Robot path planning method for fiber flexible body operation

Abstract

The invention discloses a robot path planning method for fiber flexible body operation. Aiming at the problem of low efficiency of a traditional manual fiber coiling mode, the invention provides a robot path planning method for fiber flexible body operation, and a robot is adopted to replace manual fiber coiling, so that the production efficiency of an optical fiber sensor can be improved, and the manufacturing cost can be reduced. Through actual verification, the method disclosed by the invention can accurately control optical fiber coiling according to a designed path, realizes intelligent fiber coiling, and has very high application value and economic benefit.

Description

Technical Field

[0001] The invention relates to a flexible body shape control and path planning method in the optical fiber industry, belonging to the field of flexible body shape control and path planning. Background Art

[0002] Fiber optic sensors are the most important and widely used of all sensors. In the production process of fiber optic sensors, the optical path connection tasks include fiber stripping, fiber cleaning, fiber cutting, and fiber splicing, among which fiber coiling is the most time-consuming and energy-consuming link.

[0003] At present, the fiber winding process in the fiber connection task still uses the traditional manual fiber winding method, and its fiber winding efficiency can no longer meet industrial needs, which seriously hinders the market development of fiber optic sensors. With the expansion of the application field of robot technology and the change of actual needs, robots are now able to complete the control operation of flexible deformable objects. Using robots to replace manual fiber winding tasks is an extremely effective method, and its performance fully meets the actual requirements of industrial fiber winding. Summary of the invention

[0004] The purpose of the present invention is to provide a method for controlling an optical fiber to coil a fiber by a robot, so as to improve the production efficiency of optical fiber sensors and reduce the manufacturing cost.

[0005] In order to achieve the above object, the technical solution of the present invention is to provide a robot path planning method for fiber flexible body operation, characterized in that it includes the following steps:

[0006] Step 1: The end of the fiber is randomly moved by the gripper at the end of the robot arm to obtain the point cloud data of the fiber image captured by the 3D camera and the position and posture of the gripper at the end of the robot arm;

[0007] Step 2: Input the point cloud data of the optical fiber image obtained in real time into the trained mechanism and data fusion network model, and the mechanism and data fusion network model outputs the velocity information of each key point of the optical fiber at the current time t-1. Based on Get the position information X of each key point of the optical fiber at the next time t t , where X t is the position information of each key point of the optical fiber at the current time t-1, and Δt is the length of a time step;

[0008] Step 3: Set the pre-coil position near the fiber coil box as the target position X d , repeat step 2 to repeatedly solve the optimal robot motion using model predictive control and send the result to the robot for execution until ||X(t)-X d||≤e, so that the randomly placed optical fiber can be moved to the pre-coiled position, where e is the preset target threshold;

[0009] Step 4: Use a 3D camera to collect the point cloud data of the fiber cassette, and use color filtering to filter out the target buckle point cloud. Calculate the centroid of each filtered area as the representative position of each buckle. At the same time, use the principal component analysis method to analyze the variability of the point cloud data to find the main axis to determine the direction of different target buckles, and comprehensively consider the position and direction of the buckle to determine the order in which the optical fiber passes through the target buckle to generate the initial reference trajectory.

[0010] Step 5: Construct a virtual magnetic vector field ring at each target buckle position, dynamically adjust the clamp position and recalculate the rope motion path, and insert each buckle in turn according to the initial reference trajectory, which specifically includes the following steps:

[0011] Step 501: Control the holder to make the fiber tip pass through the first target buckle; set the virtual magnetic vector field to ensure that all integral curves pass through the buckle opening. The direction of the vector field is designed so that all field lines can enter the opening from one side and leave from the other side, ensuring that there is a closed solution form at any position in the workspace, thereby speeding up the calculation process. The magnetic field induced by the current-carrying ring that follows these characteristics is constructed as follows:

[0012]

[0013] Where p represents an arbitrary point in the workspace, S represents a fixed current-carrying ring, and F(p) represents the magnetic force at point p. The position g of the optical fiber grasped by the gripper is calculated using the relationship between the opening direction of the first target ring and the direction of gravity. The expected motion of the key point on the optical fiber is predicted by calculating the velocity of the optical fiber end through the relationship between the motion of the gripper and the key point encoded by the network model and the magnetic field equation of the target ring.

