Robot dynamics physical information prediction neural network construction method based on motor and external force coupling

Abstract

The invention discloses a robot dynamics physical information prediction neural network construction method based on motor and external force coupling, and relates to the technical field of robots. Comprising the following steps: embedding a motor kinetic equation into a DeLaN network, and establishing an explicit dynamic mapping relation between a system motion state and a motor current through mapping between the motor current and a joint torque; introducing an external force item by fitting a Jacobian matrix, and constructing approximate mapping of the Jacobian matrix by adopting a data driving method; and an initial model is constructed, and after motor dynamics and external force modeling are incorporated into the initial model, an inverse dynamics model of the mechanical arm is obtained. Compared with an existing most advanced method, the method has higher modeling precision and generalization ability and better control performance. The innovative scheme provides solid technical support for realizing real-time accurate control in the field of robots.

Description

Technical Field

[0001] The present invention relates to the field of robotics technology, and more particularly to a method for constructing a robot dynamics physical information prediction neural network based on the coupling of a motor and an external force. Background Art

[0002] With the development of industrial automation and intelligent technologies, robotic arms are widely used in complex environments such as parts assembly, medical operations, and human-machine interaction. External forces in interactive scenarios can significantly alter the system's dynamic characteristics. Therefore, high-precision dynamic modeling in complex dynamic environments is crucial for achieving precise control of robotic arms.

[0003] Traditional modeling methods are mainly based on physical principles (such as the Lagrange method and the Newton-Euler method) to derive dynamic equations. They rely on precise physical parameters and complex formula derivations, and contain a lot of simplifications and linearizations, making it difficult to describe complex nonlinear external forces.

[0004] Purely data-driven methods have been widely used in dynamics modeling in recent years. These approaches, independent of physical parameters, directly learn the system's dynamics from data, demonstrating excellent performance in complex nonlinear scenarios. However, the black-box nature of these methods prevents them from capturing the underlying physical structure, limiting the ability to gain a deeper understanding of the system and interpret the results. Overfitting and poor physical interpretability also hinder their integration into classical control frameworks.

[0005] To overcome the limitations of purely data-driven approaches, physically informed neural networks (PINNs), which combine physical principles with data-driven approaches, are increasingly being used for robot dynamics modeling. Physically informed neural networks (PINNs) embed fundamental physical principles as constraints into the architecture and training of neural networks, allowing the network to naturally follow the laws of physics during learning. For example, the Deep Lagrangian Network (DeLaN) embeds Lagrangian mechanics into the network, maintaining the symmetry and positive definiteness of the inertia matrix during learning. This not only effectively learns the system's dynamic equations, but also ensures physical interpretability. Due to the conservative nature of Lagrangian mechanics, DeLaNs are unable to model non-conservative forces with complex nonlinearities, such as friction and external forces.

[0006] In order to address the limitations of DeLaN in modeling non-conservative forces, some studies are now incorporating friction models into the system's dynamic model and adding dissipative force networks to deal with friction, achieving better modeling results when dealing with friction scenarios. However, for external contact forces, current attempts to embed contact forces into physical information neural networks are usually targeted at specific scenarios with restrictive assumptions, and therefore lack wide applicability. Other studies have introduced feedforward neural networks (FFNNs) based on the DeLaN model, which can learn all uncertainties that DeLaN cannot describe. However, due to the black box nature of FFNNs, the learned uncertainties lack physical explanations, and their outputs may violate known physical laws. In addition, since many robotic arms are not equipped with joint torque sensors, accurate joint torque data cannot be obtained. Summary of the Invention

[0007] In view of this, the present invention provides a method for constructing a robot dynamics physical information prediction neural network based on the coupling of motors and external forces, integrating motor dynamics and external force modeling into the deep Lagrangian network (DeLaN) framework; establishing an explicit dynamic mapping relationship between the system motion state and the motor current to make up for the lack of joint torque data; and proposing two methods for fitting the Jacobian matrix estimation method to introduce external force terms, thereby constructing two PINNs-based dynamic models. Experiments on different robotic arms have shown that the proposed models can achieve high-precision torque estimation under external disturbances without the need for joint torque sensors. Compared with the most advanced existing methods, it has higher modeling accuracy and generalization ability, as well as better control performance. This innovative solution provides solid technical support for achieving real-time and precise control in the field of robotics.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A method for constructing a neural network for predicting robot dynamics physical information based on the coupling of motors and external forces, comprising the following steps:

[0010] The motor dynamics equations are embedded into the DeLaN network, and an explicit dynamic mapping relationship between the system motion state and the motor current is established by mapping the motor current to the joint torque.

