A path planning method and system of a prefabricated component demolding transfer robot

Abstract

The application provides a prefabricated component demolding transfer robot path planning method and system, relates to the technical field of robot path planning, and the method comprises the following steps: acquiring electrostatic signal data, and performing inverse finite element analysis and manifold learning dimension reduction to generate gradient distribution data and geometric contour data; then the gradient distribution data is subjected to smoothing treatment to generate a first repulsion potential field; then a second repulsion potential field is constructed on the component surface region according to three-dimensional laser point clouds; then an attractive potential field is generated based on the three-dimensional pose information of the prefabricated component and the geometric constraint of the grabbing path, and the three potential fields are adaptively weighted and superimposed to construct a composite artificial potential field; finally, combined with a hybrid A* algorithm and a Reeds-Shepp curve model, node connection and trajectory feasibility checking are performed to complete state node evaluation and exploration in the composite artificial potential field, so as to generate trajectory planning data. The application can realize undisturbed accurate grabbing path planning.

Description

Technical Field

[0001] This application relates to the technical field of robot path planning, and in particular to a path planning method and system for a robot for demolding and transferring prefabricated components. Background Technology

[0002] In the field of automated production of precast components, path planning technology in the demolding and transfer process directly affects the finished product quality and transfer efficiency of the components. For dry components with fine substances adhering to their surfaces, non-contact adaptive gripping path planning can be achieved, which has application value for improving the level of production automation.

[0003] Currently, the demolding and transfer of precast components mainly adopts teaching programming or predefined path planning methods, and the transfer operation of components is achieved by manually guiding the robot to repeatedly execute fixed trajectories or by pre-setting the grasping path based on the 3D model.

[0004] However, existing methods rely on preset trajectories or geometric features, making it difficult to effectively perceive the state of subtle attachments on the surface of components. Furthermore, during the grasping process, fixed-pattern path planning cannot dynamically adjust the contact position and approach trajectory according to the surface state, which can easily lead to local attachment disturbances and affect the surface quality of components and grasping stability. Summary of the Invention

[0005] The purpose of this application is to provide a path planning method and system for a precast component demolding and transfer robot, so as to solve the problem of adaptive path planning in the prior art that is difficult to take into account both surface state perception and environmental constraints.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a path planning method for a precast component demolding and transfer robot, comprising:

[0007] Electrostatic signal data of the surface of the precast component is acquired, and gradient distribution data and geometric contour data of the precast component are generated by performing inverse finite element analysis and manifold learning dimensionality reduction on the electrostatic signal data. The gradient distribution data is used to characterize the dust risk distribution on the surface of the precast component.

[0008] The gradient distribution data is smoothed by Gaussian convolution to generate a first repulsive potential field.

[0009] Based on the static obstacle information represented by the three-dimensional laser point cloud data of the current production scene, a second repulsive potential field based on the Euclidean distance of the obstacle is constructed on the surface region of the component corresponding to the geometric contour data.

[0010] Based on the three-dimensional pose information of the prefabricated components, combined with the geometric constraints of the preset grasping path, an attractive potential field pointing towards the target point is generated.

[0011] A composite artificial potential field is constructed by adaptively weighting and superimposing the first repulsive potential field, the second repulsive potential field, and the attractive potential field.

[0012] By combining the hybrid A* algorithm and the Reeds-Shepp curve model, node connections and trajectory feasibility verification are performed to evaluate and explore state nodes in the composite artificial potential field, thereby generating trajectory planning data for the robot.

[0013] Optionally, the step of performing inverse finite element analysis and manifold learning dimensionality reduction on the electrostatic signal data to generate gradient distribution data and geometric contour data of the precast component includes:

[0014] Using the electrostatic signal data as boundary constraints, an electric potential field control equation is constructed. The electric potential field control equation is solved through a finite element analysis model to obtain the electric field intensity data on the surface of the prefabricated component.

[0015] The electric field intensity data is sampled in a neighborhood to obtain multiple electric field intensity vectors. All the electric field intensity vectors are then aggregated to obtain a high-dimensional feature set.

[0016] Based on the high-dimensional feature set, an adjacency graph is constructed using an equidistant feature mapping algorithm, and the shortest path distance between data points in the adjacency graph is calculated to obtain a geodesic distance matrix.

[0017] Multidimensional scaling analysis is performed on the geodesic distance matrix to obtain low-dimensional embedded coordinates;

[0018] Based on the low-dimensional embedded coordinates, the geometric parameter representation of the surface of the precast component is reconstructed, and the gradient field at each data point is calculated based on the geometric parameter representation to obtain gradient distribution data.

[0019] Delaunay triangulation is performed on the low-dimensional embedded coordinates to obtain the geometric contour data of the prefabricated component.

[0020] Optionally, the step of performing multidimensional scaling analysis on the geodesic distance matrix to obtain low-dimensional embedded coordinates includes:

[0021] Perform multidimensional scaling analysis on the geodesic distance matrix to obtain the first embedded coordinates, and calculate the dust risk coefficient for each data point based on the electric field intensity vector in the high-dimensional feature set;

[0022] The dust risk coefficient is used as a node attribute and together with the adjacency graph to construct a weighted adjacency graph. The spectral clustering method is then applied to perform spectral clustering on the weighted adjacency graph to obtain multiple clustering regions with different risk levels.

[0023] Using the clustered regions with different risk levels as category labels, Fisher discriminant analysis is performed on the first embedded coordinates to obtain the discriminant coordinates;

[0024] The dust risk coefficient of each data point is fused with the discrimination coordinates at the feature level to obtain a fused feature vector;

[0025] The fused feature vector is calculated using a local preservation projection algorithm to obtain low-dimensional embedded data.

[0026] Optionally, the step of performing Gaussian convolution smoothing on the gradient distribution data to generate a first repulsive potential field includes:

[0027] The gradient distribution data is mapped to a two-dimensional parametric grid corresponding to the surface of the precast component to obtain an initial gradient field, and the surface curvature information of each grid point on the two-dimensional parametric grid is extracted to obtain a curvature distribution field.

[0028] Calculate the structure tensor of each grid point in the initial gradient field, and determine the principal direction and degree of anisotropy at each grid point based on the structure tensor;

[0029] Based on the principal direction and the degree of anisotropy, the initial gradient field is convolved and smoothed using a Gaussian convolution kernel to obtain the target gradient field. The magnitude of the target gradient field is then calculated to obtain the risk intensity field.

[0030] Based on the curvature distribution field, the risk intensity field is modulated using a nonlinear transformation function to obtain the modulated risk field;

[0031] The modulated risk field is transformed using an exponential potential function to obtain the first repulsive potential field.

[0032] Optionally, the combination of the hybrid A* algorithm and the Reeds-Shepp curve model for node connection and trajectory feasibility verification is used to evaluate and explore state nodes in the composite artificial potential field, generating trajectory planning data for the robot, including:

[0033] An open list is constructed based on the starting node, and multiple motion primitives are generated according to the robot's nonholonomic kinematic constraints and dynamic characteristics.

[0034] Perform kinematic integration on each of the motion primitives to obtain multiple next nodes corresponding to the starting node, and calculate the segment cost from the starting node to each of the next nodes in the composite artificial potential field;

[0035] Based on the road segment cost value, the actual cost value corresponding to the next node is obtained, and the estimated cost value from the next node to the target node is calculated. The actual cost value and the estimated cost value are added together to obtain the comprehensive cost value.

[0036] The Reeds-Shepp curve model is used to calculate the shortest curve from the next node to the target node, and the curvature continuity of the shortest curve is verified in the composite artificial potential field.