[0014] Step 502: If the difference between the expected motion and the actual motion If the tolerance value is exceeded, the gripper is readjusted to move closer to the critical point to ensure that the gripper moves along the expected path;

[0015] Step 503: After the clamp pushes the optical fiber through the first target buckle, it needs to be re-grasped; when calculating the grasping position, it is determined by calculating the angle θ between the normal line of the target ring and the direction of gravity, and the gripping point of the clamp is selected to be beyond the final key point distance l / θ; after the tip of the optical fiber penetrates the target virtual ring, it is selected to re-grasp or not to grasp the optical fiber. This re-grasping method is combined with the clamp, and the optical fiber is passed through a series of planned buckles in this way to realize intelligent fiber winding.

[0016] Preferably, in step 2, the mechanism and data fusion network model is built based on a graph neural network model, and the mechanism module is embedded in the graph neural network model.

[0017] Preferably, in step 2, training the mechanism and data fusion network model comprises the following steps:

[0018] Step 201: The point cloud data of the obtained optical fiber image is regarded as a sample from a Gaussian mixture model, whose centroid is the key point position X(t). According to the Bayesian theorem, sampling points are extracted from the Gaussian mixture model. Probability It is expressed as:

[0019]

[0020] Where p(n) represents the weight of the nth mixture component. represents sampling from the mixture component to The corresponding probability of

[0021] The problem is transformed into a maximum likelihood estimation problem, where the mixed centroid is optimized to maximize the point cloud data Y t The log-likelihood L of the sampling is:

[0022]

[0023] Where: σ represents the standard deviation of the covariance matrix of each component in the Gaussian mixture model; represents the centroid position of the nth mixed component at time t; S represents the total number of sample points in the point cloud data; p(n) represents the weight of the nth mixed component; N+1 represents the index of the component in the mixed model, where N is the number of Gaussian components of the real data point and N+1 is an additional component used to model noise and outliers;

[0024] The positions of key points of the collected point cloud data are calculated through Gaussian mixture model and Bayesian reasoning, and then matched with the obtained posture data of the end gripper of the robot arm to obtain the training image data set;

[0025] Step 202: The optical fiber mechanism model is constructed by using abquas software to obtain the relationship between the key points and the end clamp position, and the relationship data between the key points and the end clamp position is referenced to initialize the mechanism and data fusion network weights. The optical fiber mechanism model uses the Neo-Hookean model to represent the simplest form of strain potential energy. The energy function of this Neo-Hookean model is as follows:

[0026]

[0027] Where, U is the strain energy per unit reference volume; C 10 , D1 is the material parameter of the optical fiber; is the first deviatoric strain invariant, J el is the elastic volume ratio;

[0028] Step 203: The mechanism and data fusion network model is trained by the training image data set. During the training, it first passes through the encoding module. The encoder processes the information of the vertices and edges of the graph G and projects the graph G to a low-dimensional representation G = (V, E) to improve efficiency. Among them, the vertex V corresponds to the position sequence of the key points of the optical fiber, and the edge E corresponds to the interaction between the key points. The output of the encoder is the low-dimensional representation of the graph, that is, the encoded vertex features 0 v i ′ and edge features 0 e i ' ,j ,These outputs serve as inputs to the processor module;

[0029] The processor module consists of multiple message passing modules. The first message passing module receives the output from the encoder, i.e., the low-dimensional vertex features. 0 v i ′ and edge features 0 e i ' ,j ; Compute the interactions between vertices in the latent graph to simulate the dynamics of the fiber. The output after the last message passing module is the updated latent vertex features. K+1 v i ′ and edge features K+ 1 e i ' ,j , as the input of the decoder; finally, the decoder module restores the obtained potential representation to the original prediction of the fiber motion, and the output is the velocity information of each key point of the fiber at a specific moment Among them, based on the current graph state and the control input of the robot arm, the process of propagating graph vertex and edge features in the processor module is expressed as follows:

[0030]

[0031] k+1 e i ' ,j =f e ( k e i ' ,j , k v i ′, k v j ′)

[0032] Where: fv (·) and f e (·) represent the graph vertex network and graph edge network respectively; k v i ′, k e i ' ,j Represents the current vertex and edge features; k+1 v i ′, k+1 e i ' ,j represents the potential vertex and edge features after the kth message passing; represents the sum of the features of the edges connected to vertex i; k v i 'and k v′ j are the features of the two vertices i and j connected by the edge (i, j).

[0033] Preferably, step 4 comprises the following steps:

[0034] Step 401: Decentralize the data set D consisting of point cloud data, as shown in the following formula:

[0035]

[0036] Where: m is the number of data, X = [x1, x2, ..., x m ], each x is an n-dimensional column vector.

[0037] Step 402: Calculate the covariance matrix C of the data set D, as shown in the following formula:

[0038]

[0039] Find a direction in the above data space so that the variance of the projection of the data in this direction is the largest, where w is the direction with the largest projection variance, w T x i The coordinate value of each data in this direction. is the covariance matrix of the sample, represented by C, then the optimization objective is as follows:

[0040]

[0041] Step 403: Perform eigenvalue decomposition on the covariance matrix C. The obtained eigenvectors indicate the main directions of the data, and the eigenvalues ​​indicate the degree of variation along each direction. The Lagrange constant method is used to solve the problem, and we have:

[0042] L(w,λ)=w T Cw+λ(1-w T w)

[0043]

[0044] The solution is:

[0045]

[0046] maxD(x)=max{w T Cw}=max{w T λw}=maxλ

[0047] Where: w is the eigenvector of C;

[0048] Find the maximum variance, which is the maximum eigenvalue of the covariance matrix. The direction at this time is the eigenvector corresponding to the maximum eigenvalue. Select the first few eigenvectors with the largest eigenvalue. These eigenvectors point to the main direction of change of the data, which is also the main direction of the fixed device.

[0049] In view of the low efficiency of the traditional manual fiber coiling method, the present invention proposes a robot path planning method for fiber flexible body operation, which uses a robot to replace manual fiber coiling, which can improve the production efficiency of optical fiber sensors and reduce manufacturing costs. Through actual verification, the method disclosed by the present invention can accurately control the fiber coiling according to the designed path, realize intelligent fiber coiling, and has high application value and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is the algorithm flow chart of the present invention;

[0051] Figure 2 The data and mechanism fusion network model in the fiber coil path planning method of the present invention is illustrated;

[0052] Figure 3 It is a virtual magnetic vector field construction diagram in the fiber coil path planning method of the present invention. DETAILED DESCRIPTION

[0053] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0054] like Figure 1 As shown, a robot path planning method for fiber flexible body operation disclosed in an embodiment of the present invention specifically includes the following steps:

[0055] Step 1: The end of the fiber is clamped by the gripper at the end of the robot arm and moved randomly to obtain the point cloud data of the fiber image captured by the 3D camera and the position and posture of the gripper at the end of the robot arm.

[0056] Step 2: Input the point cloud data of the optical fiber image obtained in real time into the trained mechanism and data fusion network model, and the mechanism and data fusion network model outputs the shape prediction of the optical fiber.

[0057] Combination Figure 2 The mechanism and data fusion network model is built based on the graph neural network model (GNN). The mechanism module is embedded in the backbone network of GNN. The training of the mechanism and data fusion network model includes the following steps:

[0058] Step 201: The obtained point cloud data is regarded as samples from a Gaussian mixture model (GMM), whose centroid is the key point position X(t). According to Bayes' theorem, sampling points are extracted from the Gaussian mixture model Probability It can be expressed as:

[0059]

[0060] Where p(n) represents the weight of the nth mixture component. represents sampling from the mixture component to The corresponding probability of .