[0011] The external force term is introduced by fitting the Jacobian matrix, and the approximate mapping of the Jacobian matrix is ​​constructed using a data-driven method;

[0012] After constructing the initial model and incorporating the motor dynamics and external force modeling into the initial model, the inverse dynamics model of the robotic arm is obtained.

[0013] Optionally, the inverse dynamics model of the manipulator is expressed as follows:

[0014]

[0015] The network input is the robot arm joint angle q, joint angular velocity Joint angular acceleration and the terminal six-dimensional force F ext , the output is the motor current i, where The matrix representing the centrifugal and Coriolis forces,

[0016] g(q) represents the gravity vector, H(q) is the mass matrix, τ represents the generalized force acting on the generalized coordinate q of the system, i is the motor current, r is the reduction ratio, k t is the motor torque constant, τ F is the friction torque, τ ext For the external torque, the inverse dynamics model introduces the mapping relationship between the joint torque τ and the motor current i, and finally establishes the dynamic mapping relationship between the system motion state and the motor current.

[0017] Alternatively, the DC motor dynamics are described by the following equations:

[0018] τ=rk t i;

[0019] Where τ is the joint torque, i is the motor current, and k t is the motor torque constant, and r is the reduction ratio.

[0020] Optionally, the joint friction torque is:

[0021]

[0022] Among them, the constant ε is a smoothing parameter set to 10-3 based on experience, r is the reduction ratio, k v is the coefficient of friction, is the joint angular velocity; when power flows from the motor to the load, the Coulomb term T c The value of When the power direction is opposite, the Coulomb term T c The value of

[0023] Optionally, the external force term is described as follows:

[0024] τ ext =J T (q)F ext ;

[0025] τ ext is the external torque; F ext is the external force or torque of the end effector, and J(q) is the Jacobian matrix of the manipulator.

[0026] Optionally, the Jacobian matrix is ​​fitted using direct Jacobian as follows:

[0027] The robot's configuration consists of its rigid links and joint angles. Through forward kinematics, the joint angles q are converted to the end-effector pose x, i.e., x = f(q). The end-effector pose x describes the spatial configuration of the robot's end-effector in the task space and consists of a position vector and attitude parameters. The forward kinematics model is obtained by chain multiplication of a series of homogeneous transformation matrices, representing the transformation from the base coordinate system to the end-effector coordinate system. To analyze the relationship between the end-effector velocity and the joint velocity, the Jacobian matrix is ​​introduced and defined as:

[0028]

[0029] in is the joint velocity, is the terminal velocity, which is defined as the Jacobian matrix to establish the differential mapping relationship between the joint space and the task space. Therefore, the direct Jacobian method directly learns the mapping relationship between the terminal velocity and the joint velocity. The network input is the joint angle q, and the output is the Jacobian matrix J(q). The loss function is designed as follows:

[0030]

[0031] Where N represents the number of training samples, is the actual measured terminal velocity, is the terminal velocity predicted by the model, which is calculated as follows:

[0032]

[0033] in, Represents the Jacobian matrix of the network predictions.

[0034] Optionally, the Jacobian matrix can be obtained using the kinematic differentiation method as follows: first learn the system kinematic behavior and fit the forward kinematic model, that is, the mapping relationship between joint angles and end-point poses x = f(q), to capture the global nonlinearity of the kinematics and cover the entire workspace; calculate the partial derivatives of the forward kinematic model f(q) with respect to q to obtain the Jacobian matrix of the system:

[0035]

[0036] The input of the network is the joint angle q, and the output is the end effector pose x. It is trained through supervised learning, and the loss function is designed to predict the end effector pose Compared with the actual end effector pose The mean square error between:

[0037]

[0038] Optionally, the network parameter structure of the initial model is described as follows:

[0039]

[0040] θ, φ, and ψ represent the weight parameters of the network that need to be trained to learn the mass matrix H(q), potential energy V(q), and Jacobian matrix J(q), respectively. The parameter k represents the parameters that need to be learned in the motor dynamics model, including the reduction ratio r and the motor torque constant k. t 、Coulomb term T c and friction coefficient k v .

