Encasement path planning method and system based on deep learning

Abstract

The invention relates to the technical field of boxing path planning, and discloses a deep learning-based boxing path planning method and system, and the method comprises the steps: collecting the data of an object to be boxed, and constructing a three-dimensional coordinate parameter; performing spatial topology modeling by using the graph convolutional network to generate a geometric topological graph, and calculating a path density coefficient; carrying out graph Laplacian matrix spectrum decomposition on the path density coefficient to extract a spatial principal component vector, and generating a path planning weight through a genetic algorithm; and dynamically updating the reference path node based on the weight to generate an optimized path, and controlling the motion trail of the mechanical arm. The system comprises a three-dimensional data acquisition module, a topology modeling module, a density coefficient calculation module, a spectral decomposition module, a weight optimization module, a path generation module and a motion control module. According to the method, through combination of deep learning and an intelligent algorithm, complex geometric feature modeling and path optimization are realized, the boxing efficiency, accuracy and system adaptability are improved, and the method is suitable for boxing scenes such as logistics and warehousing.

Description

Technical Field

[0001] The present invention relates to the technical field of packing path planning, and specifically to a packing path planning method and system based on deep learning. Background Art

[0002] In logistics, warehousing, smart manufacturing, and other fields, packing is a critical and high-frequency operation. The rationality of its path planning directly affects operational efficiency, equipment loss, and cargo safety. Traditional packing path planning methods rely primarily on manual experience or classical algorithms such as genetic algorithms and simulated annealing. However, these methods have significant limitations when dealing with complex geometric shapes, dynamically changing packing spaces, and multi-objective optimization problems.

[0003] Traditional methods struggle to efficiently capture the geometric features of objects and their spatial topological relationships. For objects with irregular shapes and complex surface structures, manually designed feature extraction rules or simple geometric parameter descriptions often fail to accurately reflect the spatial constraints between objects, leading to problems such as collisions and path redundancy during path planning. For example, when packing multiple odd-shaped parts, traditional methods may generate unreasonable paths due to the inability to accurately model the adjacency relationships between parts, increasing the robot's motion energy consumption and operation time.

[0004] Traditional algorithms are less efficient when processing high-dimensional data and dynamic optimization. As packing scenarios become more complex, factors that need to be considered, such as object size, weight, spatial density, and robotic arm motion parameters, continue to increase, and the data dimension increases dramatically. The computational complexity of traditional algorithms increases exponentially, making it difficult to meet real-time requirements. In addition, traditional methods usually use fixed optimization strategies and are unable to dynamically adjust path planning weights based on real-time packing status, resulting in poor system adaptability. For example, when the object distribution density in a certain area of ​​the packing space changes, traditional methods find it difficult to quickly reallocate path planning weights, which may cause the robotic arm to frequently encounter path conflicts or detours in high-density areas.

[0005] Traditional path planning methods lack synergy with robotic arm motion control. Path planning results often fail to fully consider the robotic arm's kinematic and dynamic constraints, such as joint angle limits and speed limits. This can result in the generated path being unachievable during actual execution or requiring extensive post-process adjustments, impacting packing efficiency. For example, a path generated by traditional methods may require the robotic arm to move beyond its joint angle range, forcing the system to replan the path, increasing operational uncertainty and time costs.

[0006] The development of deep learning technology has demonstrated powerful feature learning and data processing capabilities in areas such as image recognition, natural language processing, and complex system modeling. Applying deep learning to packing path planning promises to address challenges faced by traditional methods in geometric feature modeling, high-dimensional data processing, and dynamic optimization. However, deep learning-based packing path planning technology is still in its exploratory stages. How to integrate deep learning models with the specific needs of packing path planning to build an efficient, accurate, and robust path planning system remains a pressing technical challenge. Summary of the Invention

[0007] The purpose of the present invention is to provide a packing path planning method based on deep learning to solve the problems raised in the above background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a packing path planning method based on deep learning, the method comprising:

[0009] Collect the size data, weight data and geometric feature data of the objects to be packed, and construct the three-dimensional coordinate parameters of the packing space;

[0010] A graph convolutional network is used to perform spatial topological modeling on geometric feature data, generate a geometric topological graph, and calculate the path density coefficient based on the spatial density changes between nodes in the topological graph;

[0011] The Laplacian matrix of the path density coefficients of different binned areas is spectrally decomposed to extract the spatial principal component vectors. The principal component vectors are iteratively optimized using a genetic algorithm to generate the path planning weights for each area.

[0012] The reference path nodes are dynamically updated according to the path planning weights to generate an optimized path, and the motion trajectory of the robotic arm is controlled based on the optimized path.