[0061] The problem is then transformed into a maximum likelihood estimation (MLE) problem, where we want to optimize the mixture centroid so as to maximize the point cloud data Y t The log-likelihood L of the sampling is:

[0062]

[0063] Where: σ represents the standard deviation of the covariance matrix of each component in the Gaussian mixture model; represents the centroid position of the nth mixture component (or Gaussian component) at time t, which is the key point position we hope to optimize through maximum likelihood estimation; S represents the total number of sample points in the point cloud data, that is, the number of all samples we consider when performing probability calculations; p(n) represents the weight of the nth mixture component; N+1 represents the index of the component in the mixture model, where N is the number of Gaussian components of the real data point, and N+1 is an additional component used to model noise and outliers.

[0064] Through the Gaussian mixture model and Bayesian reasoning, the positions of the key points of the collected point cloud data are calculated, and then matched with the obtained posture data of the end gripper of the robotic arm to obtain the training image data set.

[0065] Step 202: The optical fiber mechanism model is constructed by using abquas software to obtain the relationship between the key points and the end clamp position, and the relationship data between the key points and the end clamp position is referenced to initialize the mechanism and data fusion network weights. The optical fiber mechanism model uses the Neo-Hookean model to represent the simplest form of strain potential energy. The energy function of this Neo-Hookean model is as follows:

[0066]

[0067] Where, U is the strain energy per unit reference volume; C 10 , D1 is the material parameter of the optical fiber; is the first deviatoric strain invariant, J el is the elastic volume ratio.

[0068] Step 203: The mechanism and data fusion network model is trained by the training image data set. During the training, it first passes through the encoding module. The encoder processes the vertex and edge information of the graph G. The graph G is projected to a low-dimensional representation G = (V, E) to improve efficiency, where the vertex V corresponds to the position sequence of the key points of the optical fiber, and the edge E corresponds to the interaction between the key points. The output of the encoder is a low-dimensional representation of the graph, that is, the encoded vertex features 0 v i ′ and edge features 0 e i ' ,j These outputs will serve as input to the first message passing module.

[0069] Then through the processor module (composed of multiple message passing modules), the first message passing module receives the output from the encoder, i.e., the low-dimensional vertex features 0 v i ′ and edge features 0 e i ' ,j The interactions between vertices in the latent graph are calculated to simulate the dynamics of the fiber. The output after the last message passing module is the updated latent vertex features. K+1 v i ′ and edge features K+1 e i ' ,j , which will serve as the input of the decoder. Finally, the decoder module will restore the latent representation to the original prediction of the fiber motion, and the output is the velocity information of each key point of the fiber at a specific moment. This is the target output of the entire model. Thus, the training of the mechanism and data fusion network model is completed. Among them, based on the current graph state and the control input of the robot arm, the process of propagating graph vertex and edge features in the processor module is expressed as follows:

[0070]

[0071] k+1 e i ' ,j =f e ( k e i ' ,j , k v i ′, k v j ′)

[0072] Where: f v (·) and f e (·) represent the graph vertex network and graph edge network respectively; k v i ′, k e i ' ,j Represents the current vertex and edge features; k+1 v i ′, k+1 e i ' ,j represents the potential vertex and edge features after the kth message passing; represents the sum of the features of the edges connected to vertex i; k v i 'and k v′ j are the features of the two vertices i and j connected by the edge (i, j).

[0073] Step 204: Substitute the obtained speed of each key point into the formula To obtain the key point position of the optical fiber in the next time step, we set the pre-coiled fiber position close to the fiber coil box as our target position X d , use model predictive control (MPC) to repeatedly solve the optimal robot motion and send the result to the robot to execute until ||X(t)-X d ||≤e. Now it is possible to move the randomly placed optical fiber to the pre-coiled position.

