleration at the moment, It is a designable matrix. τ mpc Represents the control quantity of the nominal model, and the initial value of the nominal model is selected as

[0116] According to the nominal model, the trajectory tracking error of the robot system can be decomposed into two parts:

[0117]

[0118] in, represents the nominal tracking error, It is clear that the nominal tracking error With nominal system is related, and the estimated deviation x e With the actual system and nominal system Therefore, considering the Tube-based model predictive control framework, the nominal model predictive controller and the auxiliary controller are designed to make x e →0, thus achieving the aforementioned control goal.

[0119] The auxiliary controller may be determined based on the following formula:

[0120]

[0121] in,

[0122] Is an ideal auxiliary controller,

[0123] represents the symmetric positive definite inertia matrix, Respectively represent the actual position and actual speed at the current moment, It is a designable matrix, defined as Represent the current time t k Nominal joint positions and nominal joint velocities.

[0124] make represents the nominal control law, then for t∈[t k ,t k+1 ), the ideal actual control law of the manipulator can be constructed as:

[0125]

[0126] Then, define the basic nominal control strategy. Let Indicates the nominal state, then can be reformulated as a state-space model:

[0127]

[0128] in, represents the coefficient matrix, 0 n×n ,I n represent the zero matrix and identity matrix of order n, respectively. In this embodiment, the Tube invariant set is defined as for where γ 1i (t),γ 2i (t) is a bounded positive time-varying function that satisfies γ 1i (t)<Γ 1i ,γ 2i (t)<Γ 2i Based on this, the nominal state should satisfy the constraint

[0129] Define the time series {t k} is the solution sequence of model predictive control, where t0 = 0. At the solution time t k , the basic nominal control strategy can be expressed as:

[0130]

[0131] in represents the stage cost, represents the terminal cost, Ω p represents the terminal constraint, and T represents the prediction time domain.

[0132] To relax the terminal state Constraints, eliminating the offline design steps, first solve According to the calculation results, the state constraints Convert to input constraints. It is easy to get the differential equation The solution is:

[0133]

[0134] Assuming the nominal input τ mpc satisfy According to the equivalent standard form calculation method and Rolle's middle value theorem, the nominal state satisfies:

[0135]

[0136] Where |·| means that each element of the vector (·) takes the absolute value.

[0137] in It can be seen that if the nominal input satisfies τ mpc ∈Ω τ ,and Then the tight state constraint can be satisfied

[0138] The next step is to consider introducing the sliding mode into the objective function of the model predictive control, using the sliding surface to constrain the nominal state, thereby relaxing the constraints on the terminal state. First, define the virtual dynamic system as follows:

[0139]

[0140] in, Represents a virtual state quantity, satisfying is a positive constant, Δτ=[Δτ1,Δτ2,...,Δτ n ] T represents a virtual dynamic input, where

[0141]

[0142] Combined with the above virtual state quantity and the aforementioned nominal tracking error The sliding mode is defined as:

[0143]

[0144] Where c1 represents a positive constant, represents the augmented nominal error.

[0145] Based on the formula and the input constraint τ mpc ∈Ω τ , the basic nominal control strategy can be reconstructed as:

[0146]

[0147] Among them, L(s,τ mpc ,l|t k )=s T s represents the stage cost, Ψ(s)=s T Ps represents the terminal cost, and P represents a positive definite symmetric matrix.

[0148] For the convenience of expression, in the embodiment of the present invention, for t∈[t k ,t k +T), a set of feasible solutions of the above reconstructed basic nominal control strategy is denoted as (·)^(t|t k ), which is k The optimal solution obtained at the moment (including the optimal nominal input and the corresponding nominal state) is recorded as (·) * (t|t k ).

[0149] It is worth noting that compared with the base nominal control strategy, the reconstructed strategy:

[0150] (1) Introducing the input constraint τ mpc ∈Ω τ , by limiting the nominal input to meet the constraints, so that the nominal state (including feasible solutions and optimal solutions) Satisfy constraints

[0151] (2) Based on sliding mode The objective function of model predictive control is constructed, and the optimization goal is to set s to 0. When s converges to 0, the nominal tracking error can converge along the sliding surface (i.e., the sliding surface is used to constrain the nominal state), relaxing the constraints on the terminal state.

[0152] (3)Introduction of virtual state quantity Used to build a viable terminal controller.