[0013] Preferably, the spatial topology modeling includes:

[0014] Convert the geometric feature data into a three-dimensional space tensor and eliminate the dimension difference through normalization;

[0015] Multi-layer graph convolution layers are used to aggregate neighborhood features of the three-dimensional spatial tensor to generate an adjacency matrix between nodes;

[0016] The topological connection of the adjacency matrix is ​​optimized through a dynamic routing mechanism, and a geometric topological graph is output.

[0017] Preferably, the path density coefficient calculation includes:

[0018] Select a reference path node in any boxed area and calculate the spatial density distribution of its n neighboring nodes;

[0019] The ratio of the standard deviation of spatial density to the Euclidean distance of the reference node is used as the local density measure;

[0020] Calculate the path residual between the reference node and the associated node, and take its absolute value as the global density;

[0021] The harmonic mean of the local density and the global density is taken as the path density coefficient.

[0022] Preferably, the spatial principal component vector extraction includes:

[0023] Construct the graph Laplacian matrix of path density coefficients of different binned areas, and perform spectral clustering decomposition on it to obtain eigenvectors;

[0024] The eigenvectors whose cumulative contribution rates corresponding to the eigenvalues ​​exceed the set threshold are selected to form the principal component subspace;

[0025] Project the graph Laplacian matrix onto the principal component subspace to obtain the principal component vector of the reduced-dimensional space.

[0026] Preferably, the path planning weight generation includes:

[0027] Normalize the spatial principal component vector and calculate its cosine similarity with the historical optimal path vector;

[0028] The similarity is input into the genetic algorithm population and the initial weights are generated through selection, crossover and mutation operations;

[0029] The initial weights are locally optimized through an adaptive sliding window, and the path planning weights of each area are output.

[0030] Preferably, generating the optimization path includes:

[0031] Multiply the path planning weight by the reference path node tensor to obtain the node update parameter;

[0032] Calculate the deviation between the node update parameters and the reference path, and use the dynamic programming algorithm to correct the deviation;

[0033] Connect the corrected node sequences into a continuous trajectory to generate an optimized path.

[0034] Preferably, the method further comprises:

[0035] Construct a three-dimensional packing trajectory space based on the optimized path and extract the intersection points and conflict areas in the trajectory;

[0036] When the number of intersections exceeds the set threshold or the area of ​​the conflict zone is larger than the limit, it is determined to be a path abnormality state and a trajectory adjustment instruction is generated.

[0037] Preferably, the three-dimensional packing trajectory space construction includes:

[0038] Map the optimized path to a three-dimensional coordinate system according to the motion sequence to generate a trajectory point cloud set;

[0039] The point cloud reconstruction algorithm is used to perform three-dimensional voxel modeling on the trajectory point cloud and calculate the surface curvature of the voxel model;

[0040] The curvature distribution is enhanced by the non-maximum suppression algorithm to eliminate noise interference.

[0041] Preferably, the method further comprises:

[0042] A mixed integer programming model of path planning parameters and robot arm motion parameters is established, and the branch and bound method is used to solve the optimal motion strategy.

[0043] Spatially align the optimal motion strategy with the optimized path to generate a path plan with the best packing efficiency;

[0044] The mixed integer programming model solution includes:

[0045] The objective function is defined as a linear combination of path length and motion energy consumption, and the constraint condition is the limit of the robot arm joint angle;

[0046] Perform relaxation transformation on the objective function and decompose it into path parameter subproblem and motion parameter subproblem;

[0047] The two sub-problems are solved alternately and iteratively until convergence, and the optimal motion strategy that satisfies the constraints is output.

[0048] Preferably, the present invention further includes a packing path planning system based on deep learning, the system comprising:

[0049] 3D data acquisition module, used to obtain the size, weight and geometric feature data of the objects to be packed and construct a 3D space coordinate matrix;

[0050] The topology modeling module uses a graph convolutional network to aggregate neighborhood features of the three-dimensional spatial coordinate matrix and generates a geometric topology graph through a dynamic routing mechanism;

[0051] The density coefficient calculation module generates local density and global density according to the spatial density distribution of nodes in the geometric topology graph, and calculates their harmonic mean as the path density coefficient;

[0052] The spectral decomposition module constructs a graph Laplacian matrix for the path density coefficients of different binned areas and extracts the spatial principal component vectors through spectral clustering decomposition;

[0053] The weight optimization module uses a genetic algorithm to iteratively optimize the spatial principal component vector to generate the path planning weights for each area;

[0054] The path generation module dynamically updates the reference path nodes based on the path planning weights, corrects the path deviation through the dynamic planning algorithm, and generates the optimized path;

[0055] The motion control module converts the optimized path into control instructions for the robotic arm and drives the actuator to complete the packing operation through a real-time communication protocol.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] In terms of geometric feature modeling, spatial topological modeling of the geometric feature data of objects through graph convolutional networks can effectively capture the adjacency relationship and topological structure between objects in three-dimensional space. The geometric feature data is converted into a three-dimensional spatial tensor and normalized to eliminate the influence of dimensional differences, ensuring that the model can accurately learn the features of objects of different scales. The neighborhood feature aggregation operation of the multi-layer graph convolution layer can generate an adjacency matrix between nodes, and then optimize the topological connections through a dynamic routing mechanism to construct a geometric topological graph that accurately reflects the spatial structure. This process breaks through the limitations of traditional methods that rely on manually designed features, enabling the system to automatically learn the feature representation of complex geometric shapes, providing richer and more accurate spatial information for subsequent path planning, and effectively reducing path conflicts and collisions caused by inaccurate geometric feature descriptions.