[0074] Step 3: Use a 3D camera to collect the point cloud data of the fiber cassette, and use color filtering to filter out the target buckle point cloud. Calculate the centroid of each filtered area as the representative position of each buckle. At the same time, use the principal component analysis (PCA) to analyze the variability of the point cloud data to find the main axis to determine the direction of different target buckles. And by comprehensively considering the position and direction of the buckle. So as to determine the order of the optical fiber passing through the target buckle to generate the initial reference trajectory. The algorithm steps are as follows:

[0075] Step 301: Decentralize the data set D, as shown in the following formula:

[0076]

[0077] Where: m is the number of data, X = [x1, x2, ..., x m ], each x is an n-dimensional column vector;

[0078] Step 302: Calculate the covariance matrix C of the data set D, as shown in the following formula:

[0079]

[0080] Our goal is to find a direction in this data space that maximizes the variance of the projection of the data in this direction. Where: w is the direction with the largest projection variance, w T x i The coordinate value of each data in this direction. is the covariance matrix of the sample, which we denote by C. Our optimization goal is:

[0081]

[0082] Step 303: Perform eigenvalue decomposition on the covariance matrix C. The obtained eigenvectors indicate the main directions of the data, while the eigenvalues ​​indicate the degree of variation along each direction. We use the Lagrange constant method to solve.

[0083] L(w,λ)=w T Cw+λ(1-w T w)

[0084]

[0085] The solution is:

[0086]

[0087] maxD(x)=max{w T Cw}=max{w T λw}=maxλ

[0088] Where w is the eigenvector of C. Find the maximum variance, that is, the maximum eigenvalue of the covariance matrix, and the direction at this time is the eigenvector corresponding to the maximum eigenvalue. Select the first few eigenvectors with the largest eigenvalues, which point to the main direction of change of the data, that is, the main direction of the fixture.

[0089] Step 4: Construct a virtual magnetic vector field ring at each target buckle position, and insert each buckle in sequence according to the initial reference trajectory by dynamically adjusting the clamp position and recalculating the rope motion path. Step 4 specifically includes the following steps:

[0090] Step 401: Control the holder to make the fiber tip pass through the first target buckle. We set up the virtual magnetic vector field to ensure that all integral curves pass through the buckle opening, so that the fiber can be easily inserted into the buckle opening regardless of the initial configuration. The direction of the vector field is designed so that all field lines can enter the opening from one side and leave from the other side, ensuring that there is a closed solution form at any position in the workspace, thereby speeding up the calculation process. The magnetic field induced by the current-carrying ring that follows these characteristics is constructed as follows:

[0091]

[0092] Where p represents an arbitrary point in the workspace, S represents a fixed current-carrying ring, and F(p) represents the magnetic force at point p. The position g of the fiber grasped by the gripper is calculated using the relationship between the opening direction of the first target ring and the direction of gravity. The relationship between the gripper and the key point motion encoded by the network model and the magnetic field equation of the target ring are used to calculate the fiber end velocity to predict the expected motion of the key point on the fiber.

[0093] Step 402: Considering that the expected movement of the gripper may result in a collision, if the difference between the expected movement and the actual movement If the tolerance value is exceeded, the gripper needs to be readjusted to move closer to the critical point to ensure that the gripper moves along the expected path.

[0094] Step 403: Because the clamp cannot pass through the buckle, it is necessary to re-grasp the optical fiber after the clamp pushes it through the first target buckle. When calculating the grasping position, it is determined by calculating the angle θ between the normal line of the target ring and the direction of gravity. The gripping point of the clamp is selected to exceed the distance l / θ of the final key point (l is a constant that can be determined experimentally). After the tip of the optical fiber penetrates the target virtual ring, you can choose to re-grasp or not grasp the optical fiber, depending on the situation. This re-grasping method is combined with the clamp to pass the optical fiber through a series of planned buckles in this way to achieve intelligent fiber coiling.