[0037] If the verification fails, the next node, along with its corresponding actual value, estimated value, and parent node pointer, is added to the open list. In the next round, the node with the smallest comprehensive value in the open list is selected as the starting node, and the process of calculating the value to expand the open list is repeated until the verification passes.

[0038] If the verification is successful, the next node is taken as the connection node, and the search path of the connection node in the open list is traced back in reverse according to the parent node pointer to obtain the node sequence.

[0039] The trajectory segments corresponding to all the motion primitives are sequentially spliced ​​together with the shortest curve that has passed the verification along the node sequence to generate the trajectory planning data of the robot.

[0040] Optionally, constructing a composite artificial potential field by adaptively weighting and superimposing the first repulsive potential field, the second repulsive potential field, and the attractive potential field includes:

[0041] Discretize the first repulsive potential field, the second repulsive potential field and the attractive potential field to the same spatial grid coordinate system to obtain the first repulsive potential field value, the second repulsive potential field value and the attractive potential field value corresponding to each grid point.

[0042] The first repulsive potential field value, the second repulsive potential field value, and the attractive potential field value corresponding to each grid point are input into the multilayer perceptron network. The forward propagation calculation is performed through the multilayer perceptron network to output the first repulsive weight, the second repulsive weight, and the attractive weight of the corresponding grid point.

[0043] Multiply the first repulsive potential field value by the first repulsive weight to obtain the weighted first repulsive field value; multiply the second repulsive potential field value by the second repulsive weight to obtain the weighted second repulsive field value; and multiply the attractive potential field value by the attractive weight to obtain the weighted attractive potential field value.

[0044] The weighted first repulsive field value, the weighted second repulsive field value, and the weighted attractive field value at the same grid point are algebraically summed to generate the composite potential field value of the corresponding grid point.

[0045] The composite potential field value of all grid points is smoothed to obtain the composite artificial potential field.

[0046] Optionally, the generation of an attractive potential field pointing towards the target point based on the three-dimensional pose information of the prefabricated component, combined with the geometric constraints of the preset grasping path, includes:

[0047] The three-dimensional pose information of the prefabricated component and the relative pose relationship in space between the tool center point and the feature point coordinates of the robot end effector are obtained.

[0048] Using the environmental coordinate system defined by the three-dimensional pose information as a reference, and combining the relative pose relationship, the kinematic model of the robot, and the geometric constraints of the grasping path, the path from the center point of the tool to the coordinates of the feature point is sampled to generate multiple candidate grasping paths.

[0049] Calculate the smoothness index and execution time estimate for each candidate crawling path, and determine the attraction intensity value for each candidate crawling path based on the smoothness index and execution time estimate;

[0050] Centered on the coordinates of the feature point, the attraction intensity values ​​are diffused along the corresponding spatial direction to generate a vector direction field, and the vector direction field is converted into an attraction potential field.

[0051] Secondly, this application provides a path planning system for a precast component demolding and transfer robot, comprising:

[0052] The acquisition module is used to acquire electrostatic signal data on the surface of the precast component, and generate gradient distribution data and geometric contour data of the precast component by performing inverse finite element analysis and manifold learning dimensionality reduction on the electrostatic signal data. The gradient distribution data is used to characterize the dust risk distribution on the surface of the precast component.

[0053] The processing module is used to perform Gaussian convolution smoothing on the gradient distribution data to generate a first repulsive potential field.

[0054] The first construction module is used to construct a second repulsive potential field based on the Euclidean distance of the obstacle in the component surface area corresponding to the geometric contour data, based on the static obstacle information represented by the three-dimensional laser point cloud data of the current production scene.

[0055] The generation module is used to generate an attractive potential field pointing to the target point based on the three-dimensional pose information of the prefabricated component and the geometric constraints of the preset grasping path.

[0056] The second construction module is used to construct a composite artificial potential field by adaptively weighting the first repulsive potential field, the second repulsive potential field and the attractive potential field.

[0057] The verification module is used to combine the hybrid A* algorithm and the Reeds-Shepp curve model to verify the node connection and trajectory feasibility, so as to complete the evaluation and exploration of state nodes in the composite artificial potential field and generate trajectory planning data for the robot.

[0058] Thirdly, this application provides an electronic device, comprising:

[0059] Memory, used to store computer programs;

[0060] A processor is configured to execute the computer program to implement the steps of a path planning method for a prefabricated component demolding and transfer robot as described in the first aspect above.

[0061] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of a path planning method for a prefabricated component demolding and transfer robot as described in the first aspect above.

[0062] The path planning method for a prefabricated component demolding and transfer robot provided in this application has the following beneficial effects: This application realizes non-contact perception of dust distribution and geometric contour on the component surface through electrostatic signal data analysis and processing, which can provide accurate environmental information for path planning; then, the dust risk distribution is smoothed to form a first avoidance field for dust disturbance, and a second avoidance field based on obstacle distance is constructed by combining laser point cloud data; then, a gravitational field pointing to the target position is generated according to the component pose and grasping constraints; then, multiple constraints of dust avoidance, obstacle avoidance and target attraction are dynamically balanced through an adaptive weighted superposition method; finally, the hybrid A* algorithm and Reeds-Shepp curve model are integrated to verify the trajectory feasibility, thereby generating a smooth path that meets kinematic constraints and is executable. This method realizes adaptive path planning based on surface state perception and environmental constraints.

[0063] Furthermore, this application uses electrostatic signal data as boundary constraints and solves the electric field intensity distribution on the component surface through inverse finite element analysis. Then, it uses a manifold learning algorithm to reduce the dimensionality of the high-dimensional electric field intensity features to reconstruct the geometric parameter representation and gradient distribution data of the component surface. At the same time, it performs triangulation on the dimensionality-reduced low-dimensional embedded coordinates to accurately restore the geometric contour of the component. This mechanism realizes the accurate reconstruction from physical signals to geometric and risk information, thus laying a reliable data foundation for subsequent potential field construction and path planning. Attached Figure Description

[0064] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 A flowchart illustrating a path planning method for a precast component demolding and transfer robot provided in this application embodiment;

[0066] Figure 2 A schematic diagram illustrating a path planning method for a prefabricated component demolding and transfer robot provided in this application embodiment;

[0067] Figure 3 A schematic diagram of the path planning system for a prefabricated component demolding and transfer robot provided in this application embodiment;

[0068] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0069] Currently, the path planning for demolding and transporting precast components mostly relies on preset trajectories or geometric models, and it is difficult to perceive the distribution of fine dust on the surface of the components. This makes it easy for dust disturbance to occur during the grasping process due to improper contact position, which in turn affects the surface quality of the components and the stability of the operation.

[0070] To address this, this application introduces a path planning method for a prefabricated component demolding and transfer robot. The core of this method lies in: reconstructing the dust distribution and geometric contours of the component surface by analyzing electrostatic signals; constructing dust risk and obstacle information as two types of avoidance fields respectively; and then adaptively fusing them with the target gravitational field; furthermore, combining a hybrid A* algorithm and Reeds-Shepp curves to verify trajectory feasibility. This method not only achieves accurate non-contact surface state capture but also dynamically balances multiple constraints to generate a smooth path, thus fundamentally overcoming the shortcomings of existing technologies in simultaneously addressing surface state perception and environmental obstacle avoidance.

[0071] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0072] The core of this application is to provide a path planning method for a robot for demolding and transporting precast components. A flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0073] S101. Obtain electrostatic signal data on the surface of the precast component, and generate gradient distribution data and geometric contour data of the precast component by performing inverse finite element analysis and manifold learning dimensionality reduction on the electrostatic signal data. The gradient distribution data is used to characterize the dust risk distribution on the surface of the precast component.