[0058] In terms of path density coefficient calculation and spatial principal component analysis, the path density coefficient is calculated by comprehensively considering local and global density quantities, which can fully reflect the spatial density distribution characteristics of the packaging area. The local density quantity is based on the ratio of the spatial density standard deviation of the nodes adjacent to the reference node to the Euclidean distance, reflecting the density variation details of the local area; the global density quantity is measured by the absolute value of the path residual between the reference node and the associated nodes, reflecting the path coherence on a global scale. The harmonic mean of the two is used as the path density coefficient, taking into account both local details and global trends, providing a more reasonable density reference indicator for path planning. The Laplacian matrix of the path density coefficient graph is further spectrally decomposed to extract spatial principal component vectors whose cumulative contribution rate exceeds a threshold, achieving dimensionality reduction of high-dimensional data. While retaining the main feature information, it reduces computational complexity and improves the efficiency of path planning.

[0059] During the generation of path planning weights and optimized paths, the application of a genetic algorithm enables iterative optimization of the spatial principal component vectors. Through normalization and cosine similarity calculation, the principal component vectors are associated with the historical optimal path vectors, ensuring that the generation of initial weights is guided by a certain amount of prior knowledge. The local optimization operation of the adaptive sliding window enables the weights to be dynamically adjusted based on the real-time packing status, enhancing the system's adaptability to complex scenarios. The reference path nodes are dynamically updated based on the path planning weights, and path deviations are corrected using a dynamic planning algorithm. The resulting optimized path can more reasonably allocate path planning weights to each area, reduce path redundancy, and improve the robot's motion efficiency. At the same time, the construction of a three-dimensional packing trajectory space and the real-time detection of path anomalies can promptly identify intersections and conflicting areas. When anomalies are detected, trajectory adjustment instructions are generated, effectively avoiding the risk of collision during the robot's movement and improving the safety and reliability of the packing operation.

[0060] In terms of coordination with the robotic arm's motion control, a mixed integer programming model was established for the path planning parameters and the robotic arm's motion parameters. Using a linear combination of path length and motion energy consumption as the objective function and the robotic arm's joint angle limits as constraints, the optimal motion strategy was solved using a branch-and-bound method. This strategy was then spatially aligned with the optimized path to ensure that the generated path plan was not only geometrically reasonable but also fully accounted for the robotic arm's actual motion capabilities. This combination of path planning and motion control avoids the disconnect between path planning and execution found in traditional methods, enabling the robotic arm to complete packing operations more efficiently and safely, significantly improving packing efficiency and overall system performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a working principle diagram of the deep learning-based packing path planning method of the present invention;

[0062] Figure 2 Workflow diagram for modeling spatial topology;

[0063] Figure 3 Generate a flow chart for the optimization path;

[0064] Figure 4 This is the flow chart of path anomaly detection and adjustment. DETAILED DESCRIPTION

[0065] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0066] See also Figure 1-Figure 4 The present invention relates to a packing path planning method based on deep learning, and the specific implementation steps are as follows:

[0067] The size, weight, and geometric data of the objects to be packed are collected to construct the three-dimensional coordinate parameters of the packing space. Specifically, the length, width, and height of the object are obtained through equipment such as a 3D scanner, and the weight data is obtained using a weighing sensor. Geometric data such as the object's surface curvature and edge features are also collected and mapped into a three-dimensional coordinate system. This data forms three-dimensional coordinate parameters containing information such as the object's position and shape, providing basic spatial information for subsequent path planning.

[0068] A graph convolutional network (GCN) is used to model the spatial topology of geometric feature data, generating a geometric topology graph. The path density coefficient is then calculated based on the spatial density variations between nodes in the topology graph. The geometric feature data is fed into the GCN, where computations are used to mine the spatial relationships between nodes, constructing a geometric topology graph that reflects the geometric structure of the object. Within this topology graph, the spatial density distribution around each node is analyzed, and a specific calculation method is used to derive the path density coefficient, which is used to characterize the complexity of path planning and the distribution of feasible areas.