Claims

1. A robot path planning method for fiber flexible body operations, It is characterized in that The following steps are involved: Step 1: The end of the fiber is randomly moved by the gripper at the end of the robot arm to obtain the point cloud data of the fiber image captured by the 3D camera and the position and posture of the gripper at the end of the robot arm; Step 2: Input the point cloud data of the optical fiber image obtained in real time into the trained mechanism and data fusion network model, and the mechanism and data fusion network model outputs the velocity information of each key point of the optical fiber at the current time t-1. Based on Get the position information X of each key point of the optical fiber at the next time t t , where X t is the position information of each key point of the optical fiber at the current time t-1, and Δt is the length of a time step; Step 3: Set the pre-coil position near the fiber coil box as the target position X d , repeat step 2 to repeatedly solve the optimal robot motion using model predictive control and send the result to the robot for execution until ||X(t)-X d ‖≤e, so that the randomly placed optical fiber can be moved to the pre-coiled position, where e is the preset target threshold; Step 4: Use a 3D camera to collect the point cloud data of the fiber cassette, and use color filtering to filter out the target buckle point cloud. Calculate the centroid of each filtered area as the representative position of each buckle. At the same time, use the principal component analysis method to analyze the variability of the point cloud data to find the main axis to determine the direction of different target buckles, and comprehensively consider the position and direction of the buckle to determine the order in which the optical fiber passes through the target buckle to generate the initial reference trajectory. Step 5: Construct a virtual magnetic vector field ring at each target buckle position, dynamically adjust the clamp position and recalculate the rope motion path, and insert each buckle in turn according to the initial reference trajectory, which specifically includes the following steps: Step 501: Control the holder to make the fiber tip pass through the first target buckle; set the virtual magnetic vector field to ensure that all integral curves pass through the buckle opening. The direction of the vector field is designed so that all field lines can enter the opening from one side and leave from the other side, ensuring that there is a closed solution form at any position in the workspace, thereby speeding up the calculation process. The magnetic field induced by the current-carrying ring that follows these characteristics is constructed as follows: Where p represents an arbitrary point in the workspace, S represents a fixed current-carrying ring, and F(p) represents the magnetic force at point p. The position g of the optical fiber grasped by the gripper is calculated using the relationship between the opening direction of the first target ring and the direction of gravity. The expected motion of the key point on the optical fiber is predicted by calculating the velocity of the optical fiber end through the relationship between the motion of the gripper and the key point encoded by the network model and the magnetic field equation of the target ring. Step 502: If the difference between the expected motion and the actual motion If the tolerance value is exceeded, the gripper is readjusted to move closer to the critical point to ensure that the gripper moves along the expected path; Step 503: After the clamp pushes the optical fiber through the first target buckle, it needs to be re-grasped; when calculating the grasping position, it is determined by calculating the angle θ between the normal line of the target ring and the direction of gravity, and the gripping point of the clamp is selected to be beyond the final key point distance l / θ; after the tip of the optical fiber penetrates the target virtual ring, it is selected to re-grasp or not to grasp the optical fiber. This re-grasping method is combined with the clamp, and the optical fiber is passed through a series of planned buckles in this way to realize intelligent fiber winding.

2. A robot path planning method for fiber flexible body operation as claimed in claim 1, It is characterized in that In step 2, the mechanism and data fusion network model is built based on the graph neural network model, and the mechanism module is embedded in the graph neural network model.