[0074] In one specific implementation, such as Figure 2 As shown, step S101 includes:

[0075] Step 1011: Using the electrostatic signal data as boundary constraints, construct the electric potential field control equation, and solve the electric potential field control equation through the finite element analysis model to obtain the electric field intensity data on the surface of the prefabricated component.

[0076] Among them, the electric potential field control equation refers to the partial differential equation describing the distribution law of the electrostatic field. Its input is the electrostatic signal data on the boundary, and its output is the electric potential value at each point in the solution domain. The finite element analysis model refers to a computer model that discretizes the continuous solution domain into a finite number of elements and solves the partial differential equations approximately by numerical methods.

[0077] In step 1011, the collected electrostatic signal data is first used as known boundary conditions to establish the electric potential field control equation applicable to the current solution region. This equation is based on the Poisson equation or the Laplace equation and reflects the basic relationship between electric potential and electric field in a region without free charges. Subsequently, the surface of the component and the surrounding space are meshed using a finite element analysis model, and an interpolation function is constructed on each mesh element to transform the continuous partial differential equation into a discrete algebraic equation system. After solving this equation system, the electric potential distribution of the entire solution domain can be obtained. Then, by calculating the spatial gradient of the electric potential, the electric field intensity data at each node on the surface of the component is finally obtained. This process realizes the reverse reconstruction from the boundary sensor reading to the complete electric field distribution on the surface of the component.

[0078] Step 1012: Perform neighborhood sampling on the electric field intensity data to obtain multiple electric field intensity vectors, and aggregate all the electric field intensity vectors to obtain a high-dimensional feature set.

[0079] In step 1012, sampling points are first arranged on the surface of the component at a uniform or adaptive density. Neighborhood sampling is performed on each sampling point. Specifically, a circular or rectangular neighborhood window containing several adjacent grid nodes is defined with the current sampling point as the center. The electric field intensity data of all nodes in the window are extracted and flattened into a one-dimensional array in spatial order to obtain an electric field intensity vector. This vector completely records the electric field distribution characteristics of the local area. After traversing all sampling points, all the obtained electric field intensity vectors are gathered together to form a high-dimensional feature set. This feature set contains the original high-dimensional representation of the electric field distribution of each local area on the surface of the component.

[0080] Step 1013: Based on the high-dimensional feature set, construct an adjacency graph using the equidistant feature mapping algorithm, and calculate the shortest path distance between data points in the adjacency graph to obtain the geodesic distance matrix.

[0081] Among them, the equidistant feature mapping algorithm is a nonlinear dimensionality reduction method that approximates the geodesic distance on the manifold by constructing a neighborhood graph and calculating the shortest path distance on the graph. The geodesic distance matrix is ​​a symmetric matrix whose element (i, j) represents the cumulative edge length of the shortest path from node i to node j along the graph edge on the adjacency graph, and is used to approximate the true geodesic distance on the high-dimensional manifold.

[0082] In step 1013, each vector in the high-dimensional feature set is first regarded as a point in the high-dimensional space. Then, the isometric feature mapping algorithm is applied to perform manifold structure analysis. The first step of the algorithm is to construct an adjacency graph for these points: calculate the Euclidean distance between all pairs of points, and find the K nearest neighbors or points with a distance less than ε for each point, and establish connecting edges between these neighboring pairs of points, where the weight of the edge is set to the Euclidean distance between the two points. Subsequently, after the adjacency graph is constructed, the algorithm runs the Dijkstra or Floyd shortest path algorithm on the graph structure to calculate the cumulative distance of the shortest path along the graph edge between any two data points, and then fills these distance values ​​into the corresponding positions of the matrix to finally obtain the geodesic distance matrix. This matrix approximately reflects the intrinsic distance relationship of the data points on the low-dimensional manifold embedded in the original high-dimensional feature space.

[0083] Step 1014: Perform multidimensional scaling analysis on the geodesic distance matrix to obtain low-dimensional embedded coordinates.

[0084] Step 1014 may specifically include the following steps:

[0085] Step a1: Perform multidimensional scaling analysis on the geodesic distance matrix to obtain the first embedded coordinates, and calculate the dust risk coefficient for each data point based on the electric field intensity vector in the high-dimensional feature set.

[0086] In step a1, classical multidimensional scaling analysis is first applied to the geodesic distance matrix obtained in step 1013. This analysis converts the distance matrix into an inner product matrix, performs eigenvalue decomposition on the inner product matrix, and takes the eigenvectors corresponding to the first d largest eigenvalues ​​as coordinates, thereby obtaining the first embedding coordinates of each data point in the low-dimensional embedding space. At the same time, based on the electric field intensity vector, a dust risk coefficient is calculated for each data point. This coefficient can be obtained by weighted summation of the electric field amplitude in the electric field intensity vector or by calculating the local variance of the electric field gradient, etc., and the higher the value, the greater the dust risk at that location.

[0087] Step a2: Using the dust risk coefficient as a node attribute, construct a weighted adjacency graph together with the adjacency graph, and apply the spectral clustering method to perform spectral graph clustering on the weighted adjacency graph to obtain multiple clustering regions with different risk levels.

[0088] In step a2, the dust risk coefficient is first added to the adjacency graph as a node attribute value to form a weighted adjacency graph. In the weighted adjacency graph, the edge weight still reflects the distance in the feature space, while the node itself carries risk information. Subsequently, the spectral clustering method is applied to divide the weighted adjacency graph: the Laplacian matrix of the graph is constructed, and the eigenvectors corresponding to its first k smallest eigenvalues ​​are calculated. These eigenvectors are then arranged into a new matrix according to the rows, and K-means clustering is performed on the row vectors of the matrix. The clustering result divides the data points on the surface of the component into multiple clustering regions with different risk levels, so that the risk coefficients of points in the same region are similar, and the risk levels of different regions are significantly different.

[0089] Step a3: Using the clustered regions of different risk levels as category labels, perform Fisher discriminant analysis on the first embedded coordinates to obtain the discriminant coordinates.

[0090] In step a3, the clustered regions of different risk levels obtained in step a2 are used as known category labels and paired with the first embedding coordinates to form a supervised training dataset. Then, Fisher discriminant analysis is applied to calculate the inter-class scatter matrix and the intra-class scatter matrix, and the projection direction vector that maximizes the ratio of inter-class scatter to intra-class scatter is solved. Then, the first embedding coordinates are projected onto these optimal discriminant directions to obtain the discriminant coordinates of each data point. These coordinates not only preserve the inherent geometric structure of the manifold, but also make the regions of different risk levels more separated in the low-dimensional space, which is convenient for subsequent risk-oriented processing.

[0091] Step a4: Perform feature-level fusion of the dust risk coefficient of each data point with the discrimination coordinates to obtain a fused feature vector.

[0092] In step a4, a feature-level fusion operation is performed on each data point. Specifically, the dust risk coefficient is added as a new dimension to the end of the discriminant coordinates obtained in step a3 to form a vector with one more dimension than the original discriminant coordinates. For the entire component surface, these new vectors of all data points together form a fused feature vector. This fusion method makes the representation of each point include both its geometric position information in the discriminant space and explicitly carry the dust risk quantification value of that point, which can provide a more comprehensive feature basis for subsequent dimensionality reduction processing.

[0093] Step a5: Calculate the fused feature vector using the local preserving projection algorithm to obtain low-dimensional embedded data.