[0153] The corresponding terminal controller can be constructed as follows. First, the conditions that the terminal controller must meet are given as follows:

[0154] against like Represents its terminal controller, then for any l∈[t k +T,t k+1 +T), in the controller Under this effect, the following conditions are met:

[0155]

[0156] To construct a The terminal controller introduces a nonlinear saturation function S(u)=[S(u1),S(u2),...,S(u n )] T ,

[0157] Therefore, the terminal controller can be designed as:

[0158]

[0159] in, k s1 is a positive constant that satisfies 2k s1 λ P -1>0. Correspondingly, the aforementioned virtual dynamic system The input can be reconstructed as

[0160] When l∈[t k +T,t k+1 +T) Substitute into In the calculation, we get:

[0161]

[0162] From this we can get the form The terminal controller is designed to meet the above conditions.

[0163] The above content is related to the design of the nominal controller. The design of the auxiliary controller is carried out below.

[0164] To implement Tube invariant set constraints In order to achieve the goal of satisfying the requirements in the entire control cycle, an inverse hyperbolic tangent type nonlinear transformation is introduced in the embodiment of the present invention to transform the constrained deviation state x e Mapped to the new state quantity χ. Definition Then the inverse hyperbolic tangent type nonlinear transformation can be expressed as follows:

[0165]

[0166] The inverse transformation of the above transformation can be expressed as:

[0167] x ek,i (t) = γ ki (t)tanh(χ ki (t))k=1,2; i=1,2,...,n (2)

[0168] From the selection of the initial value of the nominal state, it can be seen that the initial value of the estimated deviation satisfies Therefore, according to formulas (1) and (2) and the properties of hyperbolic and inverse hyperbolic functions, the following lemma holds.

[0169] Lemma 1: The transformed state χ ki tends to infinity if and only if the estimated deviation satisfies |x ek,i (t)|→γ ki (t). Therefore, if χ ki is bounded, then the estimated deviation x in the entire motion cycle e All constraints can be satisfied

[0170] Lemma 2: If t ≥ T χ When the transformed state χ converges to a compact set in Then when t≥T χ When the estimated deviation x e Converges to a compact set in

[0171] The above lemma shows that the estimated deviation x between the nominal state and the actual state is e The goal of always staying within the Tube invariant set and converging to a neighborhood of the origin can be reduced to making the transformed state χ consistently and ultimately bounded.

[0172] To design a suitable auxiliary controller, the dynamic model of χ is first calculated as follows:

[0173]

[0174] in, For the sake of convenience, the time variable (t) is omitted in the above formula. Furthermore, (3) can be rewritten in matrix form:

[0175]

[0176] Among them, χ k =[χ k1 ,χ k2 ,...,χ kn ] T , a k=diag(a k1 ,a k2 ,...,a kn ), b k =[b k1 ,b k2 ,...,b kn ] T . Easy to get When a1 and a2 are positive definite symmetric matrices, the formula and Substituting into formula (4), we can get:

[0177]

[0178] in, Next, a feasible auxiliary controller is constructed through backstepping.

[0179] First, define the coordinate transformation as follows:

[0180]

[0181] Where α1 represents the virtual control amount, which will be given later.

[0182] Consider the transition state

[0183] Define the Lyapunov candidate function as Its derivative can be calculated as follows:

[0184]

[0185] when When , different from formula (2), we can define x e2 The pseudo-inverse transformation between x and χ2 is x e2,i =β 2i χ 2i ,in β 2i represents a state-dependent time-varying parameter whose boundedness is explained by L’Hôpital’s rule. e2,i ≠0, it is easy to get β 2i is a positive bounded quantity; when x e2,i →0 o'clock, This can be achieved When β 2i is a positive bounded quantity. The above pseudo-inverse transformation can be rewritten in matrix form:

[0186] x e2 =β2χ2 (8)

[0187] Where β2=diag(β 21 ,β 22,...,β 2n ). When x e Located in a tight set When , β2 is a positive definite symmetric matrix.

[0188] Substituting formula (6) (8) into formula (7), we can get:

[0189]

[0190] in, when hour, is a positive symmetric matrix. The calculation of formula (9) uses Young's inequality. According to formula (9), the virtual control quantity α1 can be designed as:

[0191] α1=-K1z χ1 -z χ1 ||b1|| 2 (10)

[0192] Here, K1>0 is a positive constant.

[0193] Furthermore, consider the transition state z χ2 .

