A method for identifying robot dynamic parameters

Abstract

The present invention discloses a method for identifying robot dynamic parameters, comprising the following steps: S1, collecting joint motion data of multiple groups of robots when executing excitation trajectories as training sets; S2, calculating the corresponding theoretical joint torque under the training set data according to the established dynamic model, calculating the mean square error between the theoretical joint torque and the collected measured joint torque, and defining it as a loss function; S3, in each round of training, calculating the gradient of the loss function with respect to the dynamic parameters based on the back propagation algorithm; S4, iteratively training the dynamic parameters along the negative gradient direction of the loss function, and as the training proceeds, the loss function gradually decreases, and when the loss function converges to a minimum value, saving the corresponding dynamic parameters at this time as the identification result, and completing parameter identification. The present invention improves the identification accuracy of parameters, and solves the problem that the least square method is susceptible to noise interference and can only identify linear system parameters.

Description

Technical Field

[0001] The invention belongs to the field of robot control and relates to a method for identifying robot dynamic parameters. Background Art

[0002] The dynamic model is a mathematical model that describes the relationship between the driving torque required for robot movement and the dynamic parameters and kinematic parameters. The kinematic parameters can be obtained through calibration, while the dynamic parameters are related to the internal friction of the joint and the coupling between adjacent joints, which are generally difficult to obtain directly. Control based on the dynamic model is an effective method to achieve high-speed and high-precision control of the robot. Therefore, it is of great significance to design a dynamic parameter identification method with sufficient accuracy and easy to identify.

[0003] The methods of dynamic parameter identification can generally be divided into three categories. Both the disassembly measurement method and the CAD measurement method ignore the dynamic characteristics of the robot during actual motion and the coupling between the joints. Therefore, the experimental identification method is the most widely used identification method. The experimental identification method substitutes the motion data of the robot under the excitation trajectory into the established dynamic model, and minimizes the objective function by executing a specifically designed identification algorithm, thereby realizing the identification of dynamic parameters. The steps of the experimental identification method mainly include establishing a model and determining identifiable dynamic parameters, designing and optimizing the excitation trajectory, collecting and processing data, and identifying parameters and model verification. The main difference between various experimental identification methods lies in the selection and design of identification algorithms.

[0004] The combination of inverse dynamics model and least squares method is widely used in parameter identification, because the inverse dynamics model is a mathematical model that can directly calculate the joint torque from the joint angle, velocity and acceleration, and there is a linear relationship between the model and the dynamic parameters. However, the least squares method needs to collect the signals of joint angle and torque, and the specific values ​​of joint velocity and acceleration need to be obtained through filters, which may introduce large noise, resulting in a biased solution. In order to improve this drawback, M. Gautier proposed a method of parameter identification using only joint torque. Under the same control law, the robot actually executes the same excitation trajectory as the simulation, and obtains the optimal dynamic parameters by minimizing the torque difference between the two. However, this method requires knowing the detailed control law of the robot and requires motion simulation experiments, and the process is relatively cumbersome. Yang Chengguang proposed an adaptive finite time parameter identification method, which integrates the error of parameter identification into the identification algorithm to ensure the convergence rate of the algorithm. KS Narendra proved that neural networks can be effectively used for the identification and control of nonlinear dynamic systems. ZhaoHui Jiang used neural network to dynamically compensate for the identification error of dynamic parameters, improving the control accuracy of the robot. Yazdizadeh A designed an adaptive time-delay neural network to achieve parameter identification for the two-joint flexible robotic arm system.

[0005] Therefore, the back propagation algorithm in the neural network has excellent nonlinear mapping ability and powerful self-learning characteristics, which provides a good idea for realizing high-precision and concise dynamic parameter identification. Therefore, the present invention proposes a method for identifying robot dynamic parameters based on the principle of back propagation learning algorithm in order to improve the identification accuracy of dynamic parameters. Summary of the invention

[0006] The purpose of the present invention is to provide a method for identifying the dynamic parameters of a robot based on a back-propagation learning algorithm, aiming to overcome the shortcomings of the prior art, solve the problem that parameter identification is easily interfered by noise, and achieve high-precision and concise dynamic parameter identification, so as to improve the robot's dynamic model and parameter identification for the robot's force control and human-computer interaction tasks.

[0007] The present invention is achieved by at least one of the following technical solutions.