[0094] In step a5, the local preservation projection algorithm is applied to perform the final dimensionality reduction of the fused feature vectors. The algorithm first constructs a nearest neighbor graph for all fused feature vectors and determines the K nearest neighbors of each vector. Then, a weight matrix is ​​constructed based on the nearest neighbor relationship. If two vectors are adjacent, they are assigned a hot kernel weight or a simple weight of 1, otherwise it is 0. Next, an objective function is constructed, requiring that this nearest neighbor weight relationship be maintained in the low-dimensional space. The projection matrix is ​​then obtained by solving the generalized eigenvalue problem. Finally, the original fused feature vectors are projected onto the first d column vectors of the projection matrix to obtain the final low-dimensional embedding data. This result preserves the local geometric structure, discriminative information, and risk quantification information of the original features, providing high-quality input for subsequent geometric reconstruction and gradient analysis.

[0095] Step 1015: Based on the low-dimensional embedded coordinates, reconstruct the geometric parameter representation of the surface of the precast component, and calculate the gradient field at each data point location based on the geometric parameter representation to obtain gradient distribution data.

[0096] Among them, geometric parameter representation refers to the parameters that describe the surface shape of a component in a mathematical way, such as surface equation coefficients, the mapping relationship between parameterized coordinates and three-dimensional spatial coordinates, etc.

[0097] In step 1015, geometric reconstruction is first performed using low-dimensional embedded coordinates. Since the low-dimensional embedded coordinates already reflect the inherent geometric structure of the component surface, a mapping function from the low-dimensional parameter space to the original three-dimensional space can be established through interpolation or fitting methods, thereby obtaining the three-dimensional spatial coordinates corresponding to each data point, i.e., the geometric parameter representation. Based on this, the dust risk coefficient is used as a scalar field defined on the component surface, and the spatial gradient of this scalar field is calculated: at each data point, the risk coefficient change of its neighborhood points is considered, and the magnitude and direction of the gradient at that point are calculated through numerical differentiation. After performing this operation on all points, the gradient distribution data describing the risk change characteristics at each location on the surface is obtained.

[0098] Step 1016: Perform Delaunay triangulation on the low-dimensional embedded coordinates to obtain the geometric contour data of the prefabricated component.

[0099] In step 1016, the low-dimensional embedded coordinates are first regarded as a set of points in a two-dimensional parameter domain, and then a triangular mesh is constructed according to the Delaunay criterion: for any two adjacent triangles, the sum of the two angles opposite their common side is less than 180 degrees, and the circumcircle of each triangle does not contain other points. This process can be efficiently implemented by point-by-point insertion or divide-and-conquer method. After the meshing is completed, a series of triangular patches and their vertex index relationships are obtained. These vertex indices are then combined with the three-dimensional spatial coordinates reconstructed in step 1015 to finally generate complete geometric contour data. This data accurately describes the surface geometry of the prefabricated component in the form of a triangular mesh.

[0100] This application achieves accurate reconstruction from physical signals to component surface geometry and risk information, providing a data foundation for subsequent path planning that simultaneously includes dust distribution characteristics and precise shape.

[0101] S102. The gradient distribution data is smoothed by Gaussian convolution to generate a first repulsive potential field.

[0102] In one specific implementation, step S102 includes:

[0103] Step 1021: Map the gradient distribution data to a two-dimensional parametric grid corresponding to the surface of the precast component to obtain an initial gradient field, and extract the surface curvature information of each grid point on the two-dimensional parametric grid to obtain a curvature distribution field.

[0104] Among them, surface curvature information refers to the geometric quantity that describes the degree of curvature of the component surface, including principal curvature, Gaussian curvature, etc., which is used to characterize whether the surface is flat, convex or concave; curvature distribution field refers to the scalar field or tensor field formed by storing the surface curvature information of each grid point at the corresponding grid position.

[0105] In step 1021, the surface of the precast component is first parametrically processed, and a mapping relationship from the three-dimensional surface to the two-dimensional plane is established to generate a regularly arranged two-dimensional parametric mesh. This process is achieved through conformal mapping or equal area mapping algorithms, ensuring that the mesh is continuous and non-overlapping on the two-dimensional plane. After the mesh construction is completed, the gradient distribution data is assigned to each mesh point through interpolation to form an initial gradient field. At the same time, based on the three-dimensional geometric model of the component surface, the surface curvature information of the three-dimensional position corresponding to each mesh point on the two-dimensional parametric mesh is calculated, and these curvature values ​​are stored according to the mesh index to obtain the curvature distribution field.

[0106] Step 1022: Calculate the structure tensor of each grid point in the initial gradient field, and determine the principal direction and anisotropy degree at each grid point based on the structure tensor.

[0107] Here, the structure tensor is a 2x2 symmetric positive semi-definite matrix calculated from the local gradient of the vector field, used to describe the directional characteristics of the gradient distribution in the local neighborhood; the principal direction refers to the direction of the eigenvector corresponding to the largest eigenvalue of the structure tensor.

[0108] In step 1022, a local structure analysis is performed on the initial gradient field. Specifically, for each grid point, a rectangular neighborhood window of fixed size is defined centered on that point, and the gradient vectors of all grid points within the window are collected. Then, based on these gradient vectors, the structure tensor J of that point is calculated using the following formula: ,in, and Let x and y represent the gradient components of each point in the neighborhood, respectively. The summation operation covers the entire neighborhood window. Then, eigenvalue decomposition is performed on this matrix to obtain two eigenvalues. and And the largest eigenvalue The direction of the corresponding eigenvector is the principal direction at that point; then the degree of anisotropy A is calculated, and its formula is: ,in, It is a very small positive number, used to avoid the denominator being zero. The largest eigenvalue after eigenvalue decomposition of the structure tensor. It is the second largest or smallest eigenvalue after eigenvalue decomposition of the structure tensor.

[0109] Step 1023: Based on the principal direction and the degree of anisotropy, use a Gaussian convolution kernel to convolve and smooth the initial gradient field to obtain the target gradient field, and calculate the magnitude of the target gradient field to obtain the risk intensity field.

[0110] In step 1023, an adaptive Gaussian convolution kernel is constructed using the principal direction and the degree of anisotropy, and anisotropic smoothing is performed on the initial gradient field. Then, at each grid point, the Gaussian kernel is rotated according to the principal direction of that point so that its major axis is aligned with the principal direction. At the same time, the standard deviation of the Gaussian kernel in the vertical principal direction is adjusted according to the degree of anisotropy A. When A is close to 1, it indicates that the standard deviation in the vertical direction is very small, and the smoothing is mainly performed along the principal direction. When A is close to 0, it indicates that the smoothing is isotropic. Then, the adaptive convolution kernel is used to perform convolution operation on each component of the initial gradient field to obtain the smoothed target gradient field.

[0111] Next, the vector magnitude of the target gradient field at each grid point is calculated: , where Q Let x and y represent the field values ​​of the risk intensity field, where x and y represent the x-coordinate and y-coordinate of the grid point, respectively. and These are the two components of the smoothed gradient; finally, the magnitude values ​​of all grid points are arranged according to their positions to obtain the risk intensity field.

[0112] Step 1024: Based on the curvature distribution field, the risk intensity field is modulated using a nonlinear transformation function to obtain the modulated risk field.

[0113] In step 1024, the curvature distribution field is used as the modulation basis to nonlinearly adjust the risk intensity field. For example, a typical transformation form is exponential modulation: ,in, Let u be the modulated risk field value, and u be an adjustable parameter to control the degree to which curvature amplifies the risk. The curvature value in the curvature distribution field for each grid point.

[0114] Step 1025: The modulated risk field is transformed using an exponential potential function to obtain a first repulsive potential field.