[0069] The Laplacian matrix of the path density coefficients for different binned areas is spectrally decomposed to extract the spatial principal component vectors. These are then iteratively optimized using a genetic algorithm to generate path planning weights for each area. For each binned area, the path density coefficients are converted into a Laplacian matrix. Spectral decomposition is then used to extract principal component vectors that reflect the key spatial characteristics. Then, leveraging the global search capabilities of the genetic algorithm, the principal component vectors are optimized to generate path planning weights applicable to each area, guiding the direction of path optimization.

[0070] The baseline path nodes are dynamically updated based on the path planning weights to generate an optimized path. The robot's trajectory is then controlled based on this optimized path. Based on the pre-set baseline path nodes and the generated path planning weights, the nodes are adjusted and updated to form a more optimal path node sequence. These node sequences are then connected into a continuous optimized path, which is then converted into control commands for the robot, driving it to complete the packing operation along the planned trajectory.

[0071] The technical solution of the present invention is further described in detail below with reference to specific embodiments.

[0072] Example 1:

[0073] Based on the above overall plan, the specific implementation of spatial topology modeling is as follows: First, the geometric feature data is converted into a three-dimensional space tensor, and dimensional differences are eliminated through normalization. Geometric feature data covers multi-dimensional information such as surface curvature, edge angles, and vertex coordinates. This data may exist in different dimensions (such as length units in meters and angle units in radians). Direct processing will lead to model training bias. Therefore, the geometric feature data needs to be mapped into a three-dimensional space tensor. For example, each geometric feature point is represented as an element in the tensor, and its coordinate value corresponds to the dimension of the tensor. Normalization processing uses linear transformation (such as scaling the data to the [0,1] interval) or standardization methods (such as subtracting the mean and dividing by the standard deviation) to put data of different dimensions on the same numerical scale, ensuring that the subsequent graph convolutional network treats all types of features equally.

[0074] Multi-layer graph convolutional layers are used to aggregate neighborhood features in a three-dimensional spatial tensor to generate an adjacency matrix between nodes. Graph convolutional networks (GCNs) are effective tools for processing non-Euclidean data. Their core concept is to update the features of a current node by propagating features from neighboring nodes. In a three-dimensional spatial tensor, each node represents a characteristic point in the object's geometric structure (e.g., a vertex or extreme point of curvature). The node's initial feature vector contains the geometric properties of that point (e.g., coordinates and curvature value). Each layer of the multi-layer graph convolutional layer performs the following operations: For each node, the feature vectors of its neighboring nodes are collected (the neighborhood range can be determined by a set radius or the k-nearest neighbor algorithm). These features are then aggregated using a weighted summation (weights learned from the convolution kernel parameters) to generate a new feature representation for the node. After multiple layers of convolution, node features not only contain their own geometric properties but also incorporate information about the neighborhood structure. The adjacency matrix is ​​generated based on the spatial correlation between nodes. If the distance between two nodes in three-dimensional space is less than a preset threshold or they are k-nearest neighbors, the corresponding position in the adjacency matrix is ​​set to 1; otherwise, it is set to 0. This matrix reflects the topological connectivity between nodes.

[0075] A dynamic routing mechanism optimizes the topological connections of the adjacency matrix and outputs a geometric topology graph. This dynamic routing mechanism draws on the concept of capsule networks and optimizes the connection strength between nodes by iteratively updating routing weights. The specific process is as follows: During initialization, each node sets the routing weights of its neighboring nodes to the same value. Then, for each node, its predicted feature vector for the neighboring nodes is calculated based on the current routing weights and similarity is matched with the actual feature vectors of the neighboring nodes (e.g., by calculating the dot product). The routing weights are updated based on the similarity results. Higher similarity indicates a greater routing weight, indicating a more important connection between the two nodes. After several iterations, the connection weights in the adjacency matrix no longer change significantly, resulting in the optimized topology. The geometric topology graph is presented as a graph structure, with nodes representing geometric feature points and edges representing optimized adjacency relationships. Edge weights reflect the importance of feature propagation between nodes. This graph provides a structured representation of geometric information for the subsequent calculation of the path density coefficient.

[0076] In specific implementations, the dimensionality of the three-dimensional spatial tensor can be set based on the complexity of the object. For example, for simple geometric shapes, a third-order tensor consisting of three-dimensional coordinates (x, y, z) can be used. For complex objects, higher-order tensors can be formed by adding dimensions such as curvature and normal vectors. The choice of normalization method should be considered based on the data distribution. If the data follows a normal distribution, a normalization method is more appropriate. If the data range is known and the boundary values ​​need to be preserved, linear normalization is more optimal. The number of graph convolutional layers is typically 2-4. Too few layers may not fully capture high-order neighborhood features, while too many layers may lead to overfitting. The number of neighborhood nodes k can be set empirically (e.g., k=8 or k=16) or can be adaptively adjusted based on the curvature of the object's surface. A smaller k value is selected for regions with greater curvature (e.g., corners) to focus on local features, while a larger k value is selected for regions with less curvature (e.g., flat surfaces) to incorporate more neighborhood information. Dynamic routing typically requires 3-5 iterations. Each iteration calculates and updates the routing weights of all nodes, and the iteration is terminated when the weight change is less than a preset threshold (e.g., 0.01).