[0194] Define the Lyapunov candidate function as Its derivative can be calculated as:

[0195]

[0196] in a2 and M -1 The definition shows that when hour, is a positive definite symmetric matrix. According to Young's inequality, the first two terms of formula (11) can be expressed as:

[0197]

[0198] Substituting (12) into (11), we can obtain:

[0199]

[0200] in, It is easy to get the virtual control quantity α1 is the variable q, γ1 and Function. It is easy to get is the variable q, γ1, and If the augmented state variable is defined as Then the function It can be expressed as

[0201] Reconstructing uncertain nonlinear functions by node-adaptive neural networks and M(q);

[0202] and M(q) can be reconstructed by the neural network as follows;

[0203]

[0204] in, represents the ideal network weight. Represents the neural network activation function. In the embodiment of the present invention, Gaussian function is selected as the activation function. ξ () Represents the estimation error of the neural network. Obviously, the above variables are bounded and satisfy where w (·)0 >0, ξ(·)>0.

[0205] In general, the ideal network weights is unknown, so we can and the approximate values ​​of M(q) are defined as and in, and Represents the estimated weight. Different from the traditional radial basis neural network with fixed intermediate layer neuron nodes, the embodiment of the present invention adopts a radial basis neural network with adaptive node adjustment to According to the activation output of the neural network, the neuron nodes in the middle layer are adaptively increased through the width learning method, and the corresponding Gaussian function mean is also adjusted accordingly.

[0206] For neural networks The formula for determining the activation rate of a neural network is as follows:

[0207]

[0208] in, is a Gaussian function, representing the activation output of the jth neuron node, X + represents the augmented state quantity, which is determined based on the mapping deviation; l f Represents the number of neuron nodes in the middle layer of the current neural network, π F represents the maximum m of neurons in the middle layer f The average of the activation outputs;

[0209] In π F <Φ fIn the case of adding a new intermediate layer neuron node, the formula of the Gaussian function of the newly added neuron node is as follows:

[0210]

[0211] in, Indicates the maximum m f The average of the Gaussian function means of the neurons that activate the output, o f ∈(0,1) represents an adjustable parameter, and the variance of the Gaussian function is σ f , Φ f is the preset threshold, the neural network activation output after the node is increased and the weight values ​​of the neural network The formula is as follows:

[0212]

[0213] in, Represents the connection weight of the newly added node, and the number of neuron nodes in the middle layer of the updated neural network is redefined as l f =l f +1.

[0214] Neural Networks The intermediate layer neuron nodes have the same adaptive adjustment strategy and will not be described in detail.

[0215] The auxiliary control law corresponding to the mapping deviation is determined based on the following formula:

[0216]

[0217] in, represents the auxiliary control law at the current moment, z χ2 represents the coordinate transformation of the mapping deviation, b2 represents the parameter in the nonlinear mapping, K2 represents the control gain, represents the weight value of the neural network, represents the activation function of the neural network, X + represents the augmented state quantity, yes The transposed matrix of .

[0218] Will Substitute into Chinese alternative and use Approximating M(q), the actual control law of the robot system can be obtained as:

[0219]

[0220] The method provided by an embodiment of the present invention can effectively solve the trajectory tracking control problem of a model-uncertain robot with state constraints. It achieves high-performance trajectory tracking of the nominal model by introducing a nominal model predictive control strategy with a sliding control mode. By introducing an auxiliary controller based on a node-adaptive neural network, a smaller-scale neural network is used to dynamically compensate for the robot's nonlinear uncertain system model. The estimated deviation between the nominal and actual states is constrained within the Tube invariant set, thereby achieving stable and efficient trajectory tracking for the robot.

[0221] In addition, regarding the judgment of the stability of the robot system, the embodiment of the present invention constructs a Lyapunov function about the nominal model predictive control objective function, the estimated deviation between the nominal state and the actual state, and the neural network weight error. According to the Lyapunov stability theorem, the parameter conditions that meet the stability of the closed-loop system are obtained, thereby ensuring the stability of the closed-loop system and ensuring that all variables in the closed-loop system are consistently and ultimately bounded.

[0222] First, we consider the convergence of the nominal tracking error, that is, the convergence of the tracking error of the nominal model to the desired trajectory. Construct the nominal control strategy as As shown, the corresponding terminal controller is designed as If the optimization problem in model predictive control is feasible at the initial time and meets the conditions 2k s1 λ P -1>0, Then the optimization problem in the above model predictive control is cyclically feasible. Under the control law obtained by the control strategy, the nominal tracking error Satisfies uniformly eventually bounded.