[0041] Through the above technical solutions, it can be seen that compared with the prior art, the present invention provides a method for constructing a robot dynamics physical information prediction neural network based on the coupling of motors and external forces, which effectively solves the limitations of physical information neural networks in modeling non-conservative forces (especially complex external force fields), as well as the problem of missing joint torque data in actual robot systems. The motor dynamics and external force modeling are integrated into the deep Lagrangian network framework. An explicit dynamic mapping relationship between the system motion state and the motor current is established to make up for the lack of joint torque data. Two methods are proposed to fit the Jacobian matrix estimation method to introduce external force terms, thereby constructing two PINNs-based dynamic models. Experiments on different robotic arms show that the proposed models can achieve high-precision torque estimation under external disturbances without the need for joint torque sensors. Compared with the most advanced existing methods, it has higher modeling accuracy and generalization ability, as well as better control performance. The model shows the following core advantages in dynamic modeling experiments:

[0042] 1. Extremely high modeling accuracy: Thanks to the explicit external force modeling, the two proposed models show excellent modeling performance in a variety of complex external force scenarios, with an average R 2 It reaches above 0.953. Compared with the DeLaN model, the modeling accuracy is improved by an average of 55.34%, and compared with the DeLaN-FFNN, the modeling accuracy is improved by 17.23%.

[0043] 2. Excellent generalization ability: Generalization experiments conducted under unseen external forces show that the generalization ability of the two proposed models is better than DeLaN and DeLaN-FFNN. In particular, as the external force increases, the degree of deterioration is significantly better than other advanced methods. When the external force reaches 20 times, R 2 It is above 0.85.

[0044] 3. Excellent control performance: Model-based inverse dynamics control experiments show that both proposed models can achieve good trajectory tracking under external disturbances. The RMSE of trajectory tracking for each joint is less than 0.8, which is better than other advanced methods and demonstrates superior control performance.

[0045] This innovative solution provides solid technical support for achieving real-time and precise control in the field of robotics. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0047] Figure 1 Schematic diagram of the network model of the present invention. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0049] In order to address the limitations of physical information neural networks in non-conservative force modeling and the problem of missing joint torque data in actual robot systems, this paper integrates motor dynamics and external force modeling into the deep Lagrangian network framework and proposes two dynamic models based on PINNs.

[0050] For robotic arms without joint torque sensors, the motor dynamics equations are embedded in the DeLaN network. By incorporating the mapping between motor current and joint torque into the training process, an explicit dynamic mapping relationship between the system motion state and motor current is finally established, achieving high-fidelity dynamic modeling without torque sensors.

[0051] In view of the limitations of DeLaN in modeling external forces, two methods of fitting the Jacobian matrix are proposed to introduce external force terms. The first method directly learns the mapping relationship between the end velocity and the joint velocity, which is called direct Jacobian. The second method first learns the kinematic behavior of the system and then obtains the Jacobian matrix by differential analysis of the learned forward kinematic model, which is called kinematic differential. The training process implicitly satisfies the kinematic geometric constraints and has stronger physical consistency and interpretability.

[0052] Multiple dynamic modeling experiments were conducted on the UR5 and UR10e six-degree-of-freedom robotic arms under various external force scenarios. The results demonstrate that the two proposed models surpass other advanced methods in modeling accuracy, generalization, and control performance. This validates the effectiveness of introducing motor dynamics and external force modeling, as well as the improved physical consistency and interpretability achieved by implicitly satisfying geometric constraints through learned kinematic behavior, which can further enhance model performance.

[0053] The dynamic modeling method based on physical information neural network proposed in this invention has both high real-time performance and high precision, is fully suitable for robot control, and can operate stably in complex external force environments, providing important technical support for the intelligent control of robots.

[0054] The embodiment of the present invention discloses a method for constructing a neural network for predicting robot dynamics physical information based on the coupling of a motor and an external force, comprising the following steps:

[0055] The motor dynamics equations are embedded into the DeLaN network, and an explicit dynamic mapping relationship between the system motion state and the motor current is established by mapping the motor current to the joint torque.

[0056] The external force term is introduced by fitting the Jacobian matrix, and the approximate mapping of the Jacobian matrix is ​​constructed using a data-driven method;

[0057] After constructing the initial model and incorporating the motor dynamics and external force modeling into the initial model, the inverse dynamics model of the robotic arm is obtained.