[0077] Through the above steps, the geometric feature data is transformed from its original scattered state into a structured geometric topology graph. This process not only retains the spatial geometric information of the object, but also enhances the semantic association between nodes through neighborhood feature aggregation and topological connection optimization, providing a more accurate spatial topology basis for the calculation of path density coefficient in subsequent path planning.

[0078] Example 2:

[0079] Based on the above overall scheme, the calculation process of the path density coefficient is as follows: First, a reference path node in any boxed area is selected and the spatial density distribution of its n neighboring nodes is calculated. The selection of the reference path node can be based on the geometric center of the boxed area or the key turning point of the initial path planning to ensure that the calculation is representative. The determination of the n neighboring nodes can be achieved through the k-nearest neighbor algorithm in three-dimensional space. That is, for the reference node, the n nearest nodes are searched in the space formed by the three-dimensional coordinate parameters (n is a preset positive integer, such as n=10 or n=20). The calculation of the spatial density distribution is centered on the reference node, and the distribution frequency of the n neighboring nodes in different spatial directions is counted. For example, the three-dimensional space is divided into several cubic grids, and the number of nodes in each grid is calculated to form a density histogram centered on the reference node. This histogram reflects the degree of aggregation and distribution pattern of the neighboring nodes in space.

[0080] The ratio of the standard deviation of spatial density to the Euclidean distance of the reference node is used as the local density measure. The standard deviation of spatial density is calculated by statistically calculating the density values ​​in the density histogram above. A larger standard deviation indicates a more dispersed spatial distribution of neighboring nodes; conversely, a smaller standard deviation indicates a more concentrated distribution. The Euclidean distance of a reference node refers to the distance between the reference node and the centroid of the set of neighboring nodes. The centroid coordinates are calculated by taking the arithmetic average of the coordinates of the neighboring nodes. The formula for calculating the local density measure is: Local density measure = Spatial density standard deviation / Euclidean distance. This ratio measures the relative size of the node distribution dispersion relative to the spatial distance within the local area centered on the reference node. For example, if the standard deviation is small and the Euclidean distance is large, it means that the nodes in the local area are concentrated and far from the reference node, and the local density is low. If the standard deviation is large and the Euclidean distance is small, it means that the nodes are dispersed but close to the reference node, and the local density is high.

[0081] Calculate the path residual of the reference node and the associated node, and take its absolute value as the global density. The associated node refers to the node that is directly connected to the reference node in the initial path planning. The path residual is the coordinate difference between the reference node and the associated node in three-dimensional space, that is, the coordinate difference in the x, y, and z axes is calculated respectively, and then the size of the path residual is calculated by the vector norm. After taking the absolute value, the global density reflects the degree of spatial difference between the reference node and the associated node on the path. For example, if the path residual is large, it means that the reference node and the associated node are far apart in space, and the path planning needs to consider a larger span, and the global density is higher; if the path residual is small, it means that the spatial distance between the two is close, the path coherence is good, and the global density is low.

[0082] Finally, the harmonic mean of the local and global density measures is used as the path density coefficient. The formula for calculating the harmonic mean is: Path Density Coefficient = 2 × (Local Density × Global Density) / (Local Density + Global Density). The harmonic mean is sensitive to small values, balancing the influence of local and global density measures and avoiding the bias of a single metric. For example, when the local density is high but the global density is low, the harmonic mean will favor the smaller global density, making the path density coefficient more reflective of global path coherence. When the two values ​​are similar, the path density coefficient is the geometric mean of the two, comprehensively reflecting the local and global density characteristics.

[0083] In specific implementations, the selection of reference path nodes can be combined with the binning strategy. For example, the binning space can be divided into several cubic subregions, with the central node of each subregion selected as the reference node. The value of k (i.e., n) in the k-nearest neighbor algorithm needs to be adjusted based on the size and density of the binned objects. For small, densely packed objects, n can be set to a larger value (e.g., n=30) to capture the distribution characteristics of more neighboring nodes; for large, sparsely packed objects, n can be set to a smaller value (e.g., n=5) to avoid including irrelevant nodes. The cubic grid granularity of the spatial density distribution can be set based on the required geometric accuracy of the objects. For example, for high accuracy, the grid side length can be set to 1 / 10 of the minimum object size; for lower accuracy, it can be set to 1 / 5 of the average object size. The calculation of path residuals must take into account the initial direction of the path plan. For example, in a Cartesian coordinate system, the x-axis is the direction of the robot's motion, and the y and z-axes are perpendicular to it. The residual components of each axis are calculated separately to ensure that the global density metric accurately reflects the spatial differences in the path direction.