[0115] In step 1025, the modulated risk field is used as input, and an exponential potential function is applied to convert it into a first repulsive potential field. The conversion formula is designed as follows: ,in, It is the upper limit threshold where the risk is negligible, and η is the first gain coefficient. When the modulated risk field value is less than or equal to... When the potential field value decreases, the potential field value increases sharply; when the modulated risk field value is greater than 1, the potential field value increases sharply. At this point, the potential field value is 0; then, this transformation is performed on all grid points to obtain the first repulsive potential field covering the entire surface of the component.

[0116] This application realizes the transformation from dust risk gradient to robot-perceptible repulsion field, and retains risk edge characteristics while suppressing sensor noise. Subsequently, the risk sensitivity is dynamically adjusted according to the geometry of the component, which can provide potential field constraints for subsequent path planning that can effectively guide the robot to avoid dust-prone areas.

[0117] S103. Based on the static obstacle information represented by the three-dimensional laser point cloud data of the current production scene, construct a second repulsive potential field based on the Euclidean distance of the obstacle in the component surface area corresponding to the geometric contour data.

[0118] In step S103, the three-dimensional laser point cloud data of the current production scene is first acquired. This data is collected in real time by the lidar deployed around the production station. Then, the raw point cloud data is preprocessed, including noise reduction, downsampling and ground point filtering. Then, the static obstacle information containing only fixed facilities is extracted by the dynamic object filtering algorithm. This process can be implemented by grid occupancy comparison or deep learning-based target segmentation method to ensure that the output obstacle information is accurate and timely.

[0119] After obtaining the static obstacle information, it is spatially registered with the geometric contour data generated in step S101 and unified to the same world coordinate system. The geometric contour data accurately describes the three-dimensional shape of the prefabricated component surface, while the static obstacle information characterizes the spatial distribution of fixed facilities in the environment around the component. After superimposing the two, the area where there are obstacles around the component surface can be identified.

[0120] Next, at each point on the component surface, the Euclidean distance from that point to the nearest static obstacle is calculated; the smaller the distance, the larger the repulsive potential field value. For any point p on the component surface, its second repulsive potential field value... The calculation formula is: Where d(p) represents the Euclidean distance from point p to the nearest static obstacle. It is the threshold distance of the obstacle's influence range. It is the second gain coefficient, when the distance d(p) is less than or equal to When the distance d(p) is greater than a certain value, the potential field value increases sharply as the distance decreases, indicating that the repulsion is stronger the closer to the obstacle; when the distance d(p) is greater than a certain value, the potential field value increases sharply as the distance decreases, indicating that the repulsion is stronger the closer to the obstacle. At that time, the influence of obstacles is considered negligible, and the potential field value is 0.

[0121] During the calculation process, it is necessary to query the three-dimensional coordinates of each grid point in the geometric contour data, and then calculate the shortest Euclidean distance from the coordinates to all obstacle surfaces described by the static obstacle information. This calculation can be achieved by constructing the distance field of the obstacle or by using a KD tree to accelerate the nearest neighbor query. Finally, the calculated potential field value is assigned to the corresponding component surface grid point, thus obtaining the second repulsive potential field covering the entire component surface.

[0122] This application enables spatial constraint modeling of fixed obstacles in the production environment, and can effectively guide the robot end effector away from collision risk areas in subsequent path planning, thereby ensuring equipment safety and smooth operation during the grasping process.

[0123] S104. Based on the three-dimensional pose information of the prefabricated components and combined with the geometric constraints of the preset grasping path, generate an attractive potential field pointing towards the target point.

[0124] In one specific implementation, step S104 includes:

[0125] Step 1041: Obtain the three-dimensional pose information of the prefabricated component, and the relative pose relationship in space between the tool center point and feature point coordinates of the robot end effector.

[0126] The tool center point refers to the reference point defined on the robot's end effector. It is usually set at a critical position of the tool, such as the center of the suction cup or the gripper's grasping point. The essence of robot movement is to move the tool center point to the target position.

[0127] In step 1041, the three-dimensional pose information of the prefabricated component is first obtained from the upper control system. This information is usually measured in real time by a vision positioning system or a laser tracker and reflects the actual spatial state of the component in the current environmental coordinate system. At the same time, the coordinates of the tool center point of the robot's current end effector are obtained. These coordinates are provided in real time by the robot controller through forward kinematics calculation based on the joint angles. Then, based on the hand-eye calibration results, the relative pose relationship between the tool center point and the coordinates of the feature points on the component surface is calculated.

[0128] Step 1042: Using the environmental coordinate system defined by the three-dimensional pose information as a reference, and combining the relative pose relationship, the kinematic model of the robot, and the geometric constraints of the grasping path, sample the path from the center point of the tool to the coordinates of the feature point to generate multiple candidate grasping paths.

[0129] Among them, the environmental coordinate system refers to the global reference coordinate system defined in the production scene, which is usually established with the robot base or fixed point of the site as the origin.

[0130] In step 1042, the environmental coordinate system defined by the three-dimensional pose information is used as the global reference. The relative pose relationship, the robot's kinematic model, and the geometric constraints of the preset grasping path are used as input conditions. Based on this, the path for the tool center point to move to the feature point coordinates is sampled and generated. The sampling process is usually carried out in the robot's joint space: first, the starting point configuration is determined according to the current joint angle, then the target joint angle corresponding to the feature point coordinates is solved by inverse kinematics, and then multiple paths are generated in the joint space by interpolation. Each path corresponds to a set of transition sequences from the starting joint angle to the target joint angle, and must meet the geometric constraints of the grasping path, such as avoiding high-risk areas on the surface of the component and maintaining the stability of the end face. Then, by adjusting the interpolation parameters or adopting different motion planning strategies, multiple candidate grasping paths are generated.

[0131] Step 1043: Calculate the smoothness index and execution time estimate for each candidate crawling path, and determine the attraction intensity value for each candidate crawling path based on the smoothness index and execution time estimate.

[0132] In step 1043, firstly, the smoothness index of each candidate grasping path is obtained by integrating the curvature or joint angular acceleration change rate at each point on the corresponding candidate grasping path, with a smaller value indicating a smoother path. Simultaneously, the execution time is estimated based on the total path length and the maximum motion speed of each robot axis. Then, the attraction intensity value of each candidate grasping path is determined based on these two indices: paths with higher smoothness and shorter time are assigned higher attraction intensity values, which can be achieved through a weighted summation function. ,in, This represents the attraction strength value of the i-th path. For smoothness index, For execution time estimation, α and β are weighting coefficients for adjusting smoothness and the contribution of execution time to attraction intensity.

[0133] Step 1044: Using the coordinates of the feature point as the center, diffuse each attraction intensity value along the corresponding spatial direction to generate a vector direction field, and convert the vector direction field into an attraction potential field.

[0134] In step 1044, with the feature point coordinates as the center, the attraction intensity value of each candidate grasping path is diffused along its direction in space. The diffusion process adopts a weighted interpolation method: for any spatial point on the surface of the component and its surroundings, the shortest projection distance from the point to each candidate grasping path is calculated, and the attraction intensity value of the corresponding path is superimposed on the point according to the distance weight. At the same time, the direction of the vector is weighted and synthesized by the directions of multiple paths. This process generates a vector direction field covering the grasping area, where the vector at each point points to the final target direction, and the magnitude reflects the attraction intensity at that position.

[0135] Subsequently, after constructing the vector direction field, it is converted into an attractive potential field. This conversion is achieved by performing a line integral of the vector direction field along the path: for any point in space, the virtual "work" required to travel from that point to the coordinates of the feature point along the vector direction field is calculated. The value of this work is the value of the attractive potential field at that point. The smaller the value, the closer the robot is to the target, and the stronger the virtual attraction force it experiences. The final generated attractive potential field achieves a global minimum at the coordinates of the feature point, and the potential field value gradually increases in the region far from the target point.