[0058] Specifically include the following:

[0059] Many robotic arms are not equipped with joint torque sensors and cannot obtain accurate torque data. Therefore, it is proposed to model the dynamics of the joint motor. Generally, the dynamics of a DC motor with a harmonic reducer is described by the following equation:

[0060] τ=rk t i;

[0061] Where τ is the joint torque, i is the motor current, k t is the motor torque constant, and r is the reduction ratio. The UR10e robot arm has a high reduction ratio and uses a harmonic drive. This results in low mechanical efficiency and a significant friction effect, making it important to select an appropriate friction model.

[0062] In the UR10e robot arm, the friction in the well-designed and maintained brushless motors is very low, so the friction in the system is mainly due to the harmonic drive. Here, we use the friction model used in the paper for the UR5e robot arm because it and the UR10e robot arm have similar motor structures. It assumes that the friction torque consists of the "standard" friction torque and the asymmetric smooth Coulomb friction torque. The friction torque on the jth motor side is described by the following formula:

[0063]

[0064] The constant ε is a smoothing parameter that is set to 10-3 based on experience. When power flows from the motor to the load, the Coulomb term T c The value of When the power direction is opposite, the Coulomb term T c The value of is the motor angular velocity, k v is the coefficient of friction.

[0065] Since the friction force of the UR10e robot arm is mainly generated by the harmonic drive, and the motor angular velocity is is the joint angular velocity, so the joint friction torque is:

[0066]

[0067] The motor actuator model is incorporated into the deep Lagrangian model, and the reduction ratio r and motor torque constant k t 、Coulomb term T c and friction coefficient k v As network weights are learned, end-to-end training is used to jointly optimize the model and parameters, ultimately establishing a dynamic mapping between the system's motion state and the motor current. Unlike traditional parameter identification, these parameters are not calibrated through offline experiments, but are dynamically adapted as part of the modeling framework.

[0068] When robots perform tasks such as grasping, assembly, and surgery, the end-point of the robot needs to come into contact with the object, generating external contact forces. Accurately modeling external contact forces is key to improving robot performance and safety. The formula for the external torque is as follows:

[0069] τ ext =J T (q)F ext ;

[0070] τ ext is the external torque, which is the torque acting on each joint of the robotic arm. extIt is the external force or torque of the end effector, which is measured by the six-dimensional force sensor at the end, including the three-dimensional external force (F x ,F y ,F z ) and three-dimensional external moment (M x ,M y ,M z ). J(q) is the Jacobian matrix of the robot arm, and the external force F acting on the end effector is converted to ext The moment τ converted to joint space ext Since robot manufacturers do not disclose various parameters of robots, the exact Jacobian matrix is ​​usually difficult to obtain. Therefore, a data-driven method is used to construct an approximate mapping of the Jacobian matrix, and two different methods are proposed to fit the Jacobian matrix.

[0071] (1) Direct Jacobian

[0072] The configuration of a robot consists of its rigid links and joint angles. Through forward kinematics, the joint angle q can be converted to the end effector's pose x, that is, x = f(q). The end pose x describes the spatial configuration of the robot's end effector in the task space, usually consisting of a position vector and attitude parameters. The forward kinematics model is obtained by chain multiplication of a series of homogeneous transformation matrices, representing the transformation from the base coordinate system to the end coordinate system. In order to analyze the relationship between the end effector velocity and the joint velocity, the Jacobian matrix is ​​introduced and defined as:

[0073]

[0074] in is the joint velocity, is the terminal velocity, which is defined as the Jacobian matrix that establishes the differential mapping relationship between the joint space and the task space. Therefore, the direct Jacobian method directly learns the mapping relationship between the terminal velocity and the joint velocity. The network input is the joint angle q, and the output is the Jacobian matrix J(q). The loss function is designed as follows:

[0075]

[0076] Where N represents the number of training samples, is the actual measured terminal velocity, is the terminal velocity predicted by the model, which is calculated as follows:

[0077]

[0078] in The Jacobian matrix represents the network's predictions. Neural Jacobi methods directly learn differential mappings between joint health and task space. Essentially, they use data to fit a local linearization of a chain structure and are unable to globally model nonlinear kinematics. If the data only covers a local workspace, large prediction errors may occur in untrained areas, and the kinematic and geometric constraints of the robot arm may be violated due to data noise and overfitting.