[0008] A method for identifying robot dynamic parameters comprises the following steps:

[0009] S1, collect joint motion data of multiple groups of robots when executing the excitation trajectory, collect relevant data at each time sampling point, and use the second-order central differential method to calculate the joint acceleration, and combine the filtered and denoised joint motion data into a training set;

[0010] S2. Establish a dynamic model to calculate the corresponding theoretical joint torque under the training set data, calculate the mean square error between the theoretical joint torque and the collected measured joint torque, and define it as a loss function;

[0011] S3. In each round of training, the gradient of the loss function with respect to the dynamics parameters is calculated based on the back propagation algorithm;

[0012] S4. The kinetic parameters are iteratively trained along the negative gradient direction of the loss function. As the training progresses, the loss function gradually decreases. When the loss function converges to the minimum value, the corresponding kinetic parameters are saved as the identification result, thereby realizing the identification of the kinetic parameters.

[0013] Preferably, the excitation trajectory described in step S1 is a periodic Fourier series:

[0014]

[0015] Where i represents the robot joint, t represents the discrete time sequence, l = 1, 2, ..., n H represents the harmonic number, q(i,t) represents the joint angle of joint i at each discrete time point, and q i,0 represents the bias term, Respectively represent the amplitude terms of the sine function and the cosine function, w f Indicates the fundamental angular frequency.

[0016] Preferably, the kinetic model described in step S2 is:

[0017] τ=τ b +τ f

[0018]

[0019]

[0020] Among them, q, They represent the robot joint angle, joint angular velocity and joint acceleration respectively; τ represents the joint driving torque; τ b It represents the resultant moment of inertia, Coriolis moment, centripetal moment and gravity moment; represents the regression matrix, which is only related to the robot's state q, Related; β represents the minimum identifiable inertial parameter set, which is obtained by reorganizing the complete dynamic parameters of the robot; τ f represents the joint friction torque, which adopts the Stribeck friction model, f c 、f s 、V s ,ξ,f v 、f prepresents the Stribeck friction coefficient set, so the minimum inertia parameter set and the friction coefficient set together constitute the dynamic parameter set to be identified; ξ represents the Stribeck index parameter.

[0021] Preferably, the dynamic model described in step S2 is a generalized neural network, the input of the generalized neural network includes joint angles, joint angular velocities and joint accelerations, the output of the network includes theoretical joint torques calculated using the dynamic model, and the dynamic parameters are connection weights of the network, which can be trained and learned.

[0022] Preferably, the overall architecture of the generalized neural network includes:

[0023] Input layer X:

[0024]

[0025] Hidden Layer:

[0026]

[0027]

[0028] Output layer:

[0029]

[0030] Where i = 1, 2, ..., 6 represents the robot joints; β l represents the minimum inertial parameter set β=[β1,β2,…,β 40 ] the lth parameter in ; Represents the regression matrix The lth element in the i-th row of ; represents the joint velocity of the i-th joint; f c,i represents the Coulomb friction parameter, f s,i represents the static friction parameter, f v,i represents the viscous friction parameter, V s,i represents the Stribeck velocity, ξ i represents the Stribeck index parameter, f p,i represents the bias parameter; It represents the theoretical calculated value of the resultant moment of inertia, Coriolis moment, centripetal moment and gravity moment of joint i. represents the theoretical calculated value of the friction torque of joint i, and The sum represents the theoretical joint torque calculated based on the dynamic model;

[0031] The loss function of the generalized neural network is defined as the mean square error between the theoretical joint torque and the measured joint torque of the robot's six joints. The generalized neural network has the following properties:

[0032] (1) The connection weights of the network are the parameters to be identified or functions of the parameters to be identified;

[0033] (2) The default value of the neuron bias is 0;

[0034] (3) The activation function is determined by the functional relationship in the established dynamic model.

[0035] Preferably, the loss function described in step S2 is:

[0036]

[0037] Among them, τ i represents the measured joint torque of joint i collected, Represents the calculated theoretical joint torque of joint i.

