A robot compliant control method, system and device with parameter self-learning

Abstract

The present invention belongs to the field of robot control, and specifically relates to a robot compliance control method, system and device with parameter self-learning. The method includes constructing a joint dynamics model and a joint impedance control model of the robot to obtain an unknown dynamic model of the robot; constructing a Gaussian process that is structurally consistent with the unknown dynamic model of the robot, recorded as a structurally consistent Gaussian process; constructing a training set based on a data optimization method, and using the training set to offline train the hyperparameters of the structurally consistent Gaussian process; observing the robot state in real time, predicting the robot's unknown dynamic model online through a structurally consistent Gaussian process, and outputting the Gaussian prediction value of the unknown dynamic model; according to the Gaussian prediction value of the unknown dynamic model, outputting the control torque through the robot's joint impedance controller and sending it to the robot; judging whether the robot has reached the expected position, and if so, ending; otherwise continuing to observe the robot state. The present invention realizes high-precision control of the robot.

Description

Technical Field

[0001] The present invention belongs to the field of robot control, and in particular relates to a parameter self-learning robot compliance control method, system and device. Background Art

[0002] With the continuous advancement of robotics technology, robots have been widely used in various contact tasks, such as massage, human-machine interaction, and polishing. In unstructured environments, robots not only need to have high-precision motion control capabilities, but also need to significantly improve compliance performance to ensure the efficiency and effectiveness of contact operations while ensuring safety during the interaction process.

[0003] Impedance control is one of the basic control methods for robots to cope with complex contact tasks. Its application depends on accurate robot dynamic parameters. However, in actual operation, it is difficult to obtain these parameters. Therefore, intelligent impedance control methods that combine learning technology to cope with unknown dynamic models of robots are gradually becoming a research focus.

[0004] Adaptive impedance control is one of the most popular methods for dealing with unknown robot dynamics. The key idea of ​​adaptive impedance control is to design an appropriate adaptive law to estimate the robot dynamic parameters. The estimated parameters are then used by the impedance controller in order to ensure the stability of the closed-loop robot system and the convergence of the parameter estimator using the Lyapunov theorem. Adaptive impedance control methods are divided into four types according to the estimation objectives. The first type: directly designing adaptive laws for the robot dynamic model parameters, i.e., the robot's inertia, Coriolis / eccentricity matrix, and gravity vector. However, this type of adaptive control law can be quite complex due to the diversity of estimated parameters. The second type: based on the property that the robot dynamic model can be expressed in a linear form, i.e., a basis matrix multiplied by a weight parameter. However, collecting the basis matrix can be very complex. The third type: various bounds on the robot dynamic model parameters are estimated. Compared with the previous type, the estimated parameters of this type are scalars, which reduces the estimation complexity and may speed up the convergence. The fourth type: using a neural network to fit the unknown part of the robot dynamics, where the weight parameters of the neural network will be estimated by the adaptive law. Due to the strong fitting ability of neural networks, they can be used to fit different parts of the robot dynamics, thus different neural network-based adaptive impedance control methods are proposed.

[0005] Although adaptive impedance control uses adaptive techniques to estimate robot dynamic parameters, the main goal of parameter adaptation is not to reduce the estimation error, but to ensure the stability of the closed-loop system. Therefore, the adaptive control law is not intuitive enough and the dynamic control accuracy may be limited. A more intuitive and effective method is to directly use regression methods to fit the unknown robot dynamics and use the fitting results for impedance control. Due to the flexibility of the model and the ability to provide prediction uncertainty, Gaussian processes (GPs) have become popular in capturing unknown robot dynamics for robot control. Therefore, it is particularly important to use Gaussian processes to predict robot dynamic models to achieve high-precision robot control. However, most of the Gaussian process-based methods do not consider the special structure of the robot dynamics model in learning, which to some extent affects the reliability of learning and reduces the accuracy and stability of robot control. Therefore, considering the special structure of the robot dynamics model, it is of great significance to design a Gaussian process that is consistent with its structure to learn the unknown dynamic model. Summary of the invention

[0006] The object of the present invention is to provide a robot compliance control method, system and device with parameter self-learning to achieve high-precision control of the robot.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] A first aspect: The present invention provides a robot compliance control method with parameter self-learning, comprising:

[0009] Construct the robot's joint dynamics model and joint impedance control model to obtain the robot's unknown dynamic model;

[0010] Construct a Gaussian process that is consistent with the unknown dynamic model structure of the robot, denoted as a structurally consistent Gaussian process;

[0011] A training set is constructed based on data optimization methods, and the hyperparameters of the structurally consistent Gaussian process are trained offline using the training set.