[0079] (2) Kinematic differentials

[0080] The kinematic differential method first learns the system's kinematic behavior and fits the forward kinematic model, which is the mapping relationship between joint angles and end-point poses x = f(q). This method can capture the global nonlinearity of kinematics and cover the entire workspace. Then, the partial derivatives of the forward kinematic model f(q) with respect to q are calculated to obtain the system's Jacobian matrix:

[0081]

[0082] The input of the network is the joint angle q, and the output is the end effector pose x. It is trained through supervised learning, and the loss function is designed to predict the end effector pose Compared with the actual end effector pose The mean square error between:

[0083]

[0084] The forward kinematics model is essentially the chain rule for the transformation of the robot arm's link coordinate system. In the process of learning the kinematic model, since the model itself is constrained by the robot's geometric structure (such as the orthogonality of the rotation matrix, the invariance of the link length, etc.), the network implicitly satisfies these geometric constraints during training. As the analytical derivative of f(q), each partial derivative in the Jacobian matrix element The geometric constraints of the robot arm link structure are inherited, making the learned Jacobian matrix highly physically consistent.

[0085] Each element in the Jacobian matrix has a clear kinematic interpretation, intuitively explaining the specific contribution of joint angle changes to the end effector's position and posture. It directly reflects the spatial coupling relationship between the manipulator's configuration parameters and possesses a high degree of physical interpretability. This method yields a Jacobian matrix with high physical consistency and interpretability, enabling seamless integration of external force modeling into PINN.

[0086] The specific network parameter structure of the model is described as follows:

[0087]

[0088] θ, φ, and ψ represent the weight parameters of the network that need to be trained to learn the mass matrix H(q), potential energy V(q), and Jacobian matrix J(q), respectively. The parameter k represents the parameters that need to be learned in the motor dynamics model, including the reduction ratio r and the motor torque constant k. t 、Coulomb term T c and friction coefficient k v The inverse dynamics model of the robotic arm can be expressed as:

[0089]

[0090] The training process not only learns the system dynamics, but also the Jacobian matrix. Therefore, in order to avoid the Jacobian matrix dominating the training process and ignoring the learning of system dynamics, an additional penalty term λ is introduced. The optimization objective of DeLaN-MKF is defined as follows:

[0091]

[0092] The optimization objective of DeLaN-MJF is defined as follows:

[0093]

[0094] The ultimate goal is to learn the dynamics of the system. Therefore, when optimizing the model, it is necessary to balance the loss of dynamics and Jacobian matrix as much as possible to ensure that the system dynamics can be learned to the maximum extent possible on the basis of learning the accurate Jacobian matrix.

[0095] After incorporating motor dynamics modeling and external force modeling into the model, the inverse dynamics model of the robotic arm is finally obtained:

[0096]

[0097] The network structures of DeLaN-MJF and DeLaN-MKF are as follows Figure 1 As shown, the network input is the robot arm joint angle q, joint angular velocity Joint angular acceleration and the terminal six-dimensional force F ext , the output is the motor current i.

[0098] The orange part represents the structure of DeLaN. The mass matrix H(q) and potential energy V(q) are parameterized to learn the conservative forces of the system. At the same time, the mass matrix H(q) is decomposed into a lower triangular matrix to ensure its symmetry and positive definiteness.

[0099] The blue portion represents external force modeling, mapping the task-space forces to the joint space via the Jacobian transpose. Two neural network approaches are used to fit the system Jacobian matrix. The DeLaN-MKF model, based on an analytical kinematic approach, first learns the system's forward kinematic model and then differentiates it to obtain the Jacobian matrix J(q). The DeLaN-MJF model, based on a direct Jacobian approach, directly learns the differential mapping between the joint space and task space to fit the Jacobian matrix J(q).

[0100] The purple portion represents motor dynamics modeling. By introducing a mapping relationship between joint torque τ and motor current i, a dynamic mapping relationship between the system's motion state and motor current is established, compensating for the lack of joint torque sensors. Since actual robotic arms have a significant amount of joint friction, joint friction modeling is also included.

[0101] The green part is the loss function of the two models, where the loss function of DeLaN-MKF is defined as: in represents the motor current predicted by the model. The loss function of DeLaN-MJF is defined as The loss functions of both models include the dynamics loss and the Jacobian loss of the entire model. λ is used to balance the proportion of the two in the total loss value.

[0102] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0103] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a neural network for predicting robot dynamics physical information based on the coupling of motors and external forces, characterized in that: The following steps are involved: The motor dynamics equations are embedded into the DeLaN network, and an explicit dynamic mapping relationship between the system motion state and the motor current is established by mapping the motor current to the joint torque. The external force term is introduced by fitting the Jacobian matrix, and the approximate mapping of the Jacobian matrix is ​​constructed using a data-driven method; After constructing the initial model and incorporating the motor dynamics and external force modeling into the initial model, the inverse dynamics model of the robotic arm is obtained.