[0038] Preferably, in step S3, in each round of training, the gradient of the loss function E with respect to the dynamics parameter is calculated as:

[0039] For the minimum inertia parameter set β l :

[0040]

[0041] where β l represents the minimum inertial parameter set β=[β1,β2,…,β 40 ] the lth parameter in ; Represents the regression matrix The lth element in the i-th row of ; τ i represents the measured joint torque of joint i collected, represents the calculated theoretical joint torque of joint i;

[0042] For the friction parameter set, δ is defined as the error term corresponding to the friction parameter set:

[0043]

[0044] For the Coulomb friction parameter f c,i :

[0045]

[0046] where ξ i represents the Stribeck index parameter, Represents the joint velocity of joint i:

[0047] For the static friction parameter f s,i :

[0048]

[0049] For the viscous friction parameter f v,i :

[0050]

[0051] For Stribeck speed V s,i :

[0052]

[0053] For the Stribeck index parameter ξ i :

[0054]

[0055] For the bias parameter f p,i :

[0056]

[0057] Preferably, in step S4, according to the gradient of the loss function with respect to the kinetic parameters, the kinetic parameters are iteratively updated in the direction of the negative gradient of the loss function based on the gradient descent method, and the loss function value gradually decreases; when the training is completed, the loss function value converges to the minimum value, and the corresponding kinetic parameter set at this time is saved as the identification result, thereby completing the parameter identification.

[0058] Preferably, the method for iterative training and updating of the dynamic parameters along the negative gradient direction of the loss function is the gradient descent method:

[0059] Θ j (k+1)=θ j (k)-ΔΘ j (k)

[0060]

[0061] Among them, k represents the number of training times, Θ t represents the set of dynamic parameters, η represents the learning rate, ΔΘ j (k) represents the amount of gradient descent during the kth training. Respectively represent the modified first-order and second-order gradient momentum, and their update method is:

[0062] m t (k) = α1·m t (k-1)+(1-α1)·g(k)

[0063]

[0064] V t (k) = α2·V t (k-1)+(1-α2)·g 2 (k)

[0065]

[0066] Where k represents the number of training times, α1 and α2 represent the parameters for controlling the decay of the first-order and second-order gradient momentum, g(k) represents the gradient of the loss function with respect to the set of dynamic parameters, and m t (k), V t (k) represent the first and second order gradient momentum before correction.

[0067] Preferably, the robot comprises a rotary articulated robot, and the rotary articulated robot comprises a serial articulated industrial robot.

[0068] Compared with the prior art, the present invention has the following advantages and effects:

[0069] (1) The present invention utilizes the powerful data mapping capability and self-learning function of the back propagation algorithm to train the dynamic parameters under multiple groups of excitation trajectories and automatically find the optimal value. This overcomes the limitations of traditional methods that are easily affected by noise and easily fall into local optimality under big data.

[0070] (2) The present invention trains the kinetic parameters based on a large amount of training set data based on the back propagation algorithm, so that the kinetic parameters learn the mapping relationship between the input and output of the kinetic model. The trained kinetic parameters can not only fit the modeled kinetic characteristics, but also compensate for the unmodeled kinetic factors. Therefore, the present invention can effectively improve the integrity of the model and the recognition accuracy of the parameters.

[0071] (3) The present invention is not only applicable to parameter identification of linear models, but also to parameter identification of nonlinear models, thus improving practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 is a schematic diagram of the robot DH in an embodiment of the present invention;

[0073] Figure 2 is a flow chart of a method for identifying robot dynamic parameters in an embodiment of the present invention;

[0074] Figure 3 It is a schematic diagram of a generalized neural network established in an embodiment of the present invention. DETAILED DESCRIPTION

[0075] The present invention will be further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0076] This example uses a six-axis serial articulated robot as the example object. Figure 1 shown.

[0077] like Figure 2 As shown, this example provides a method for identifying robot dynamic parameters, comprising the steps of:

[0078] S1, collecting joint motion data of multiple groups of robots when executing the excitation trajectory, collecting joint angle, joint velocity, and joint torque data at each time sampling point, and using the second-order central differential method to calculate the joint acceleration, and combining the filtered and denoised joint motion data into a training set; specifically, the excitation trajectory described in step S1 is a periodic Fourier series,

[0079]

[0080] i represents the robot joint, i represents the robot joint, t represents the discrete time sequence, l=1,2,…,n H represents the harmonic number, q(i,t) represents the joint angle of joint i at each discrete time point, and q i,0 represents the bias term, Respectively represent the amplitude terms of the sine function and the cosine function, and the harmonic number n H Defined as 6, the fundamental angular frequency w f Defined as 0.16*π.