[0012] Observe the robot state in real time, predict the robot's unknown dynamic model online through the structural consistency Gaussian process, and output the Gaussian prediction value of the unknown dynamic model;

[0013] According to the Gaussian prediction value of the unknown dynamic model, the control torque is output through the robot's joint impedance controller and sent to the robot;

[0014] Determine whether the robot has reached the expected position. If so, end the process; otherwise, continue to observe the robot status.

[0015] Preferably, the step of constructing a joint dynamics model and a joint impedance control model of the robot to obtain an unknown dynamic model of the robot includes:

[0016] Build The joint dynamics model of the DOF robot is:

[0017] ;

[0018] In the formula, Respectively represent the position, velocity and acceleration of the robot joints, represents the inertia matrix of the robot, represents the Coriolis force and centrifugal force matrix of the robot, represents the gravity torque of the robot, represents the friction torque of the robot, They represent the robot control torque and contact torque respectively, express dimensional vector, Expressed as Matrix of

[0019] The joint impedance control model is constructed as:

[0020] ;

[0021] In the formula, They represent the inertia matrix, damping matrix and stiffness matrix in the joint impedance control model respectively. They represent the position error, velocity error and acceleration error of the robot joints respectively. Representing the expected trajectory of the robot joint, we can get:

[0022] ;

[0023] In the formula, is the expected acceleration of the robot joint, and the robot joint impedance control rate is obtained as:

[0024] ;

[0025] In the formula, They represent the components containing impedance parameters and unknown dynamic models respectively, is the expected velocity of the robot joint, let:

[0026] ;

[0027] In the formula, , As a constant matrix, we get:

[0028] ;

[0029] In the formula, take Respectively The nominal value of Rewritten as:

[0030] ;

[0031] In the formula, is the unknown dynamic model of the robot, expressed as:

[0032] ;

[0033] In the formula, They are the inertia force, Coriolis force and friction force in the unknown dynamic model of the robot, and the unknown dynamic model of the robot satisfies the following structure:

[0034] ;

[0035] It can be seen from the above formula that when the input hour, ;when or hour, ;when hour, .

[0036] Preferably, the Gaussian process consistent with the unknown dynamic model structure of the robot is constructed as follows:

[0037] Due to the inertial forces in the unknown dynamic model of the robot Coriolis force and friction Both Dimension, so design independent Gaussian processes predict various dimensions of

[0038] Defining parameters ,parameter and parameters ,set up They obey the following Gaussian distributions respectively:

[0039] ;

[0040] In the formula, represents a Gaussian distribution, represents the transpose operation of the matrix, represents the position of the robot joints, represents the velocity of the robot joint, is the desired velocity of the robot joint, is the expected acceleration of the robot joint, represents the variance, , and denote the inputs corresponding to inertial force, Coriolis force and friction force from the training set, respectively. represents the variance of the inertial force in the unknown dynamic model, represents the variance of the Coriolis force in the unknown dynamic model, represents the variance of the friction force in the unknown dynamic model and is expressed as:

[0041] ;

[0042] In the formula, represents the expected value, represents the mean;

[0043] Inertial force Coriolis force and friction Each item is independent of each other, that is, the covariance between the three items is 0:

[0044] ;

[0045] In the formula, represents the covariance, then the compensation term The variance function of is:

[0046] ;

[0047] In the formula, and denote the actual value input and the input of the unknown dynamic model from the training set respectively;

[0048] Similarly, They are:

[0049] ;

[0050] So, for the prior The joint probability distribution is as follows:

[0051] ;

[0052] In the formula, represents the probability, represents the inertial force prior of the unknown dynamic model, represents the Coriolis force prior of the unknown dynamic model, represents the friction prior of the unknown dynamic model, The inertial force prior The mean of Coriolis force prior The mean of Denotes the friction prior The mean of represents the training set, Indicates the output, represents a Gaussian distribution, Indicates the length of the robot's position data. , and denote the prior kernel function values ​​corresponding to inertial force, Coriolis force and friction force respectively, , , denote the kernel function vectors corresponding to inertial force, Coriolis force and friction force, respectively. represents the kernel function matrix, represents the noise variance, represents the unit matrix, , , Indicates the training set data, Indicates the corresponding The kernel function value of represents the training set, so:

[0053] ;

[0054] In the formula, represents the Gaussian prediction of the inertial force in the unknown dynamic model, represents the Gaussian prediction of the Coriolis force in the unknown dynamic model, represents the Gaussian prediction of friction in the unknown dynamic model, represents the joint variance, where:

[0055] ;

[0056] Therefore, for the unknown dynamic model , whose Gaussian process prediction is:

[0057] ;

[0058] In the formula, represents the Gaussian prediction value of the unknown dynamic model, represents the Gaussian prediction variance of the unknown dynamic model;

[0059] And for inertial force Coriolis force and friction Design the corresponding kernel functions respectively as follows:

[0060] ;

[0061] ;

[0062] ;

[0063] In the formula, , , , , and All represent intermediate variables. represents the exponential function, , and represents the hyperparameters of the kernel function corresponding to the inertial force, , and represents the hyperparameters of the kernel function corresponding to the Coriolis force, , and Hyperparameters of the kernel function corresponding to the friction force.