[0223] The following is a proof of the above description:

[0224] First, we explain the cyclic feasibility of the optimization problem in model predictive control. Without loss of generality, we assume that the optimization problem In t k There is a solution at all times. When l∈[t,t+T),t∈(t k ,t k+1 ], the control input can be constructed as follows:

[0225]

[0226] It is easy to get formula (13) to meet the input constraints Therefore, for t∈(t k ,t k+1 ], the optimization problem has a feasible solution. Due to the arbitrariness of the value of k, it can be seen that the above optimization problem is cyclically feasible.

[0227] The next step is to consider the stability of the nominal model and first analyze the convergence of the sliding mode s. To this end, the Lyapunov function is defined as In particular, when t=t k hour,

[0228] 1) When t∈(t k ,t k+1 ), from time t to time t k Moment V s The difference of (t) can be calculated as follows:

[0229]

[0230] in,

[0231] According to the definition of the nominal model and the virtual dynamic system, the initial value of the optimization problem at time t satisfies s(t|t)=s(t|t k ). From this we can get that when l∈[t,t k +T), s(l|t)=s(l|t k ), namely ΔV s1 =0.

[0232] For ΔV s2 , since the terminal controller Satisfy the formula and Therefore, the initial value s(t k +T|t)=s(t k +T|t k ), the formula From t k Integrate from t+T to t+T to get ΔV s2 ≤0.

[0233] For ΔV s3 , which obviously satisfies ΔV s3 <0. Therefore, V s (t)-V s (t k )≤0.

[0234] 2) When t = t k+1 When the objective function of model predictive control satisfies We can get:

[0235]

[0236] In summary, we can get the control strategy Under the control law obtained, the sliding mode s converges gradually. n×1 as well as Substitute into In the equation, it is easy to get the augmented nominal error It converges gradually and converges to the origin along the sliding surface s.

[0237] Next step is to analyze the virtual state The boundedness of . Define the Lyapunov function as Its derivative can be calculated as follows:

[0238]

[0239] Among them, the parameters satisfy

[0240] From the definition of Δτ, we can see that for any t∈[t k ,t k+1 ), when l∈[t,t k +T), Δτ=0 n×1 ; When l∈[t k +T,t+T), Δτ is bounded (due to the constraints Available is bounded, and thus Δτ is bounded). In summary, at any time, Δτ satisfies Where Δτ0>0. Therefore, according to Knowable virtual state Stay consistent and eventually bounded.

[0241] From the above analysis, we can see that the augmented error state Gradual convergence, while the virtual state The consistency is eventually bounded. Since the nominal tracking error satisfies:

[0242]

[0243] Nominal tracking error is easy to obtain Also remains consistently eventually bounded.

[0244] In practical applications, according to The obtained control law The input constraint is always satisfied, that is, the virtual dynamic system input Δτ is always 0. Therefore, in the control process, the virtual state quantity always satisfies According to the formula It can be seen that the nominal tracking error Maintaining and augmenting nominal error Same convergence.

[0245] Next, we prove that the estimated deviation x eThe estimation errors of the transformed states χ1, χ2 and the neural network weights The consistency final boundedness can be maintained. For the formula The described robot system, the nominal model is defined as The estimation deviation between the nominal state and the actual state is carried out by the formula Nonlinear transformation. Based on the mapping deviation, the neural network is updated in weight value, and the weight update law is shown in the following formula:

[0246]

[0247] If the condition And Under the action of The estimation errors of the transformed states χ1, χ2 and the neural network weights Maintain the consistency of the final boundedness.

[0248] The following expands the proof:

[0249] Define the Lyapunov function as:

[0250]

[0251] Wherein, The derivative of the Lyapunov function is obtained:

[0252]

[0253] Substitute Into the first term of , we can get:

[0254]

[0255] Wherein, In the calculation of formula (14), Young's inequality is considered, and the normalized signal Satisfies

[0256] Similarly, the second term of formula Can be simplified as:

[0257]

[0258] Wherein

[0259] Substitute α1=-K1z χ1 -z χ1 ||b1|| 2 Into , then the third term of formula Can be simplified as:

[0260]

[0261] in,

[0262] will Substitute into In the formula The fourth item is:

[0263]

[0264] in because Bounded, we can get ξ and θ B Bounded, satisfying ‖ξ‖≤ξ0, ‖θ B ‖≤θ B0 , ξ0 and θ B0 is a positive constant.

[0265] Substituting formulas (14)-(17) into In the above equation, we can get:

[0266]

[0267] in

[0268] The above parameters satisfy κ χ1 >0, κ f1 >0 and κ B1 > 0. From this we can conclude that the state quantity and the estimated error of the neural network weights Stay consistent and eventually bounded.