[0084] When calculating the harmonic mean, care must be taken to avoid situations where the denominator is zero. When both the local and global density measures are zero (i.e., the reference node has no neighboring nodes or associated nodes, a situation that typically occurs at isolated nodes or at the start or end of a path), the path density coefficient can be set to a preset minimum value (e.g., 0.1) to ensure algorithm robustness. Furthermore, the dimensions of the local and global density measures must be consistent. This can be achieved through normalization (e.g., scaling both to the [0, 1] range) to ensure the physical meaning of the harmonic mean is clear.

[0085] Through the above steps, the path density coefficient can comprehensively reflect the local spatial distribution characteristics and global path connection characteristics of the benchmark nodes in the boxed area, providing a quantitative basis for the subsequent construction of the graph Laplacian matrix and the extraction of spatial principal component vectors.

[0086] Example 3:

[0087] On the basis of the above-mentioned overall scheme, regarding the extraction of spatial principal component vectors, the specific steps are as follows: first, construct a graph Laplace matrix of the path density coefficients of different binning areas, and perform spectral clustering decomposition on it to obtain eigenvectors. The path density coefficient matrix is ​​composed of the path density coefficients of each reference node calculated in Example 2. The rows of the matrix correspond to different binning areas, and the columns correspond to the reference nodes within the area. The construction of the graph Laplace matrix is ​​based on this coefficient matrix. The specific process is: first calculate the similarity matrix between nodes. The similarity is defined as the Gaussian kernel function value of the path density coefficient of the two nodes, that is, similarity = exp(-|| coefficient i - coefficient j||² / (2σ²)), where σ is the bandwidth parameter, which can be automatically determined by the median of the coefficient difference of all node pairs. Then construct a degree matrix, whose diagonal elements are the sum of the similarities of the corresponding nodes, and the non-diagonal elements are zero. Finally, the graph Laplace matrix is ​​defined as the degree matrix minus the similarity matrix. Perform spectral clustering decomposition on the matrix, that is, solve its characteristic equation to obtain the eigenvalues ​​and corresponding eigenvectors.

[0088] The principal component subspace is constructed by selecting eigenvectors whose cumulative contribution rates, corresponding to eigenvalues, exceed a set threshold. Eigenvalues ​​are arranged in descending order, and the cumulative contribution rate is the ratio of the sum of the first k eigenvalues ​​to the sum of all eigenvalues. After setting a threshold (such as 85% or 90%), the smallest k is selected so that the cumulative contribution rate exceeds the threshold, and the corresponding k eigenvectors form the principal component subspace. This space retains the key information of the original path density coefficient matrix while achieving data dimensionality reduction. For example, if the original matrix has 100 nodes, 100 eigenvectors may be obtained through spectral decomposition, but only the first 10 eigenvectors have a cumulative contribution rate exceeding 90%. In this case, these 10 eigenvectors are selected to form the principal component subspace, reducing the data dimension from 100 to 10.

[0089] Finally, the graph Laplacian matrix is ​​projected onto the principal component subspace to obtain the principal component vectors of the reduced-dimensional space. This projection is accomplished through matrix multiplication, which involves multiplying the graph Laplacian matrix with the eigenvector matrix of the principal component subspace. The resulting vectors are the principal component vectors of the reduced-dimensional space. These vectors retain the key features of the original data while removing noise and redundant information. This allows subsequent genetic algorithm optimization to be performed in a lower-dimensional space, improving computational efficiency and optimization accuracy.

[0090] In specific implementations, the similarity matrix can be calculated using a parallel computing strategy, assigning node pairs to multiple computing units for simultaneous processing to accelerate computation. The selection of the bandwidth parameter σ is crucial for constructing the similarity matrix. A too large value will cause the similarities between all nodes to approach 1, while a too small value will make the similarity matrix approach a sparse matrix, both of which are detrimental to spectral decomposition. The initial σ value can be calculated by taking the median of the coefficient differences between all node pairs and fine-tuning it based on the data distribution.

[0091] Spectral clustering can be used to solve for eigenvalues ​​and eigenvectors using iterative methods such as the power method and the Lanczos algorithm. These methods are highly efficient when processing large matrices. When selecting eigenvectors whose cumulative contribution exceeds a threshold, care should be taken to avoid selecting too many eigenvectors, which may lead to ineffective dimensionality reduction, and to avoid selecting too few eigenvectors, which may result in loss of important information.