[0064] Preferably, the step of constructing a training set based on a data optimization method comprises:

[0065] Get the location of the robot input data ,speed and acceleration , and get the robot output data at the same time , represents the control torque of the robot, represents the nominal value of the unknown dynamic model;

[0066] Position As the input data of the convex optimization method, it is denoted as , is the data dimension, is the robot's degree of freedom, express OK Column matrix, the data denoising process is transformed into a convex optimization problem as follows:

[0067] ;

[0068] In the formula, Representation Matrix of norm, Represents input data The result after denoising, that is, the position matrix after denoising, are intermediate variables, which are defined as follows:

[0069] ;

[0070] In the formula, The dimension is The unit array, The dimension is The unit array, express dimensional vector, are convex optimization parameters respectively, and through numerical calculation we can get:

[0071] ;

[0072] In the formula, represents the first position matrix after noise reduction List, The dimension is The unit array, represents the convex optimization parameter No. List, represents the convex optimization parameter No. List, Represents input data No. List, ;

[0073] Therefore, the velocity matrix after denoising is And the acceleration matrix after denoising It is expressed as:

[0074] ;

[0075] ;

[0076] In the formula, express OK A matrix of columns, express OK A matrix of columns, represents the training set, is the sampling period.

[0077] Preferably, the joint impedance controller of the robot is expressed as follows:

[0078] ;

[0079] In the formula, is the control torque of the robot, is the intermediate variable, and is a constant matrix, represents the position error of the robot joint, represents the velocity error of the robot joint, represents the gravity torque of the robot, represents the nominal value of the robot's inertia matrix, represents the desired acceleration of the robot joints, represents the nominal value of the Coriolis force and centrifugal force matrix of the robot, represents the desired velocity of the robot joint, represents the nominal value of the friction torque of the robot, Represents the Gaussian prediction value of the unknown dynamic model.

[0080] In the second aspect, the present invention provides a parameter self-learning robot compliance control system, comprising a Gaussian process building module, an offline training module and an online observation control module, wherein:

[0081] The Gaussian process building module is configured to: build a joint dynamics model and a joint impedance control model of the robot to obtain an unknown dynamic model of the robot; build a Gaussian process that is structurally consistent with the unknown dynamic model of the robot, which is recorded as a structurally consistent Gaussian process;

[0082] The offline training module is configured to: construct a training set based on a data optimization method, and use the training set to offline train the hyperparameters of the structural consistency Gaussian process;

[0083] The online observation control module is configured to: observe the robot state in real time, predict the unknown dynamic model of the robot online through a structural consistency Gaussian process, and output the Gaussian prediction value of the unknown dynamic model; output the control torque through the robot's joint impedance controller and send it to the robot according to the Gaussian prediction value of the unknown dynamic model; determine whether the robot has reached the expected position, and if so, end; otherwise, continue to observe the robot state.

[0084] A third aspect: The present invention provides a robot compliance control device with parameter self-learning, comprising a processor and a memory storing a plurality of computer instructions, wherein the computer instructions, when executed by the processor, implement the steps of the robot compliance control method with parameter self-learning.

[0085] In order to reduce the performance impact caused by the uncertainty of the dynamic model of the robot during operation and take into account the compliance performance of the robot, the present invention is based on Gaussian process regression, proposes an impedance control method, system and device based on a structurally consistent Gaussian process, and designs a Gaussian process consistent with the dynamic model of the robot. The robot dynamic model is predicted and compensated online through the Gaussian process, which effectively improves the comprehensive performance of the robot control system, thereby achieving high-precision control in the robot processing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 A flow chart of a robot compliance control method with parameter self-learning according to the present invention;

[0087] Figure 2 This is a diagram of the prediction results of the unknown dynamic model in the experiment of the present invention;

[0088] Figure 3 This is a comparison chart of trajectory tracking errors in the experiment of the present invention. DETAILED DESCRIPTION

[0089] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0090] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0091] Example 1: Figure 1 As shown, this embodiment provides a parameter self-learning robot compliance control method, which is applied to a robot control system and specifically includes the following steps:

[0092] Step 1: Construct the joint dynamics model and joint impedance control model of the robot to obtain the unknown dynamic model of the robot.