[0269] Next, we will explain the boundedness of the transition states χ1 and χ2. Obviously, χ1 satisfies the uniform final boundedness (χ1 = z χ1 ). According to Lemma 1, we can know that the estimation deviation x e1 Meet the conditions In addition, considering γ1 and are all bounded variables, it is easy to get that the virtual control quantity α1 is bounded. Then according to It can be seen that χ2 also satisfies the uniformly eventually bounded condition.

[0270] Based on the above analysis, combined with Lemma 1 and Lemma 2, we can conclude that The estimated deviation x between the nominal state and the actual state is e Can be kept in Tube unchanged Furthermore, the actual tracking error of the robot system can be obtained as It also satisfies the uniformly eventually bounded condition.

[0271] In addition, to verify the effectiveness of the designed Tube and sliding mode-based adaptive robot model predictive control method, a trajectory tracking simulation experiment of the UR5 robot can be performed based on the CoppeliaSim (V-REP) software. The computer configuration of the simulation environment is a @3.20GHz Intel i5-6500 CPU, 4GB RAM and Windows 10 operating system. The control program is written in MATLAB. For the sake of simplicity, this embodiment of the present invention only considers the tracking control problem of joints 4 to 6, and joints 1 to 3 are locked. In this embodiment of the present invention, joints 4 to 6 are represented as q1, q2 and q3 respectively. Define q = [q1, q2, q3] T , The robot state can be expressed as

[0272] Define the UR5 robot state constraints as The expected trajectory of the moving joint is set as The initial values ​​of joint angle and angular velocity are set to and The overall movement time is set to T s =20s.

[0273] For the UR5 robot simulation model defined above, the key parameters of the adaptive model predictive control method based on Tube and sliding mode are set as follows. For the nominal control strategy: the coefficient matrix of its nominal prediction system is set to The nominal input constraint is set to The sliding mode parameter in the objective function is set to c = 5, the parameter P in the terminal cost is set to P = diag (2, 2, 2), and the virtual dynamic system parameter is set to The prediction time domain of model predictive control can be selected as T = 0.05s, and the time interval for solving the optimization problem can be selected as Δ t =0.01s.

[0274] For auxiliary controllers: Tube invariant set The time-varying function in is set to γ 1i =γ 2i =e (-0.4t) +0.2, i=1,2,3. Virtual control quantity α1 and auxiliary controller The control gains in can be set to K1=5, K2=diag(19,10,8) respectively. 1i From the definition of , we can see that and can be expressed linearly by γ1, so the neural network The input can be simplified to The activation function center value can be set to [-1,1]×[-1,1]×[-1,1]×[-1]×[1]×[0]×[0]×[0]×[0]×[0]×[0]×[0]×[0], and the corresponding initial number of middle layer nodes is set to 8. Neural Network The initial number of nodes in the middle layer is set to 2, the activation function center value is set to [-1.6, 1.6] × [0] × [0], and the variance of the neural network activation function is set to 25. The parameters related to the update of the middle layer nodes of the neural network can be: m f0 =6,Φ f =0.8,o f =0.4;m B0 =3,Φ B =0.7,o B =0.2.

[0275] In addition, to further illustrate the effectiveness of the adaptive model predictive control method based on Tube and sliding control modes in robot trajectory tracking control, the control method of the robot system is compared with the PD control method, the adaptive neural network control method and the constraint model predictive control method.

[0276] In the constrained model predictive control method, the overall control framework is the same as that of the robot system, but the nominal control strategy is constructed using the general constrained model predictive control method.

[0277] In the PD controller design, the tracking error is first constructed based on the backstepping method. Define z1 = q d -q, The auxiliary control quantity The PD controller is designed to be τ PD =K p The auxiliary control quantity and the control gain in the PD controller are designed as K1 = 5 and K2 = diag(18, 10, 5) respectively.

[0278] In the adaptive neural network control method, the tracking error construction method is the same as the PD control method, and the controller is designed as The control gain K1=5, K p =diag(18,10,8); is a neural network used to estimate Item. Activation function The Gaussian function is selected and its input is set to The number of nodes in the middle layer of the neural network is set to 64, its variance is set to 25, and the activation function center value is set to [-1,1]×[-1,1]×[-1,1]×[0]×[0]×[0]×[-1,1]×[-1,1]×[0]×[0]×[0]. Neural network weights The adaptive update law design and the neural network in this invention The weight update law is the same.