[0092] Projection operations into the principal component subspace must ensure that the matrix dimensions match, which can be verified using matrix transposition and multiplication rules. The reduced spatial principal component vectors can be visualized, for example, by plotting two-dimensional or three-dimensional scatter plots to observe the distribution of different binned regions in the principal component space, providing intuitive guidance for subsequent genetic algorithm optimization.

[0093] The entire spatial principal component vector extraction process is based on rigorous mathematical operations and data processing steps, ensuring the accuracy and repeatability of the conversion from path density coefficients to principal component vectors. Dimensionality reduction not only reduces data redundancy but also highlights key features, providing more effective input data for subsequent path planning weight optimization.

[0094] Example 4:

[0095] Based on the above overall scheme, the process of generating path planning weights is as follows: First, the spatial principal component vector is standardized and its cosine similarity with the historical optimal path vector is calculated. The spatial principal component vector is extracted by Example 3, and its dimension has been reduced to the principal component subspace (such as 10 dimensions) through spectral decomposition. The standardization is achieved by subtracting the vector mean and dividing it by the standard deviation, so that the numerical distribution of each dimension has zero mean and unit variance, avoiding the influence of dimensional differences between dimensions on the similarity calculation. The historical optimal path vector is stored in the system database. It is a set of principal component vectors with the best optimized path performance in the historical packing task, reflecting the characteristic patterns of past successful cases. The cosine similarity is obtained by calculating the dot product of two vectors divided by the module length product. The value range is [-1,1]. The larger the value, the more similar the vector direction is, that is, the higher the characteristic matching degree between the current principal component vector and the historical optimal path.

[0096] Similarity is input into the genetic algorithm population, and initial weights are generated through selection, crossover, and mutation. The genetic algorithm population consists of multiple individuals, each of which represents a set of candidate weight parameters (e.g., a vector of the same dimension as the principal component vector, where each element represents the weight coefficient of the corresponding principal component). The initial population can be generated randomly or by subjecting historical optimal weights to small perturbations. Selection uses a roulette wheel or tournament selection method to select high-quality individuals based on their fitness (e.g., similarity), with individuals with higher fitness being more likely to be selected. Crossover is implemented using single-point crossover or uniform crossover, combining the weight parameters of the selected individuals to generate offspring individuals. For example, single-point crossover randomly selects a crossover point in the weight vector and swaps the parameter values ​​of the two parent individuals after that point. Mutation randomly perturbs the individual weight parameters with a low probability (e.g., 0.01), such as by adding a small random number from a normal distribution, to maintain population diversity and avoid falling into local optima. After several generations (e.g., 50 generations), initial weights with high fitness are generated.

[0097] Finally, an adaptive sliding window is used to locally optimize the initial weights, outputting the path planning weights for each region. The size of the adaptive sliding window is dynamically adjusted based on the distribution density of the principal component vectors. A larger window (e.g., containing 10 individuals) is used in areas where weights vary gently to accelerate optimization, while a smaller window (e.g., containing 3 individuals) is used in areas where gradients vary sharply to allow for finer adjustments. The optimization algorithm within the window can employ a local search method. For example, multiple candidate solutions are generated within the neighborhood of the current weight vector (e.g., within a range of plus or minus 0.1 for each parameter). The corresponding path planning performance evaluation values ​​(e.g., a combined index of path length and smoothness) are calculated, and the optimal solution is selected as the local optimal solution within the window. The window is gradually slid across the weight parameter space until all regions are traversed or a stopping condition is met (e.g., the optimal solution within five consecutive window generations remains unchanged). The resulting path planning weights for each region are a set of coefficients corresponding to the dimensions of the principal component vectors, which are used to adjust the update direction and amplitude of the reference path nodes.

[0098] In specific implementations, normalization requires preserving the mean and standard deviation parameters from the training phase to allow consistent normalization of new data during the testing phase. The storage of historical optimal path vectors must include corresponding binning scenario labels (such as object size range and binning space shape) to match the most similar historical vectors in the current scenario. The genetic algorithm population size is typically set to 50-200 individuals, and the number of iterations is typically set to 50-100 generations, depending on computing resources and accuracy requirements. The crossover probability is typically set to 0.6-0.9, and the mutation probability to 0.001-0.01, to balance search speed and global exploration capabilities.

[0099] The dynamic adjustment mechanism of the adaptive sliding window is achieved by calculating the variance of the weight parameters within the current window. A larger variance indicates a more dispersed distribution, requiring a narrower window to focus on local details; a smaller variance indicates a more concentrated distribution, requiring a wider window to explore a wider area. The calculation of the path planning performance evaluation value must be consistent with the subsequent optimization path generation objectives, such as minimizing path length and minimizing the number of turns. A comprehensive evaluation function can be constructed, but without the need for actual experimental data, relative values ​​can be calculated based solely on the mathematical relationships between parameters.