[0093] Build The joint dynamics model of the DOF robot is:

[0094] ;

[0095] In the formula, Respectively represent the position, velocity and acceleration of the robot joints, represents the inertia matrix of the robot, represents the Coriolis force and centrifugal force matrix of the robot, represents the gravity torque of the robot, represents the friction torque of the robot, They represent the robot control torque and contact torque respectively, express dimensional vector, Expressed as The matrix of .

[0096] The joint impedance control model is constructed as:

[0097] ;

[0098] In the formula, They represent the inertia matrix, damping matrix and stiffness matrix in the joint impedance control model respectively. They represent the position error, velocity error and acceleration error of the robot joints respectively. Representing the expected trajectory of the robot joint, we can get:

[0099] ;

[0100] In the formula, is the expected acceleration of the robot joint, so the robot joint impedance control rate is:

[0101] ;

[0102] In the formula, are intermediate variables, representing the components containing impedance parameters and unknown dynamic models, is the expected speed of the robot joint. In order to express the formula clearly, the following content of this embodiment will Omit, that is, take .make:

[0103] ;

[0104] In the formula, , As a constant matrix, we get:

[0105] ;

[0106] Due to the strong coupling and nonlinear characteristics of the robot, It is difficult to obtain accurate information in practical applications, so Respectively represent their nominal values, then Rewritten as:

[0107] ;

[0108] In the formula, is the unknown dynamic model of the robot, expressed as:

[0109] ;

[0110] In the formula, They are the inertia force, Coriolis force and friction force in the unknown dynamic model of the robot. It can be seen that the unknown dynamic model of the robot satisfies the following structure:

[0111] ;

[0112] It can be seen from the above formula that when the input hour, ;when or hour, ;when hour, Based on this, the present invention designs a Gaussian process with consistent structure to conform to the structure. Indicates that the input is and The inertial force when Indicates that the input is , and The Coriolis force at Indicates that the input is , and The Coriolis force at Indicates that the input is The friction force when express dimensional zero-valued matrix.

[0113] Step 2: Construct a Gaussian process that is consistent with the unknown dynamic model structure of the robot, denoted as a structurally consistent Gaussian process.

[0114] Due to the inertial forces in the unknown dynamic model of the robot Coriolis force and friction Both Dimension, so design independent Gaussian processes predict various dimensions of

[0115] Defining parameters ,parameter and parameters ,set up They obey the following Gaussian distributions respectively:

[0116] ;

[0117] In the formula, represents a Gaussian distribution, represents the transpose operation of the matrix, represents the position of the robot joints, represents the velocity of the robot joint, is the desired velocity of the robot joint, is the expected acceleration of the robot joint, represents the variance, , and denote the inputs corresponding to inertial force, Coriolis force and friction force from the training set, respectively. represents the variance of the inertial force in the unknown dynamic model, represents the variance of the Coriolis force in the unknown dynamic model, represents the variance of the friction force in the unknown dynamic model and is expressed as:

[0118] ;

[0119] In the formula, represents the expected value, represents the mean, Indicates that the input is The variance of , the other similar understanding, Indicates that the input is The expected value of Indicates that the input is The variance of , and the other understandings are similar.

[0120] Inertial force Coriolis force and friction Each item is independent of each other, that is, the covariance between the three items is 0:

[0121] ;

[0122] In the formula, represents the covariance, Indicates that the input is The covariance of , and the other is understood in the same way. Then the compensation term The variance function of is:

[0123] ;

[0124] In the formula, and represent the actual value input and the input of the unknown dynamic model from the training set, respectively.

[0125] Similarly, They are:

[0126] ;

[0127] So, for the prior The joint probability distribution is as follows:

[0128] ;

[0129] In the formula, represents the probability, represents the inertial force prior of the unknown dynamic model, represents the Coriolis force prior of the unknown dynamic model, represents the friction prior of the unknown dynamic model, The inertial force prior The mean of Coriolis force prior The mean of Denotes the friction prior The mean of represents the training set, Indicates the output, represents a Gaussian distribution, Indicates the robot's position data length, , and denote the prior kernel function values ​​corresponding to inertial force, Coriolis force and friction force respectively, , , denote the kernel function vectors corresponding to inertial force, Coriolis force and friction force, respectively. represents the kernel function matrix, represents the noise variance, represents the unit matrix, , , Indicates the training set data, Indicates the corresponding The kernel function value of represents the training set, so:

[0130] ;

[0131] In the formula, represents the Gaussian prediction of the inertial force in the unknown dynamic model, represents the Gaussian prediction of the Coriolis force in the unknown dynamic model, represents the Gaussian prediction of friction in the unknown dynamic model, represents the joint variance, where:

[0132] ;

[0133] Therefore, for the unknown dynamic model , whose Gaussian process prediction is:

[0134] ;