[0136] This application realizes the transformation from component pose information to a robot-sensible gravitational field. The generated attractive potential field can effectively guide the robot end effector to approach the target point along the optimal grasping path, thereby providing accurate target guidance constraints for subsequent composite potential field construction and path planning.

[0137] S105. Construct a composite artificial potential field by adaptively weighting and superimposing the first repulsive potential field, the second repulsive potential field, and the attractive potential field.

[0138] In one specific implementation, step S105 includes:

[0139] Step 1051: Discretize the first repulsive potential field, the second repulsive potential field, and the attractive potential field to the same spatial grid coordinate system to obtain the first repulsive potential field value, the second repulsive potential field value, and the attractive potential field value corresponding to each grid point.

[0140] In step 1051, the range and grid resolution of the robot's workspace are first determined, and a spatial grid coordinate system covering the entire work area is established. Then, the first repulsive potential field, the second repulsive potential field, and the attractive potential field are mapped to the grid coordinate system. For each grid point, the potential field value at that point is obtained by interpolation from the continuous definition of each potential field according to its spatial coordinates, thereby obtaining the first repulsive potential field value, the second repulsive potential field value, and the attractive potential field value corresponding to each grid point. This discretization operation enables the three potential fields from different sources to complete numerical alignment on the same set of spatial discrete points, providing a unified data foundation for subsequent joint processing.

[0141] Step 1052: Input the first repulsive potential field value, the second repulsive potential field value, and the attractive potential field value corresponding to each grid point into the multilayer perceptron network, perform forward propagation calculation through the multilayer perceptron network, and output the first repulsive weight, the second repulsive weight, and the attractive weight of the corresponding grid point.

[0142] Among them, the multilayer perceptron network refers to a feedforward neural network structure, which includes an input layer, several hidden layers and an output layer, and the layers are connected in a fully connected manner. In this scheme, the input layer of the network receives three potential field values, and the output layer generates three weight coefficients to adaptively adjust the contribution of each potential field in the composite potential field.

[0143] In step 1052, the first repulsive potential value, the second repulsive potential value, and the attractive potential value at each grid point are combined into a three-dimensional input vector and fed into a pre-trained multilayer perceptron network. The network structure is designed as follows: the input layer contains 3 neurons, corresponding to three potential values; there are two hidden layers, with the first hidden layer containing 8 neurons and the second hidden layer containing 8 neurons, both using ReLU activation function; the output layer contains 3 neurons, corresponding to three weights, using Softmax activation function; then the forward propagation process of the network passes through the input layer, the first hidden layer, and the second hidden layer in sequence, and finally the output layer is normalized by Softmax to obtain the first repulsive weight, the second repulsive weight, and the attractive weight of the grid point. This process allows the weight allocation to be dynamically adjusted according to the specific values ​​of the local potential field.

[0144] The multilayer perceptron network needs to be trained in advance. The training process adopts a supervised learning method, collecting a large amount of potential field data and corresponding optimal weight combinations in typical grasping scenarios as training samples. The success rate of path planning and trajectory smoothness are used as evaluation indicators. The network weight parameters are adjusted through the backpropagation algorithm so that the weights output by the network can effectively guide the robot to complete the grasping task through the composite potential field.

[0145] Step 1053: Multiply the first repulsive potential field value by the first repulsive weight to obtain a weighted first repulsive field value; multiply the second repulsive potential field value by the second repulsive weight to obtain a weighted second repulsive field value; and multiply the attractive potential field value by the attractive weight to obtain a weighted attractive potential field value.

[0146] In step 1053, the first repulsive potential field value of the grid point is multiplied by the first repulsive weight to obtain the weighted first repulsive field value; the second repulsive potential field value is multiplied by the second repulsive weight to obtain the weighted second repulsive field value; and the attractive potential field value is multiplied by the attractive weight to obtain the weighted attractive potential field value. This operation realizes the dynamic adjustment of the intensity of the three original potential fields, so that the influence of a certain type of potential field in dust avoidance, obstacle avoidance or target attraction can be highlighted in different spatial locations according to the local environmental characteristics.

[0147] Step 1054: Perform algebraic summation of the weighted first repulsion field value, the weighted second repulsion field value, and the weighted attraction field value at the same grid point to generate the composite potential field value of the corresponding grid point.

[0148] In step 1054, the weighted first repulsion field value, the weighted second repulsion field value, and the weighted attraction potential field value of the grid point are directly added together to obtain the composite potential field value of the point. This value comprehensively reflects the combined effect of three factors: dust risk, static obstacle avoidance, and target attraction. The larger the value, the less suitable the position is for the robot end to stop or pass through, and the smaller the value, the more suitable it is as a point on the path.

[0149] Step 1055: Smooth the composite potential field values ​​of all grid points to obtain the composite artificial potential field.

[0150] In step 1055, a Gaussian filtering method is used to perform a weighted average of the composite potential field values ​​in the neighborhood of each grid point to eliminate local noise and abrupt changes introduced by discrete sampling or weight calculation; the final smoothed potential field is continuous and smooth in space, which is the final composite artificial potential field.

[0151] This application achieves intelligent fusion of multiple constraints such as dust risk, obstacle avoidance, and target attraction. It can adjust the weight of each constraint in real time according to local environmental characteristics, thereby making the generated composite potential field more in line with actual operation requirements and providing accurate and continuous comprehensive navigation information for subsequent path planning.

[0152] S106. Combining the hybrid A* algorithm and the Reeds-Shepp curve model, node connection and trajectory feasibility verification are performed to evaluate and explore state nodes in the composite artificial potential field and generate trajectory planning data for the robot.

[0153] In one specific implementation, step S106 includes:

[0154] Step 1061: Construct an open list based on the starting node, and generate multiple motion primitives according to the robot's nonholonomic kinematic constraints and dynamic characteristics.

[0155] In step 1061, the robot's initial pose is first defined as the starting node, and an empty open list is initialized based on this. Then, the starting node is added to the list. Subsequently, based on the robot's nonholonomic kinematic constraints and dynamic characteristics, a set of motion primitives is predefined. These motion primitives cover the basic actions that the robot may perform, such as moving forward, moving backward, turning left and forward, turning right and forward, etc., and each primitive corresponds to a fixed path length and curvature change.

[0156] Step 1062: Perform kinematic integration on each of the motion primitives to obtain multiple next nodes corresponding to the starting node, and calculate the segment cost from the starting node to each of the next nodes in the composite artificial potential field.

[0157] In step 1062, kinematic integration is performed on each motion primitive, that is, starting from the state of the initial node, integrating point by point along the trajectory defined by the primitive to obtain the pose after the primitive is executed, thus generating a next node. This process is repeated for all primitives to obtain multiple next nodes. Subsequently, for each generated next node, the corresponding road segment cost is calculated in the composite artificial potential field. The calculation formula is as follows: Where L is the path length of the motion primitive. Let be the composite artificial potential field value at position s on the path, where s is the position variable on the path. and The integral value is used to adjust the weighting coefficients of the basic cost and the composite potential cost. It reflects the comprehensive cost of the journey from the current node to the next node.

[0158] Step 1063: Based on the road segment cost value, obtain the actual cost value corresponding to the next node, and calculate the estimated cost value from the next node to the target node. Add the actual cost value and the estimated cost value to obtain the comprehensive cost value.