[0081] S2. Calculate the corresponding theoretical joint torque under the training set data according to the established dynamic model, calculate the mean square error between the theoretical joint torque and the collected measured joint torque, and define it as the loss function; specifically, the established dynamic model is,

[0082] τ=τ b ++τ f

[0083]

[0084]

[0085] in They represent the robot joint angle, joint angular velocity and joint acceleration respectively; represents the joint torque; It represents the resultant moment of inertia, Coriolis moment, and gravity moment, which can be simplified to a linear equation of two matrices multiplied together; Represents the regression matrix, which is only related to the robot's state q, related; represents the minimum inertial parameter set, which is obtained by reorganizing the complete dynamic parameters of the robot; represents the joint friction torque, which adopts the Stribeck friction model. represents the Stribeck friction coefficient set. Therefore, the minimum inertia parameter set and the friction coefficient set together constitute the dynamic parameter set to be identified.

[0086] Specifically, the dynamic model described in step S2 can be analogized to a generalized neural network, because the essence of the dynamic model is the functional relationship between joint torque and joint angle, joint velocity and joint acceleration, so the input of the generalized neural network is analogized to joint angle, joint angular velocity and joint acceleration, and the output of the network is analogized to the theoretical joint torque calculated by the dynamic model. The dynamic parameters are analogous to the connection weights of the network and can be trained and learned.

[0087] The overall architecture of a generalized neural network is as follows Figure 3 As shown, including:

[0088] Input Layer:

[0089]

[0090] Hidden Layer:

[0091]

[0092]

[0093] Output layer:

[0094]

[0095] Where i = 1, 2, ..., 6 represents the number of joints; β l represents the minimum inertial parameter set β=[β1,β2,…,β 40 ] the lth parameter in ; Represents the regression matrix The lth element in the i-th row of ; represents the joint velocity of joint i; It represents the theoretical calculated value of the resultant moment of inertia, Coriolis moment, centripetal moment and gravity moment of joint i. represents the theoretical calculated value of the friction torque of joint i, and The sum represents the theoretical joint torque calculated based on the dynamic model.

[0096] Unlike traditional neural networks, the generalized neural network has the following properties:

[0097] (1) The connection weights of the network are the parameters to be identified or functions of the parameters to be identified.

[0098] (2) The default value of a neuron’s bias is 0.

[0099] (3) The activation function in this paper is determined by the functional relationship in the established dynamic model, and is not necessarily a function of the weighted sum of the inputs.

[0100] Specifically, the loss function of the network is defined as the mean square error between the theoretical joint torque and the measured joint torque of the six joints of the robot, so the loss function described in step S2 is:

[0101]

[0102] Among them, τ i represents the measured joint torque of joint i collected, Represents the calculated theoretical joint torque of joint i.

[0103] S3. In each round of training, the gradient of the loss function with respect to the dynamic parameters is calculated based on the back propagation algorithm; specifically, the gradient of the loss function with respect to the dynamic parameters is,

[0104] For the minimum inertia parameter set β l :

[0105]

[0106] where β l represents the minimum inertial parameter set β=[β1,β2,…,β 40 ] the lth parameter in ; Represents the regression matrix The lth element in the i-th row of ; τ i represents the measured joint torque of joint i collected, represents the calculated theoretical joint torque of joint i;

[0107] For the friction parameter set, define δ as the error term corresponding to the friction parameter set,

[0108]

[0109] For the Coulomb friction parameter f c,i :

[0110]

[0111] For the static friction parameter f s,i :

[0112]

[0113] For the viscous friction parameter f v,i :

[0114]

[0115] For Stribeck speed V s,i :

[0116]

[0117] For the Stribeck index parameter ξ i :

[0118]

[0119] For the bias parameter f p,i :

[0120]

[0121] S4, the kinetic parameters are iteratively trained along the negative gradient direction of the loss function. As the training progresses, the loss function gradually decreases. When the loss function converges to the minimum value, the corresponding kinetic parameters at this time are saved as the identification result, thereby realizing the identification of the kinetic parameters. Specifically, in step S4, according to the calculated gradient of the loss function with respect to the kinetic parameters, the kinetic parameters are iteratively updated in the negative gradient direction of the loss function based on the gradient descent method, and the loss function value gradually decreases; when the training is completed, the loss function value converges to the minimum value, at which time the kinetic parameter set implicitly learns the mapping relationship between the input and output of the kinetic model, and the corresponding kinetic parameter set at this time is saved as the identification result to complete the parameter identification.