[0135] In the formula, represents the Gaussian prediction value of the unknown dynamic model, represents the Gaussian prediction variance of the unknown dynamic model;

[0136] In order to make the Gaussian process consistent with the unknown dynamic model of the robot, this embodiment targets the inertial force Coriolis force and friction Design the corresponding kernel functions respectively To satisfy the structure of the unknown dynamic model, as follows:

[0137] ;

[0138] ;

[0139] ;

[0140] In the formula, , , , , and All represent intermediate variables. represents the exponential function, , and represents the hyperparameters of the kernel function corresponding to the inertial force, , and represents the hyperparameters of the kernel function corresponding to the Coriolis force, , and Hyperparameters of the kernel function corresponding to the friction force.

[0141] Step 3: Construct a training set based on the data optimization method, and use the training set to offline train the hyperparameters of the structural consistency Gaussian process.

[0142] Before using the structural consistency Gaussian process for impedance control, it is necessary to collect relevant data and construct a training set. The training results of the hyperparameters directly affect the accuracy of the robot's unknown dynamic model, and the training set is the fundamental factor affecting the hyperparameter training. Therefore, it is necessary to perform noise reduction on the constructed training set to improve the hyperparameter training effect. This embodiment uses a convex optimization method as a data optimization method. Based on the convex optimization method, the position obtained by the sensor is ,speed and acceleration Perform noise reduction and construct a training set as follows:

[0143] Based on the robot joint dynamics model, obtain the position of the robot input data ,speed and acceleration , and get the robot output data at the same time , represents the control torque of the robot, represents the nominal value of the unknown dynamic model. Where:

[0144] ;

[0145] So, for , whose input items are , , In practical applications, sensors can only obtain position ,speed , acceleration ,in is the sampling time. Very small, this paper uses time , in the calculation Therefore, when constructing the training set, the original data needs to be denoised. Indicates Speed ​​data, Indicates location data, Indicates location data, No. Speed ​​data.

[0146] Therefore, the position As the input data of the convex optimization method, it is denoted as , is the data dimension, is the robot's degree of freedom, express OK Column matrix, the data denoising process is transformed into a convex optimization problem as follows:

[0147] ;

[0148] In the formula, Representation Matrix of norm, Represents input data The result after denoising, that is, the position matrix after denoising, are intermediate variables, which are defined as follows:

[0149] ;

[0150] In the formula, The dimension is The unit array, The dimension is The unit array, express dimensional vector, are convex optimization parameters respectively, and through numerical calculation we can get:

[0151] ;

[0152] In the formula, represents the first position matrix after noise reduction List, The dimension is The unit array, represents the convex optimization parameter No. List, represents the convex optimization parameter No. List, Represents input data No. List, ;

[0153] Therefore, the velocity matrix after denoising is And the acceleration matrix after denoising , and the obtained training set is expressed as:

[0154] ;

[0155] ;

[0156] In the formula, express OK A matrix of columns, express OK A matrix of columns, represents the training set, is the sampling period. Therefore, the selection of the hyperparameters of the Gaussian process can be obtained by the maximum likelihood estimation of the training set.

[0157] Step 4: Observe the robot state in real time, predict the unknown dynamic model of the robot online through the structural consistency Gaussian process, and output the Gaussian prediction value of the unknown dynamic model .

[0158] Step 5: According to the Gaussian prediction value of the unknown dynamic model, the control torque is output through the joint impedance controller of the robot and sent to the robot. The joint impedance controller of the robot is expressed as follows:

[0159] ;

[0160] In the formula, is the control torque of the robot, is the intermediate variable, and is a constant matrix, represents the position error of the robot joint, represents the velocity error of the robot joint, represents the gravity torque of the robot, represents the nominal value of the robot's inertia matrix, represents the desired acceleration of the robot joints, represents the nominal value of the Coriolis force and centrifugal force matrix of the robot, represents the desired velocity of the robot joint, represents the nominal value of the friction torque of the robot, Represents the Gaussian prediction value of the unknown dynamic model.

[0161] Step 6: Determine whether the robot has reached the expected position. If so, end; otherwise, continue to observe the robot status and return to step 4.

[0162] The present invention predicts the robot dynamic model online by designing a Gaussian process consistent with the robot dynamic model structure And compensate the control torque of the robot joint impedance control , achieving high-precision control of the robot.

[0163] In the experiment, a training set was constructed based on a massage robot platform to verify the effectiveness of the proposed method in predicting the unknown dynamic model of the robot. The robot has 7 degrees of freedom, and the prediction results of the unknown dynamic model are as follows: Figure 2 As shown, the vertical axis is the unknown dynamic model of the robot, and the horizontal axis steps represents the iteration steps. For the robot The actual value of the unknown dynamic model of the joints, The robot output of the structurally consistent Gaussian process Gaussian prediction value of the unknown dynamic model of the joints, It can be seen that the fluctuation of the predicted value closely follows the actual value, which shows that the method of the present invention is effective in predicting the unknown dynamic model of the robot.