[0159] In step 1063, firstly, the road segment cost is added to the cumulative cost from the starting point to the current starting node to obtain the actual cost corresponding to the next node; at the same time, based on the state of the current next node and the state of the target node, the estimated cost is calculated. The estimated cost is calculated using the Reeds-Shepp curve model to solve for the shortest curve length from the current pose to the target pose that satisfies the robot's minimum turning radius, as a heuristic estimate; then, the actual cost and the estimated cost are added to obtain the comprehensive cost of the next node.

[0160] Step 1064: Calculate the shortest curve from the next node to the target node using the Reeds-Shepp curve model, and verify the curvature continuity of the shortest curve in the composite artificial potential field.

[0161] The Reeds-Shepp curve model is a mathematical model that describes the shortest path for a vehicle to travel from any pose to any pose in a plane. The path consists of a combination of straight line segments and circular arc segments with fixed curvature, and satisfies the minimum turning radius constraint.

[0162] In step 1064, the Reeds-Shepp curve model is used to solve for the shortest path curve that satisfies the robot's minimum turning radius constraint based on the pose of the current next node and the target node. After obtaining the curve, curvature continuity is checked in the composite artificial potential field. The check includes two aspects: first, checking whether the curvature change of the curve is within the robot's allowable range and whether there are abrupt change points; second, densely sampling the points on the curve and checking whether the composite artificial potential field value at each sampling point exceeds the safety threshold to ensure that the entire curve does not pass through high-risk areas or near obstacles.

[0163] Step 1065: If the verification fails, add the next node, its corresponding actual value, estimated value, and parent node pointer to the open list. In the next round, select the node with the smallest comprehensive value in the open list as the starting node and repeat the process of calculating the value to expand the open list until the verification is passed.

[0164] In step 1065, if the curvature continuity check in step 1064 fails, it means that the current next node cannot be directly connected to the target node through a curve that satisfies the constraints. At this time, the next node, along with its actual cost value, estimated cost value, and the pointer to the parent node pointing to the current starting node, is added to the open list. Subsequently, the algorithm enters the next iteration: the node with the smallest current comprehensive cost value is selected from the open list as the new starting node, and the process from step 1062 to step 1065 is repeated, and the search tree is continuously expanded. This cycle continues until, in a certain iteration, the Reeds-Shepp curve from a certain next node to the target node passes the curvature continuity check.

[0165] Step 1066: If the verification is successful, take the next node as the connection node, and backtrack the search path of the connection node in the open list according to the parent node pointer to obtain the node sequence.

[0166] The connecting node refers to the next node that ultimately passes the verification and successfully connects to the target node via the Reeds-Shepp curve.

[0167] In step 1066, after the verification in step 1064 passes, the current next node is determined as the connection node. Starting from the connection node, the node is traced forward step by step along the parent node pointers it records, and a parent node is obtained at each step of the tracing, until the original starting node is reached. Then, these nodes are arranged in order from the starting node to the connection node to obtain a complete node sequence. This sequence records all intermediate states from the starting point to the final successful connection point.

[0168] Step 1067: Along the node sequence, sequentially splice the trajectory segments corresponding to all the motion primitives with the shortest curve that has passed the verification to generate the trajectory planning data of the robot.

[0169] In step 1067, along the node sequence obtained in step 1066, the corresponding motion element trajectory segments between each pair of adjacent nodes in the sequence are connected sequentially. After reaching the connection node, the Reeds-Shepp shortest curve that passed the verification in step 1064 is spliced ​​at the end so that the path finally reaches the target node. Then, all segments are smoothly connected in sequence to generate trajectory planning data containing information such as position, attitude, velocity, and acceleration. This data is the final output, which is used to drive the robot to complete the demolding and transfer operation.

[0170] This application achieves efficient path search guided by a composite potential field. The generated trajectory avoids high-risk dust areas and static obstacles, strictly meets the kinematic feasibility of the robot, and ensures the stability and safety of the grasping process.

[0171] Figure 3 This is a schematic diagram of a specific embodiment of a path planning system for a precast component demolding and transfer robot provided in this application, with reference to... Figure 3 The system may include:

[0172] The acquisition module 31 is used to acquire electrostatic signal data on the surface of the precast component, and generate gradient distribution data and geometric contour data of the precast component by performing inverse finite element analysis and manifold learning dimensionality reduction on the electrostatic signal data. The gradient distribution data is used to characterize the dust risk distribution on the surface of the precast component.

[0173] The processing module 32 is used to perform Gaussian convolution smoothing on the gradient distribution data to generate a first repulsive potential field.

[0174] The first construction module 33 is used to construct a second repulsive potential field based on the Euclidean distance of the obstacle in the component surface area corresponding to the geometric contour data, based on the static obstacle information characterized by the three-dimensional laser point cloud data of the current production scene.

[0175] The generation module 34 is used to generate an attractive potential field pointing to the target point based on the three-dimensional pose information of the prefabricated component and the geometric constraints of the preset grasping path.

[0176] The second construction module 35 is used to construct a composite artificial potential field by adaptively weighting and superimposing the first repulsive potential field, the second repulsive potential field, and the attractive potential field.

[0177] The verification module 36 is used to combine the hybrid A* algorithm and the Reeds-Shepp curve model to verify the node connection and trajectory feasibility, so as to complete the evaluation and exploration of state nodes in the composite artificial potential field and generate trajectory planning data for the robot.

[0178] The path planning system of the precast component demolding and transfer robot in this application embodiment is used to implement the aforementioned path planning method of the precast component demolding and transfer robot. Therefore, the specific implementation of the path planning system of the precast component demolding and transfer robot can be seen in the embodiment section of the path planning method of the precast component demolding and transfer robot above. The specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0179] like Figure 4 This application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of the path planning method for a prefabricated component demolding and transfer robot described above.

[0180] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the path planning method for a prefabricated component demolding and transfer robot described above.

[0181] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0182] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the path planning method for a prefabricated component demolding and transfer robot.

[0183] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0184] The path planning method and system for a precast component demolding and transfer robot provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A path planning method for a precast component demolding and transfer robot, characterized in that, include: Electrostatic signal data of the surface of the precast component is acquired, and gradient distribution data and geometric contour data of the precast component are generated by performing inverse finite element analysis and manifold learning dimensionality reduction on the electrostatic signal data. The gradient distribution data is used to characterize the dust risk distribution on the surface of the precast component. The gradient distribution data is smoothed by Gaussian convolution to generate a first repulsive potential field. Based on the static obstacle information represented by the three-dimensional laser point cloud data of the current production scene, a second repulsive potential field based on the Euclidean distance of the obstacle is constructed on the surface region of the component corresponding to the geometric contour data. Based on the three-dimensional pose information of the prefabricated components, combined with the geometric constraints of the preset grasping path, an attractive potential field pointing towards the target point is generated. A composite artificial potential field is constructed by adaptively weighting and superimposing the first repulsive potential field, the second repulsive potential field, and the attractive potential field. By combining the hybrid A* algorithm and the Reeds-Shepp curve model, node connections and trajectory feasibility verification are performed to evaluate and explore state nodes in the composite artificial potential field, thereby generating trajectory planning data for the robot.