[0279] In the constrained model predictive control method, the overall control framework is the same as the control method of the robot system. The nominal model predictive controller is constructed using the constrained model predictive control method, and its objective function is set as The stage cost design is Where Q = diag(200,100,100,10,10,10), R = diag(0.01,0.01,0.01); the terminal cost design is where Q N =diag(2000,1000,1000,10,10,10).

[0280] The control device of the robot system provided by the present invention is described below. The control device of the robot system described below and the control method of the robot system described above can be referenced to each other.

[0281] Based on any of the above embodiments, an embodiment of the present invention provides a control device for a robot system. Figure 3 Schematic diagram of the structure of the control device of the robot system provided by the present invention, such as Figure 3 As shown, the device includes:

[0282] An acquisition unit 310 is configured to acquire an actual state at a current moment and an expected state within a preset period after the current moment;

[0283] a calculation unit 320 for calculating a nominal control law and a nominal state corresponding to the desired state within the preset time period based on a nominal model predictive control strategy, wherein the nominal model predictive control strategy is constructed by introducing a sliding mode based on a nominal model of the robot system;

[0284] an estimation deviation determining unit 330, configured to determine an estimation deviation based on the actual state and the nominal state;

[0285] An auxiliary control law determination unit 340 is configured to determine an auxiliary control law corresponding to the estimated deviation based on an auxiliary controller, wherein the auxiliary controller is a neural network that adaptively adjusts neuron nodes and / or weight values ​​based on the estimated deviation;

[0286] The control unit 350 is configured to determine an actual control law in the preset time period based on the auxiliary control law and the nominal control law, and control the robot system based on the actual control law.

[0287] The device provided by the embodiment of the application determines an actual control law in a preset time period based on an auxiliary control law and a nominal control law, and controls a robot system based on the actual control law. The auxiliary controller is a neural network that adaptively adjusts the neuron nodes and / or weight values based on the estimated deviation. Therefore, the auxiliary control law output by the auxiliary controller can dynamically compensate for non-limited uncertain items of the robot system, thereby improving the accuracy of subsequent robot system control. In addition, the nominal model predictive control strategy is constructed based on the sliding mode of the nominal model of the robot system, and can guarantee the control performance of the system under the condition of relaxing the constraint condition of the terminal state.

[0288] Based on any one of the above embodiments, the auxiliary control law determination unit is specifically configured to:

[0289] The mapping unit is configured to perform nonlinear mapping on the estimated deviation to obtain a mapping deviation.

[0290] The node updating unit is configured to determine an activation rate based on the mapping deviation, and update the neuron nodes of the neural network based on the activation rate to obtain an updated neural network.

[0291] The auxiliary control law determination sub-unit is configured to determine the auxiliary control law corresponding to the mapping deviation based on the updated neural network.

[0292] Based on any one of the above embodiments, the node updating unit is specifically configured to:

[0293] The formula for determining the activation rate of the neural network is as follows:

[0294]

[0295] wherein, is a Gaussian function, represents the activation output of the jth neuron node, X + represents an augmented state quantity, which is determined based on the mapping deviation; l f represents the number of intermediate layer neuron nodes in the current neural network, π F represents the maximum m f activation output of the intermediate layer neuron;

[0296] In the case of π F <Φ f , a new intermediate layer neuron node is added, and the formula of the Gaussian function of the newly added neuron node is as follows:

[0297]

[0298] in, Indicates the maximum m f The average of the Gaussian function means of the neurons that activate the output, o f ∈(0,1) represents an adjustable parameter, and the variance of the Gaussian function is σ f , Φ f is the preset threshold, the neural network activation output after the node is increased and the weight values ​​of the neural network The formula is as follows:

[0299]

[0300] in, Represents the connection weight of the newly added node, and the number of neuron nodes in the middle layer of the updated neural network is redefined as l f =l f +1.

[0301] Based on any of the above embodiments, determining the auxiliary control law subunit includes:

[0302] The auxiliary control law corresponding to the mapping deviation is determined based on the following formula:

[0303]

[0304] in, represents the auxiliary control law at the current moment, z χ2 represents the coordinate transformation of the mapping deviation, b2 represents the parameter in the nonlinear mapping, K2 represents the control gain, represents the weight value of the neural network, represents the activation function of the neural network, X + represents the augmented state quantity, yes The transposed matrix of .

[0305] Based on any of the above embodiments, after determining the auxiliary control law subunit, the method further includes:

[0306] Based on the mapping deviation, the weight value of the neural network is updated, and the updated weight value is shown in the following formula:

[0307]

[0308] Among them, α f ,α B Represents the learning rate preset by the neural network, Represents state-related parameters, z χ2The coordinate transformation representing the mapping deviation, It is z χ2 The transposed matrix, η f >0,η B >0,k f >0,k B >0 indicates learning parameters.