[0164] In order to verify the effectiveness and superiority of the method of the present invention, the uncompensated dynamic model and the robot impedance control method based on the Gaussian process are compared based on the massage robot platform. The trajectory tracking error is compared. Figure 3 As shown, the vertical axis error represents the trajectory tracking error, in radians, and the horizontal axis steps represents the iteration steps, where represents the robot impedance control method under the uncompensated unknown dynamic model, represents the conventional Gaussian process-based robot impedance control method, represents the impedance control method of the present invention, Indicates running Method The trajectory tracking error of each joint is Indicates running Method The trajectory tracking error of each joint is Indicates running Method The trajectory tracking error of the joints, subscript It can be seen that the present invention effectively reduces the trajectory tracking error of each joint and realizes high-precision control of the robot.

[0165] Embodiment 2: This embodiment provides a parameter self-learning robot compliance control system, including a Gaussian process construction module, an offline training module and an online observation control module, wherein:

[0166] The Gaussian process building module is configured to: build a joint dynamics model and a joint impedance control model of the robot to obtain an unknown dynamic model of the robot; build a Gaussian process that is structurally consistent with the unknown dynamic model of the robot, which is recorded as a structurally consistent Gaussian process;

[0167] The offline training module is configured to: construct a training set based on a data optimization method, and use the training set to offline train the hyperparameters of the structural consistency Gaussian process;

[0168] The online observation control module is configured to: observe the robot state in real time, predict the unknown dynamic model of the robot online through the structural consistency Gaussian process, and output the Gaussian prediction value of the unknown dynamic model; output the control torque through the robot's joint impedance controller according to the Gaussian prediction value of the unknown dynamic model and send it to the robot; determine whether the robot has reached the expected position, and if so, end; otherwise, continue to observe the robot state.

[0169] For the specific limitations of the robot compliant control system with parameter self-learning, please refer to the limitations of the robot compliant control method with parameter self-learning in the above text, which will not be repeated here.

[0170] Embodiment 3: This embodiment provides a robot compliance control device with parameter self-learning, including a processor and a memory storing a plurality of computer instructions, wherein the computer instructions, when executed by the processor, implement the steps of the robot compliance control method with parameter self-learning.

[0171] The specific limitations on the parameter self-learning robot compliance control device can be found in the above limitations on the parameter self-learning robot compliance control method, which will not be repeated here.

[0172] The memory and the processor are electrically connected directly or indirectly to achieve data transmission or interaction. For example, these elements can be electrically connected to each other through one or more communication buses or signal lines. The memory stores a computer program that can be run on the processor, and the processor implements the method of the present invention by running the computer program stored in the memory.

[0173] The memory may be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), etc. The memory is used to store a program, and the processor executes the program after receiving an execution instruction.