2. The method according to claim 1, characterized in that, The process of performing inverse finite element analysis and manifold learning dimensionality reduction on the electrostatic signal data to generate gradient distribution data and geometric contour data of the precast components includes: Using the electrostatic signal data as boundary constraints, an electric potential field control equation is constructed. The electric potential field control equation is solved through a finite element analysis model to obtain the electric field intensity data on the surface of the prefabricated component. The electric field intensity data is sampled in a neighborhood to obtain multiple electric field intensity vectors. All the electric field intensity vectors are then aggregated to obtain a high-dimensional feature set. Based on the high-dimensional feature set, an adjacency graph is constructed using an equidistant feature mapping algorithm, and the shortest path distance between data points in the adjacency graph is calculated to obtain a geodesic distance matrix. Multidimensional scaling analysis is performed on the geodesic distance matrix to obtain low-dimensional embedded coordinates; Based on the low-dimensional embedded coordinates, the geometric parameter representation of the surface of the precast component is reconstructed, and the gradient field at each data point is calculated based on the geometric parameter representation to obtain gradient distribution data. Delaunay triangulation is performed on the low-dimensional embedded coordinates to obtain the geometric contour data of the prefabricated component.

3. The method according to claim 2, characterized in that, The multidimensional scaling analysis of the geodesic distance matrix to obtain low-dimensional embedded coordinates includes: Perform multidimensional scaling analysis on the geodesic distance matrix to obtain the first embedded coordinates, and calculate the dust risk coefficient for each data point based on the electric field intensity vector in the high-dimensional feature set; The dust risk coefficient is used as a node attribute and together with the adjacency graph to construct a weighted adjacency graph. The spectral clustering method is then applied to perform spectral clustering on the weighted adjacency graph to obtain multiple clustering regions with different risk levels. Using the clustered regions with different risk levels as category labels, Fisher discriminant analysis is performed on the first embedded coordinates to obtain the discriminant coordinates; The dust risk coefficient of each data point is fused with the discrimination coordinates at the feature level to obtain a fused feature vector; The fused feature vector is calculated using a local preservation projection algorithm to obtain low-dimensional embedded data.

4. The method according to claim 1, characterized in that, The step of smoothing the gradient distribution data using Gaussian convolution to generate a first repulsive potential field includes: The gradient distribution data is mapped to a two-dimensional parametric grid corresponding to the surface of the precast component to obtain an initial gradient field, and the surface curvature information of each grid point on the two-dimensional parametric grid is extracted to obtain a curvature distribution field. Calculate the structure tensor of each grid point in the initial gradient field, and determine the principal direction and degree of anisotropy at each grid point based on the structure tensor; Based on the principal direction and the degree of anisotropy, the initial gradient field is convolved and smoothed using a Gaussian convolution kernel to obtain the target gradient field. The magnitude of the target gradient field is then calculated to obtain the risk intensity field. Based on the curvature distribution field, the risk intensity field is modulated using a nonlinear transformation function to obtain the modulated risk field; The modulated risk field is transformed using an exponential potential function to obtain the first repulsive potential field.

5. The method according to claim 1, characterized in that, The method combines the hybrid A* algorithm and the Reeds-Shepp curve model to perform node connection and trajectory feasibility verification, thereby evaluating and exploring state nodes in the composite artificial potential field and generating trajectory planning data for the robot, including: An open list is constructed based on the starting node, and multiple motion primitives are generated according to the robot's nonholonomic kinematic constraints and dynamic characteristics. Perform kinematic integration on each of the motion primitives to obtain multiple next nodes corresponding to the starting node, and calculate the segment cost from the starting node to each of the next nodes in the composite artificial potential field; Based on the road segment cost value, the actual cost value corresponding to the next node is obtained, and the estimated cost value from the next node to the target node is calculated. The actual cost value and the estimated cost value are added together to obtain the comprehensive cost value. The Reeds-Shepp curve model is used to calculate the shortest curve from the next node to the target node, and the curvature continuity of the shortest curve is verified in the composite artificial potential field. If the verification fails, the next node, along with its corresponding actual value, estimated value, and parent node pointer, is added to the open list. In the next round, the node with the smallest comprehensive value in the open list is selected as the starting node, and the process of calculating the value to expand the open list is repeated until the verification passes. If the verification is successful, the next node is taken as the connection node, and the search path of the connection node in the open list is traced back in reverse according to the parent node pointer to obtain the node sequence. The trajectory segments corresponding to all the motion primitives are sequentially spliced ​​together with the shortest curve that has passed the verification along the node sequence to generate the trajectory planning data of the robot.

6. The method according to claim 1, characterized in that, The step of constructing a composite artificial potential field by adaptively weighting and superimposing the first repulsive potential field, the second repulsive potential field, and the attractive potential field includes: Discretize the first repulsive potential field, the second repulsive potential field and the attractive potential field to the same spatial grid coordinate system to obtain the first repulsive potential field value, the second repulsive potential field value and the attractive potential field value corresponding to each grid point. The first repulsive potential field value, the second repulsive potential field value, and the attractive potential field value corresponding to each grid point are input into the multilayer perceptron network. The forward propagation calculation is performed through the multilayer perceptron network to output the first repulsive weight, the second repulsive weight, and the attractive weight of the corresponding grid point. Multiply the first repulsive potential field value by the first repulsive weight to obtain the weighted first repulsive field value; multiply the second repulsive potential field value by the second repulsive weight to obtain the weighted second repulsive field value; and multiply the attractive potential field value by the attractive weight to obtain the weighted attractive potential field value. The weighted first repulsive field value, the weighted second repulsive field value, and the weighted attractive field value at the same grid point are algebraically summed to generate the composite potential field value of the corresponding grid point. The composite potential field value of all grid points is smoothed to obtain the composite artificial potential field.

7. The method according to claim 1, characterized in that, The three-dimensional pose information based on the prefabricated components, combined with the geometric constraints of the preset grasping path, generates an attractive potential field pointing towards the target point, including: The three-dimensional pose information of the prefabricated component and the relative pose relationship in space between the tool center point and the feature point coordinates of the robot end effector are obtained. Using the environmental coordinate system defined by the three-dimensional pose information as a reference, and combining the relative pose relationship, the kinematic model of the robot, and the geometric constraints of the grasping path, the path from the center point of the tool to the coordinates of the feature point is sampled to generate multiple candidate grasping paths. Calculate the smoothness index and execution time estimate for each candidate crawling path, and determine the attraction intensity value for each candidate crawling path based on the smoothness index and execution time estimate; Centered on the coordinates of the feature point, the attraction intensity values ​​are diffused along the corresponding spatial direction to generate a vector direction field, and the vector direction field is converted into an attraction potential field.

8. A path planning system for a precast component demolding and transfer robot, characterized in that, include: The acquisition module is used to acquire electrostatic signal data on the surface of the precast component, and generate gradient distribution data and geometric contour data of the precast component by performing inverse finite element analysis and manifold learning dimensionality reduction on the electrostatic signal data. The gradient distribution data is used to characterize the dust risk distribution on the surface of the precast component. The processing module is used to perform Gaussian convolution smoothing on the gradient distribution data to generate a first repulsive potential field. The first construction module is used to construct a second repulsive potential field based on the Euclidean distance of the obstacle in the component surface area corresponding to the geometric contour data, based on the static obstacle information represented by the three-dimensional laser point cloud data of the current production scene. The generation module is used to generate an attractive potential field pointing to the target point based on the three-dimensional pose information of the prefabricated component and the geometric constraints of the preset grasping path. The second construction module is used to construct a composite artificial potential field by adaptively weighting the first repulsive potential field, the second repulsive potential field and the attractive potential field. The verification module is used to combine the hybrid A* algorithm and the Reeds-Shepp curve model to verify the node connection and trajectory feasibility, so as to complete the evaluation and exploration of state nodes in the composite artificial potential field and generate trajectory planning data for the robot.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the path planning method for a prefabricated component demolding and transfer robot as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables a path planning method for a prefabricated component demolding and transfer robot as described in any one of claims 1 to 7.

Images