2. The method for constructing a robot dynamics physical information prediction neural network based on motor and external force coupling according to claim 1, characterized in that: The expression of the inverse dynamics model of the robotic arm is as follows: The network input is the robot arm joint angle q, joint angular velocity Joint angular acceleration and the terminal six-dimensional force F ext , the output is the motor current i, where represents the matrix of centrifugal force and Coriolis force, g(q) represents the gravity vector, H(q) is the mass matrix, τ represents the generalized force acting on the generalized coordinate q of the system, i is the motor current, r is the reduction ratio, k t is the motor torque constant, τ F is the friction torque, τ ext For the external torque, the inverse dynamics model introduces the mapping relationship between the joint torque τ and the motor current i, and finally establishes the dynamic mapping relationship between the system motion state and the motor current.

3. The method for constructing a neural network for predicting robot dynamics physical information based on motor and external force coupling according to claim 1, characterized in that: The DC motor dynamics are described by the following equations: τ=rk t i; Where τ is the joint torque, i is the motor current, k t is the motor torque constant, and r is the reduction ratio.

4. The method for constructing a robot dynamics physical information prediction neural network based on motor and external force coupling according to claim 3, characterized in that: The joint friction torque is: Among them, the constant ε is a smoothing parameter set to 10-3 based on experience, r is the reduction ratio, k v is the coefficient of friction, is the joint angular velocity; when power flows from the motor to the load, the Coulomb term T c The value of When the power direction is opposite, the Coulomb term T c The value of 5. The method for constructing a robot dynamics physical information prediction neural network based on motor and external force coupling according to claim 1, characterized in that: The formula for the external force term is described as follows: τ ext =J T (q)F ext ; τ ext is the external torque; F ext is the external force or torque of the end effector, and J(q) is the Jacobian matrix of the manipulator.

6. The method for constructing a robot dynamics physical information prediction neural network based on motor and external force coupling according to claim 5, characterized in that: The Jacobian matrix is ​​fitted using direct Jacobian, as follows: The robot's configuration consists of its rigid links and joint angles. Through forward kinematics, the joint angles q are converted to the end-effector pose x, i.e., x = f(q). The end-effector pose x describes the spatial configuration of the robot's end-effector in the task space and consists of a position vector and attitude parameters. The forward kinematics model is obtained by chain multiplication of a series of homogeneous transformation matrices, representing the transformation from the base coordinate system to the end-effector coordinate system. To analyze the relationship between the end-effector velocity and the joint velocity, the Jacobian matrix is ​​introduced and defined as: in is the joint velocity, is the terminal velocity, which is defined as the Jacobian matrix to establish the differential mapping relationship between the joint space and the task space. Therefore, the direct Jacobian method directly learns the mapping relationship between the terminal velocity and the joint velocity. The network input is the joint angle q, and the output is the Jacobian matrix J(q). The loss function is designed as follows: Where N represents the number of training samples, is the actual measured terminal velocity, is the terminal velocity predicted by the model, which is calculated as follows: in, Represents the Jacobian matrix of the network predictions.

7. The method for constructing a neural network for predicting robot dynamics physical information based on motor and external force coupling according to claim 5, characterized in that: The Jacobian matrix is ​​obtained using the kinematic differentiation method as follows: First, the kinematic behavior of the system is learned, and the forward kinematic model, that is, the mapping relationship between joint angles and end-point poses x = f(q), is fitted to capture the global nonlinearity of the kinematics and cover the entire workspace; the Jacobian matrix of the system is obtained by calculating the partial derivatives of the forward kinematic model f(q) with respect to q: The input of the network is the joint angle q, and the output is the end effector pose x. It is trained through supervised learning, and the loss function is designed to predict the end effector pose Compared with the actual end effector pose The mean square error between:

8. The method for constructing a robot dynamics physical information prediction neural network based on motor and external force coupling according to claim 1, characterized in that: The network parameter structure of the initial model is described as follows: θ, φ, and ψ represent the weight parameters of the network that need to be trained to learn the mass matrix H(q), potential energy V(q), and Jacobian matrix J(q), respectively. The parameter k represents the parameters that need to be learned in the motor dynamics model, including the reduction ratio r and the motor torque constant k. t 、Coulomb term T c and friction coefficient k v .

Images