[0174] The processor may be an integrated circuit chip with data processing capabilities. The above-mentioned processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc. The methods, steps and logic block diagrams disclosed in the embodiments of the present invention may be implemented or executed. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0175] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0176] The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A robot compliance control method with parameter self-learning, characterized in that: The parameter self-learning robot compliance control method comprises: Construct the robot's joint dynamics model and joint impedance control model to obtain the robot's unknown dynamic model; Construct a Gaussian process that is consistent with the unknown dynamic model structure of the robot, denoted as a structurally consistent Gaussian process; A training set is constructed based on data optimization methods, and the hyperparameters of the structurally consistent Gaussian process are trained offline using the training set. Observe the robot state in real time, predict the robot's unknown dynamic model online through the structural consistency Gaussian process, and output the Gaussian prediction value of the unknown dynamic model; According to the Gaussian prediction value of the unknown dynamic model, the control torque is output through the robot's joint impedance controller and sent to the robot; Determine whether the robot has reached the expected position. If so, the process ends; otherwise, continue to observe the robot status. The Gaussian process consistent with the unknown dynamic model structure of the robot is constructed as follows: Due to the inertial forces in the unknown dynamic model of the robot Coriolis force and friction Both Dimension, so design independent Gaussian processes predict various dimensions of Defining parameters ,parameter and parameters ,set up They obey the following Gaussian distributions respectively: ； In the formula, represents a Gaussian distribution, represents the transpose operation of the matrix, represents the position of the robot joints, represents the velocity of the robot joint, is the desired velocity of the robot joint, is the expected acceleration of the robot joint, represents the variance, , and denote the inputs corresponding to inertial force, Coriolis force and friction force from the training set, respectively. represents the variance of the inertial force in the unknown dynamic model, represents the variance of the Coriolis force in the unknown dynamic model, represents the variance of the friction force in the unknown dynamic model and is expressed as: ； In the formula, represents the expected value, represents the mean; Inertial force Coriolis force and friction Each item is independent of each other, that is, the covariance between the three items is 0: ； In the formula, represents the covariance, then the compensation term The variance function of is: ； In the formula, and denote the actual value input and the input of the unknown dynamic model from the training set respectively; Similarly, They are: ； So, for the prior The joint probability distribution is as follows: ； In the formula, represents the probability, represents the inertial force prior of the unknown dynamic model, represents the Coriolis force prior of the unknown dynamic model, represents the friction prior of the unknown dynamic model, The inertial force prior The mean of Coriolis force prior The mean of Denotes the friction prior The mean of represents the training set, Indicates the output, represents a Gaussian distribution, Indicates the robot's position data length, , and denote the prior kernel function values ​​corresponding to inertial force, Coriolis force and friction force respectively, , , denote the kernel function vectors corresponding to inertial force, Coriolis force and friction force, respectively. represents the kernel function matrix, represents the noise variance, represents the unit matrix, , , Indicates the training set data, Indicates the corresponding The kernel function value of represents the training set, so: ； In the formula, represents the Gaussian prediction of the inertial force in the unknown dynamic model, represents the Gaussian prediction of the Coriolis force in the unknown dynamic model, represents the Gaussian prediction of friction in the unknown dynamic model, represents the joint variance, where: ； Therefore, for the unknown dynamic model , whose Gaussian process prediction is: ； In the formula, represents the Gaussian prediction value of the unknown dynamic model, represents the Gaussian prediction variance of the unknown dynamic model; And for inertial force Coriolis force and friction Design the corresponding kernel functions respectively as follows: ； ； ； In the formula, , , , , and All represent intermediate variables. represents the exponential function, , and represents the hyperparameters of the kernel function corresponding to the inertial force, , and represents the hyperparameters of the kernel function corresponding to the Coriolis force, , and Hyperparameters of the kernel function corresponding to the friction force.

2. The robot compliance control method of parameter self-learning according to claim 1 is characterized in that: The step of constructing a joint dynamics model and a joint impedance control model of the robot to obtain an unknown dynamic model of the robot includes: Build The joint dynamics model of the DOF robot is: ； In the formula, Respectively represent the position, velocity and acceleration of the robot joints, represents the inertia matrix of the robot, represents the Coriolis force and centrifugal force matrix of the robot, represents the gravity torque of the robot, represents the friction torque of the robot, They represent the robot control torque and contact torque respectively, express dimensional vector, Expressed as Matrix of The joint impedance control model is constructed as: ； In the formula, They represent the inertia matrix, damping matrix and stiffness matrix in the joint impedance control model respectively. They represent the position error, velocity error and acceleration error of the robot joints respectively. Representing the expected trajectory of the robot joint, we can get: ； In the formula, is the expected acceleration of the robot joint, and the robot joint impedance control rate is obtained as: ； In the formula, They represent the components containing impedance parameters and unknown dynamic models respectively, is the expected velocity of the robot joint, let: ； In the formula, , As a constant matrix, we get: ； In the formula, take Respectively The nominal value of Rewritten as: ； In the formula, is the unknown dynamic model of the robot, expressed as: ； In the formula, They are the inertia force, Coriolis force and friction force in the unknown dynamic model of the robot, and the unknown dynamic model of the robot satisfies the following structure: ； It can be seen from the above formula that when the input hour, ;when or hour, ;when hour, .

3. The robot compliance control method of parameter self-learning according to claim 1 is characterized in that: The method of constructing a training set based on the data optimization method comprises: Get the location of the robot input data ,speed and acceleration , and get the robot output data at the same time , represents the control torque of the robot, represents the nominal value of the unknown dynamic model; Position As the input data of the data method, it is denoted as , is the data dimension, is the robot's degree of freedom, express OK Column matrix, transforming the data optimization process into a convex optimization problem, as follows: ； In the formula, Representation Matrix of norm, Represents input data The result after denoising, that is, the position matrix after denoising, are intermediate variables, which are defined as follows: ； In the formula, The dimension is The unit array, The dimension is The unit array, express dimensional vector, are convex optimization parameters respectively, and through numerical calculation we can get: ； In the formula, represents the first position matrix after noise reduction List, The dimension is The unit array, represents the convex optimization parameter No. List, represents the convex optimization parameter No. List, Represents input data No. List, ; Therefore, the velocity matrix after denoising is And the acceleration matrix after denoising It is expressed as: ； ； In the formula, express OK A matrix of columns, express OK A matrix of columns, represents the training set, is the sampling period.