[0122] Specifically, the method for updating the dynamic parameter set in step S4 along the negative gradient direction of the loss function is the gradient descent method:

[0123] Θ j (k+1)=θ j (k)-ΔΘ j (k)

[0124]

[0125] Among them, k represents the number of training times, Θ j represents the general term for dynamic parameters, η represents the learning rate, ΔΘ j (k) represents the amount of gradient descent during the kth training. Respectively represent the modified first-order and second-order gradient momentum, and the update method is,

[0126] m t (k) = α1·m t(k-1)+(1-α1)·g(k)

[0127]

[0128] V t (k) = α2·V t (k-1)+(1-α2)·g 2 (k)

[0129]

[0130] Where k represents the number of training times, α1 and α2 represent the parameters for controlling the decay of the first-order and second-order gradient momentum, g(k) represents the gradient of the loss function with respect to the set of dynamic parameters, and m t (k), V t (k) represent the first and second order gradient momentum before correction.

[0131] Specifically, the gradient of the loss function with respect to the dynamic parameters calculated in step S3 is substituted into the gradient descent method in step S4, and a complete training update formula for the dynamic parameters in each round of training can be obtained; when step S4 is trained, the least squares identification value is selected as the initial value of the dynamic parameter set at the beginning of training; the training data is composed of the excitation trajectory data collected in step S1, and the training data includes the regression matrix, the measured joint velocity and the joint torque; a part of the training data is taken as the training set, and a part of the data is taken as the test set to verify the training effect; the optimization method used for training is the gradient descent method, and the initial learning rate is set to 0.0001, and the dynamic parameter set is optimized with the help of the Adam optimizer in the PyTorch library; when the loss function converges, the dynamic parameter set corresponding to this round of training is saved as the final identified dynamic parameter set.

[0132] The present invention trains the kinetic parameters under a large amount of training set data based on the back propagation algorithm, so that the kinetic parameters learn the mapping relationship between the input and output of the kinetic model during the training process. The trained kinetic parameters can not only fit the modeled kinetic characteristics, but also compensate for the unmodeled kinetic factors. Therefore, the present invention can effectively improve the integrity of the model and the identification accuracy of the parameters. The present invention utilizes the powerful data mapping ability and self-learning function of the back propagation learning algorithm, overcomes the limitation that the least square method is easily affected by noise, and can provide guidance for the identification of nonlinear parameters.

[0133] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that deviate from the spirit and principles of the present invention shall be equivalent replacement methods and shall be included in the protection scope of the present invention.