[0309] Based on any of the above embodiments, the nominal model predictive control strategy is specifically:

[0310]

[0311] in, represents the state of the nominal model, represents the state-space form of the nominal model, represents the coefficient matrix, 0 n×n ,I n represent the zero matrix and identity matrix of order n, respectively. represents the augmented control quantity, Ω τ represents the nominal input constraint, L(s,τ mpc ,l|t k )=s T s represents the stage cost, Ψ(s)=s T Ps represents the terminal cost, P represents a positive definite symmetric matrix, s represents the sliding mode defined according to the nominal error, Represents a virtual state quantity.

[0312] Based on any of the above embodiments, the nominal model is specifically:

[0313]

[0314] in, Respectively represent t k The nominal joint angle, angular velocity, and angular acceleration at the moment, It is a designable matrix. τ mpc Represents the control quantity of the nominal model, and the initial value of the nominal model is selected as

[0315] Figure 4 An example of a physical structure diagram of an electronic device is shown below. Figure 4As shown, the electronic device may include: a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other via the communication bus 440. The processor 410 may call logic instructions in the memory 430 to execute a control method for a robot system, the method comprising: obtaining an actual state at a current moment and an expected state within a preset period after the current moment; calculating a nominal control law and a nominal state corresponding to the expected state within the preset period based on a nominal model predictive control strategy, wherein the nominal model predictive control strategy is constructed by introducing a sliding mode based on a nominal model of the robot system; determining an estimated deviation based on the actual state and the nominal state; determining an auxiliary control law corresponding to the estimated deviation based on an auxiliary controller, wherein the auxiliary controller is a neural network that adaptively adjusts neuron nodes and / or weight values ​​based on the estimated deviation; determining an actual control law within the preset period based on the auxiliary control law and the nominal control law, and controlling the robot system based on the actual control law.

[0316] In addition, the logic instructions in the above-mentioned memory 430 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0317] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the control method of the robot system provided by the above methods, which includes: obtaining the actual state at the current moment and the expected state within a preset time period after the current moment; based on the nominal model predictive control strategy, calculating the nominal control law and nominal state corresponding to the expected state within the preset time period, and the nominal model predictive control strategy is constructed by introducing a sliding mode based on the nominal model of the robot system; based on the actual state and the nominal state, determining the estimated deviation; based on the auxiliary controller, determining the auxiliary control law corresponding to the estimated deviation, and the auxiliary controller is a neural network that adaptively adjusts neuron nodes and / or weight values ​​based on the estimated deviation; based on the auxiliary control law and the nominal control law, determining the actual control law within the preset time period, and controlling the robot system based on the actual control law.

[0318] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the control method of the robot system provided by the above-mentioned methods, the method comprising: obtaining the actual state at the current moment and the expected state within a preset time period after the current moment; calculating the nominal control law and nominal state corresponding to the expected state within the preset time period based on the nominal model predictive control strategy, the nominal model predictive control strategy is constructed by introducing a sliding mode based on the nominal model of the robot system; determining an estimated deviation based on the actual state and the nominal state; determining an auxiliary control law corresponding to the estimated deviation based on an auxiliary controller, the auxiliary controller being a neural network that adaptively adjusts neuron nodes and / or weight values ​​based on the estimated deviation; determining the actual control law within the preset time period based on the auxiliary control law and the nominal control law, and controlling the robot system based on the actual control law.

[0319] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

Claims

1. A control method for a robot system, characterized in that: include: Obtaining the actual state at the current moment and the expected state within a preset period after the current moment; Calculating a nominal control law and a nominal state corresponding to the desired state within the preset time period based on a nominal model predictive control strategy, wherein the nominal model predictive control strategy is constructed by introducing a sliding mode based on a nominal model of the robot system; determining an estimated deviation based on the actual state and the nominal state; determining an auxiliary control law corresponding to the estimated deviation based on an auxiliary controller, wherein the auxiliary controller is a neural network that adaptively adjusts neuron nodes and / or weight values ​​based on the estimated deviation; determining an actual control law within the preset time period based on the auxiliary control law and the nominal control law, and controlling the robot system based on the actual control law; The determining, based on the auxiliary controller, an auxiliary control law corresponding to the estimated deviation includes: Performing nonlinear mapping on the estimated deviation to obtain a mapping deviation; Determining an activation rate based on the mapping deviation, and updating neuron nodes of the neural network based on the activation rate to obtain a neural network after node update; Based on the neural network after the node update, an auxiliary control law corresponding to the mapping deviation is determined.