4. The robot compliance control method of parameter self-learning according to claim 1 is characterized in that: The joint impedance controller of the robot is expressed as follows: ； In the formula, is the control torque of the robot, is the intermediate variable, and is a constant matrix, represents the position error of the robot joint, represents the velocity error of the robot joint, represents the gravity torque of the robot, represents the nominal value of the robot's inertia matrix, represents the desired acceleration of the robot joints, represents the nominal value of the Coriolis force and centrifugal force matrix of the robot, represents the desired velocity of the robot joint, represents the nominal value of the friction torque of the robot, Represents the Gaussian prediction value of the unknown dynamic model.

5. A robot compliant control system with self-learning parameters, characterized in that: The parameter self-learning robot compliance control system includes a Gaussian process building module, an offline training module and an online observation control module, wherein: The Gaussian process building module is configured to: build a joint dynamics model and a joint impedance control model of the robot to obtain an unknown dynamic model of the robot; build a Gaussian process that is structurally consistent with the unknown dynamic model of the robot, which is recorded as a structurally consistent Gaussian process; The offline training module is configured to: construct a training set based on a data optimization method, and use the training set to offline train the hyperparameters of the structural consistency Gaussian process; The online observation control module is configured to: observe the robot state in real time, predict the unknown dynamic model of the robot online through the structural consistency Gaussian process, and output the Gaussian prediction value of the unknown dynamic model; output the control torque through the joint impedance controller of the robot according to the Gaussian prediction value of the unknown dynamic model and send it to the robot; determine whether the robot has reached the expected position, and if so, end; otherwise, continue to observe the robot state; The Gaussian process consistent with the unknown dynamic model structure of the robot is constructed as follows: Due to the inertial forces in the unknown dynamic model of the robot Coriolis force and friction Both Dimension, so design independent Gaussian processes predict various dimensions of Defining parameters ,parameter and parameters ,set up They obey the following Gaussian distributions respectively: ； In the formula, represents a Gaussian distribution, represents the transpose operation of the matrix, represents the position of the robot joints, represents the velocity of the robot joint, is the desired velocity of the robot joint, is the expected acceleration of the robot joint, represents the variance, , and denote the inputs corresponding to inertial force, Coriolis force and friction force from the training set, respectively. represents the variance of the inertial force in the unknown dynamic model, represents the variance of the Coriolis force in the unknown dynamic model, represents the variance of the friction force in the unknown dynamic model and is expressed as: ； In the formula, represents the expected value, represents the mean; Inertial force Coriolis force and friction Each item is independent of each other, that is, the covariance between the three items is 0: ； In the formula, represents the covariance, then the compensation term The variance function of is: ； In the formula, and denote the actual value input and the input of the unknown dynamic model from the training set respectively; Similarly, They are: ； So, for the prior The joint probability distribution is as follows: ； In the formula, represents the probability, represents the inertial force prior of the unknown dynamic model, represents the Coriolis force prior of the unknown dynamic model, represents the friction prior of the unknown dynamic model, The inertial force prior The mean of Coriolis force prior The mean of Denotes the friction prior The mean of represents the training set, Indicates the output, represents a Gaussian distribution, Indicates the robot's position data length, , and denote the prior kernel function values ​​corresponding to inertial force, Coriolis force and friction force respectively, , , denote the kernel function vectors corresponding to inertial force, Coriolis force and friction force, respectively. represents the kernel function matrix, represents the noise variance, represents the unit matrix, , , Indicates the training set data, Indicates the corresponding The kernel function value of represents the training set, so: ； In the formula, represents the Gaussian prediction of the inertial force in the unknown dynamic model, represents the Gaussian prediction of the Coriolis force in the unknown dynamic model, represents the Gaussian prediction of friction in the unknown dynamic model, represents the joint variance, where: ； Therefore, for the unknown dynamic model , whose Gaussian process prediction is: ； In the formula, represents the Gaussian prediction value of the unknown dynamic model, represents the Gaussian prediction variance of the unknown dynamic model; And for inertial force Coriolis force and friction Design the corresponding kernel functions respectively as follows: ； ； ； In the formula, , , , , and All represent intermediate variables. represents the exponential function, , and represents the hyperparameters of the kernel function corresponding to the inertial force, , and represents the hyperparameters of the kernel function corresponding to the Coriolis force, , and Hyperparameters of the kernel function corresponding to the friction force.

6. A robot compliance control device with parameter self-learning, comprising a processor and a memory storing a plurality of computer instructions, characterized in that: When the computer instructions are executed by the processor, the steps of the robot compliant control method with parameter self-learning described in any one of claims 1 to 4 are implemented.

Images