Claims

1. A method for identifying robot dynamic parameters, characterized in that: The following steps are involved: S1, collect joint motion data of multiple groups of robots when executing the excitation trajectory, collect relevant data at each time sampling point, and use the second-order central differential method to calculate the joint acceleration, and combine the filtered and denoised joint motion data into a training set; S2. Establish a dynamic model to calculate the corresponding theoretical joint torque under the training set data, calculate the mean square error between the theoretical joint torque and the collected measured joint torque, and define it as a loss function; the dynamic model is: τ=τ b +t f Among them, q, They represent the robot joint angle, joint angular velocity and joint acceleration respectively; τ represents the joint driving torque; τ b It represents the resultant moment of inertia, Coriolis moment, centripetal moment and gravity moment; Represents the regression matrix, which is only related to the robot's state q, Related; β represents the minimum identifiable inertial parameter set, which is obtained by reorganizing the complete dynamic parameters of the robot; τ f represents the joint friction torque, which adopts the Stribeck friction model, f c 、f s 、V s ,ξ,f v 、f p represents the Stribeck friction coefficient set, so the minimum inertia parameter set and the friction coefficient set together constitute the dynamic parameter set to be identified; ξ represents the Stribeck index parameter; The overall architecture of a generalized neural network consists of: Input layer X: Hidden Layer: Output layer: Where i = 1, 2, ..., 6 represents the robot joints; βl represents the minimum inertial parameter set β = [β1, β2, ..., β 40 ] the lth parameter in ; Represents the regression matrix The lth element in the i-th row of ; represents the joint velocity of the i-th joint; f c，i represents the Coulomb friction parameter, f s，i represents the static friction parameter, f v，i represents the viscous friction parameter, V s，i represents the Stribeck velocity, ξ i represents the Stribeck index parameter, f p，i represents the bias parameter; It represents the theoretical calculated value of the resultant moment of inertia, Coriolis moment, centripetal moment and gravity moment of joint i. represents the theoretical calculated value of the friction torque of joint i, and The sum represents the theoretical joint torque calculated based on the dynamic model; The loss function of the generalized neural network is defined as the mean square error between the theoretical joint torque and the measured joint torque of the robot's six joints. The generalized neural network has the following properties: (1) The connection weights of the network are the parameters to be identified or functions of the parameters to be identified; (2) The default value of the neuron bias is 0; (3) The activation function is determined by the functional relationship in the established dynamic model; The loss function is: Among them, τ i represents the measured joint torque of joint i collected, represents the calculated theoretical joint torque of joint i; S3. In each round of training, the gradient of the loss function with respect to the dynamics parameters is calculated based on the back propagation algorithm; In each round of training, the gradient of the loss function E with respect to the dynamics parameters is calculated as: For the minimum inertia parameter set β l : where β l represents the minimum inertial parameter set β=[β1,β2,…,β 40 ] the lth parameter in ; Represents the regression matrix The lth element in the i-th row of ; τ i represents the measured joint torque of joint i collected, represents the calculated theoretical joint torque of joint i; For the friction parameter set, δ is defined as the error term corresponding to the friction parameter set: For the Coulomb friction parameter f c,i : where ξ i represents the Stribeck index parameter, Represents the joint velocity of joint i: For the static friction parameter f s,i : For the viscous friction parameter f v,i : For Stribeck speed V s,i : For the Stribeck index parameter ξ i : For the bias parameter f p,i : S4, the kinetic parameters are iteratively trained along the negative gradient direction of the loss function. As the training progresses, the loss function gradually decreases. When the loss function converges to the minimum value, the corresponding kinetic parameters are saved as the identification results, thereby realizing the identification of the kinetic parameters; According to the gradient of the loss function with respect to the dynamic parameters, the dynamic parameters are iteratively updated in the direction of the negative gradient of the loss function based on the gradient descent method, and the loss function value gradually decreases; when the training is completed, the loss function value converges to the minimum value, and the corresponding dynamic parameter set at this time is saved as the identification result, completing the parameter identification; The method for iterative training and updating of dynamic parameters along the negative gradient direction of the loss function is the gradient descent method: I j (k+1)=Θ j (k)-DTH j (k) Among them, k represents the number of training times, Θ t represents the set of dynamic parameters, η represents the learning rate, ΔΘ j (k) represents the amount of gradient descent during the kth training. Respectively represent the modified first-order and second-order gradient momentum, and their update method is: m t (k)=α1·m t (k-1)+(1-α1)·g(k) V t (k)=α2·V t (k-1)+(1-α2)·g 2 (k) Where k represents the number of training times, α1 and α2 represent the parameters for controlling the decay of the first-order and second-order gradient momentum, g(k) represents the gradient of the loss function with respect to the set of dynamic parameters, and m t (k), V t (k) represent the first and second order gradient momentum before correction.

2. A method for identifying robot dynamic parameters according to claim 1, characterized in that: The excitation trajectory described in step S1 is a periodic Fourier series: Where i represents the robot joint, t represents the discrete time sequence, l = 1, 2, ..., n H represents the harmonic number, q(i,t) represents the joint angle of joint i at each discrete time point, and q i,0 represents the bias term, Respectively represent the amplitude terms of the sine function and the cosine function, w f Indicates the fundamental angular frequency.

3. A method for identifying robot dynamic parameters according to claim 1, characterized in that: The dynamic model described in step S2 is a generalized neural network. The input of the generalized neural network includes joint angles, joint angular velocities and joint accelerations. The output of the network includes theoretical joint torques calculated using the dynamic model. The dynamic parameters are the connection weights of the network and can be trained and learned.

4. A method for identifying robot dynamic parameters according to claim 1, characterized in that: The robot includes a rotary articulated robot, and the rotary articulated robot includes a serial articulated industrial robot.

Images