2. The control method of the robot system according to claim 1, characterized in that: The updating of neuron nodes of the neural network based on the activation rate includes: The formula for determining the activation rate of the neural network is as follows: in, is a Gaussian function, representing the activation output of the jth neuron node, X + represents the augmented state quantity, which is determined based on the mapping deviation; l f Represents the number of neuron nodes in the middle layer of the current neural network, π F represents the maximum m of neurons in the middle layer f The average of the activation outputs; In π F <Φ f In the case of adding a new intermediate layer neuron node, the formula of the Gaussian function of the newly added neuron node is as follows: in, Indicates the maximum m f The average of the Gaussian function means of the neurons that activate the output, o f ∈(0,1) represents an adjustable parameter, and the variance of the Gaussian function is σ f , Φ f is the preset threshold, the neural network activation output after the node is increased and the weight values ​​of the neural network The formula is as follows: in, Represents the connection weight of the newly added node, and the number of neuron nodes in the middle layer of the updated neural network is redefined as l f =l f +1.

3. The control method of the robot system according to claim 1, characterized in that: The determining of the auxiliary control law corresponding to the mapping deviation based on the neural network after the node update includes: The auxiliary control law corresponding to the mapping deviation is determined based on the following formula: in, represents the auxiliary control law at the current moment, z χ2 represents the coordinate transformation of the mapping deviation, b2 represents the parameter in the nonlinear mapping, K2 represents the control gain, represents the weight value of the neural network, represents the activation function of the neural network, X + represents the augmented state quantity, yes The transposed matrix of .

4. The control method of the robot system according to claim 1, characterized in that: The step of determining the auxiliary control law corresponding to the mapping deviation based on the neural network after the node update further includes: Based on the mapping deviation, the weight value of the neural network is updated, and the updated weight value is shown in the following formula: Among them, α f ,α B Represents the learning rate preset by the neural network, Represents state-related parameters, z χ2 The coordinate transformation representing the mapping deviation, It is z χ2 The transposed matrix, η f >0,η B >0,k f >0,k B >0 indicates learning parameters.

5. The control method of the robot system according to claim 1, characterized in that: The nominal model predictive control strategy is specifically: in, represents the state of the nominal model, represents the state-space form of the nominal model, represents the coefficient matrix, 0 n×n ,I n represent the zero matrix and identity matrix of order n, respectively. represents the augmented control quantity, Ω τ represents the nominal input constraint, L(s,τ mpc ,l|t k )=s T s represents the stage cost, Ψ(s)=s T Ps represents the terminal cost, P represents a positive definite symmetric matrix, s represents the sliding mode defined according to the nominal error, and θ1, θ2 represent virtual state quantities.

6. The control method of a robot system according to any one of claims 1 to 5, characterized in that: The nominal model is specifically: in, Respectively represent t k The nominal joint angle, angular velocity, and angular acceleration at the moment, It is a designable matrix. τ mpc Represents the control quantity of the nominal model, and the initial value of the nominal model is selected as 7. A control device for a robot system, characterized in that: include: an acquisition unit, configured to acquire an actual state at a current moment and an expected state within a preset period after the current moment; a calculation unit, configured to calculate a nominal control law and a nominal state corresponding to the desired state within the preset time period based on a nominal model predictive control strategy, wherein the nominal model predictive control strategy is constructed by introducing a sliding mode based on a nominal model of the robot system; an estimation deviation determining unit, configured to determine an estimation deviation based on the actual state and the nominal state; an auxiliary control law determination unit, configured to determine an auxiliary control law corresponding to the estimated deviation based on an auxiliary controller, wherein the auxiliary controller is a neural network that adaptively adjusts neuron nodes and / or weight values ​​based on the estimated deviation; a control unit, configured to determine an actual control law within the preset time period based on the auxiliary control law and the nominal control law, and control the robot system based on the actual control law; The auxiliary control law determination unit is specifically used to: Performing nonlinear mapping on the estimated deviation to obtain a mapping deviation; Determining an activation rate based on the mapping deviation, and updating neuron nodes of the neural network based on the activation rate to obtain a neural network after node update; Based on the neural network after the node update, an auxiliary control law corresponding to the mapping deviation is determined.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the control method of the robot system according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the control method of the robot system according to any one of claims 1 to 6 is implemented.

Images