3. A robot path planning method for fiber flexible body operation as claimed in claim 2, It is characterized in that In step 2, training the mechanism and data fusion network model includes the following steps: Step 201: The point cloud data of the obtained optical fiber image is regarded as a sample from a Gaussian mixture model, whose centroid is the key point position X(t). According to the Bayesian theorem, sampling points are extracted from the Gaussian mixture model. Probability It is expressed as: Where p(n) represents the weight of the nth mixture component, represents sampling from the mixture component to The corresponding probability of The problem is transformed into a maximum likelihood estimation problem, where the mixed centroid is optimized to maximize the point cloud data Y t The log-likelihood L of the sampling is: Where: σ represents the standard deviation of the covariance matrix of each component in the Gaussian mixture model; represents the centroid position of the nth mixed component at time t; S represents the total number of sample points in the point cloud data; p(n) represents the weight of the nth mixed component; N+1 represents the index of the component in the mixed model, where N is the number of Gaussian components of the real data point and N+1 is an additional component used to model noise and outliers; The positions of key points of the collected point cloud data are calculated through Gaussian mixture model and Bayesian reasoning, and then matched with the obtained posture data of the end gripper of the robot arm to obtain the training image data set; Step 202: The optical fiber mechanism model is constructed by using abquas software to obtain the relationship between the key points and the end clamp position, and the relationship data between the key points and the end clamp position is referenced to initialize the mechanism and data fusion network weights. The optical fiber mechanism model uses the Neo-Hookean model to represent the simplest form of strain potential energy. The energy function of this Neo-Hookean model is as follows: Where, U is the strain energy per unit reference volume; C 10 , D 1 is the material parameter of the optical fiber; is the first deviatoric strain invariant, J el is the elastic volume ratio; Step 203: The mechanism and data fusion network model is trained by the training image data set. During the training, it first passes through the encoding module. The encoder processes the information of the vertices and edges of the graph G and projects the graph G to a low-dimensional representation G = (V, E) to improve efficiency. Among them, the vertex V corresponds to the position sequence of the key points of the optical fiber, and the edge E corresponds to the interaction between the key points. The output of the encoder is the low-dimensional representation of the graph, that is, the encoded vertex features 0 v′ i and edge features 0 e′ i,j ,These outputs serve as inputs to the processor module; The processor module consists of multiple message passing modules. The first message passing module receives the output from the encoder, i.e., the low-dimensional vertex features. 0 v′ i and edge features 0 e′ i,j ; Compute the interactions between vertices in the latent graph to simulate the dynamics of the fiber. The output after the last message passing module is the updated latent vertex features. K+1 v′ i and edge features K+1 e′ i,j , as the input of the decoder; finally, the decoder module restores the obtained potential representation to the original prediction of the fiber motion, and the output is the velocity information of each key point of the fiber at a specific moment Among them, based on the current graph state and the control input of the robot arm, the process of propagating graph vertex and edge features in the processor module is expressed as follows: k+1 yes i ′ ,j =f e ( k yes i ′ ,j , k v i ′, k v j ′) Where: f v (·) and f e (·) represent the graph vertex network and graph edge network respectively; k v i ′, k e i ' ,j Represents the current vertex and edge features; k+1 v′ i , k+1 e′ i,j represents the potential vertex and edge features after the kth message passing; represents the sum of the features of the edges connected to vertex i; k v′ i and k v′ j are the features of the two vertices i and j connected by the edge (i, j).

4. A robot path planning method for fiber flexible body operation as claimed in claim 1, It is characterized in that The step 4 comprises the following steps: Step 401: Decentralize the data set D consisting of point cloud data, as shown in the following formula: Where: m is the number of data, X = [x 1 ,x 2 ,…,x m ], each x is an n-dimensional column vector. Step 402: Calculate the covariance matrix C of the data set D, as shown in the following formula: Find a direction in the above data space so that the variance of the projection of the data in this direction is the largest, where w is the direction with the largest projection variance, w T x i The coordinate value of each data in this direction. is the covariance matrix of the sample, represented by C, then the optimization objective is as follows: Step 403: Perform eigenvalue decomposition on the covariance matrix C. The obtained eigenvectors indicate the main directions of the data, and the eigenvalues ​​indicate the degree of variation along each direction. The Lagrange constant method is used to solve the problem, and we have: L(w,λ)=w T Cw+λ(1-w T In) The solution is: maxD(x)=max{w T Cw}=max{w T λw}=maxλ Where: w is the eigenvector of C; Find the maximum variance, which is the maximum eigenvalue of the covariance matrix. The direction at this time is the eigenvector corresponding to the maximum eigenvalue. Select the first few eigenvectors with the largest eigenvalue. These eigenvectors point to the main direction of change of the data, which is also the main direction of the fixed device.

Images