[0100] The entire path planning weight generation process combines historical experience data with intelligent optimization algorithms. It unifies the data scale through standardization, uses the global search capability of genetic algorithms to explore the weight space, and then performs local fine-tuning through adaptive sliding windows to ensure that the generated weights can both draw on historical successful models and adapt to the unique needs of the current packing scenario.

[0101] Example 5:

[0102] Based on the above overall scheme, the generation of the optimized path and the related extension steps are as follows: the path planning weights are tensor-multiplied with the reference path nodes to obtain the node update parameters. The path planning weights are generated by Example 4, which are a set of coefficients corresponding to the dimensions of the principal component vector (such as a weight vector with a dimension of 10). The reference path nodes are discrete points in the initial planning path, and each node contains three-dimensional coordinates (x, y, z) and attributes such as timestamps. The tensor multiplication operation multiplies the coordinate components of each reference path node by the corresponding weight coefficient element by element. For example, for the node coordinate vector and weight vector , node update parameters The calculation formula is:

[0103]

[0104] Among them: When the weight value When , the coordinate value of the corresponding dimension is magnified; when the weight value When , reduce the coordinate value of the corresponding dimension; Represents the element-wise multiplication (Hadamard product) of vectors. It reflects the adjustment amplitude and direction of the weight on the node coordinates.

[0105] Calculate the deviation between the node update parameters and the reference path, and use the dynamic programming algorithm to correct the deviation. The deviation is the difference between the updated node coordinates and the original reference path node coordinates, that is, The dynamic programming algorithm is used to optimize the adjusted node sequence under the premise of considering the path continuity and the motion constraints of the robot arm. The specific steps are: arrange the reference path nodes in chronological order to form a state space, each state is the updated coordinates of the node and the cumulative deviation value; define the transfer cost as the weighted sum of the Euclidean distance and the deviation change between adjacent nodes (the weight can be set according to the path smoothness requirement); recursively calculate the minimum cost path from the start point to the end point to obtain the corrected node coordinate sequence. For example, for a path containing n nodes, the dynamic programming table It represents the minimum cumulative cost when the i-th node among the first i nodes is in state j. The optimal solution is obtained by traversing all possible state transitions.

[0106] The corrected node sequence is connected into a continuous trajectory to generate an optimized path. The corrected node sequence is arranged sequentially on the time axis, and adjacent nodes are interpolated using cubic spline interpolation or Bezier curve fitting to form a smooth curve, ensuring that the robot's motion trajectory has no abrupt changes. The interpolation process must meet the robot's kinematic constraints, such as maximum joint velocity and acceleration limits. Interpolation parameters (such as the spline curve tension coefficient) are adjusted to ensure that the trajectory meets physically feasible conditions. The resulting optimized path is represented as a continuous sequence of coordinate points and corresponding time parameters, which can be directly used as input instructions for the robot's motion control.

[0107] The following steps are also included: constructing a 3D packing trajectory space based on the optimized path, extracting intersections and conflict areas in the trajectory. The specific implementation is: mapping the optimized path to a 3D coordinate system according to the motion sequence, and generating a trajectory point cloud set with timestamps ,in is the trajectory point at time t. A point cloud reconstruction algorithm (such as the voxelized grid method) is used to perform 3D voxel modeling on the trajectory point cloud. The 3D space is divided into uniform cubic voxels. The voxel units are filled according to the distribution of the trajectory points to form a voxel model that contains the trajectory shape. The surface curvature of the voxel model is calculated, and the normal vector of each voxel surface point is estimated using the finite difference method. The Gaussian curvature and mean curvature are then calculated. The curvature value reflects the degree of curvature of the trajectory surface. The curvature distribution is enhanced using a non-maximum suppression algorithm. That is, for each voxel point, if its curvature value is not a local maximum, it is suppressed (set to zero). Points with significant curvature are retained, eliminating noise interference, and highlighting feature areas such as sharp corners and intersections in the trajectory.

[0108] When the number of intersections exceeds the set threshold or the area of ​​the conflict zone is larger than the limit, it is determined to be a path abnormality state and a trajectory adjustment instruction is generated. Intersection detection traverses the trajectory point cloud set and calculates the spatial distance and time difference between any two points. If the distance is less than the preset threshold (such as 1.5 times the size of the robot end effector) and the time difference is less than the minimum safe time interval, it is determined to be a potential intersection. The conflict area is defined as a continuous voxel area in the voxel model where the curvature value exceeds the threshold, and its area is calculated by counting the number of conflicting voxels. If the number of intersections exceeds the threshold (such as 5) or the area of ​​the conflict zone is larger than the limit (such as 100 voxel units), the exception handling mechanism is triggered, and a trajectory adjustment instruction including the adjustment direction and amplitude is generated, returning to the path correction stage for re-optimization.

[0109] A mixed integer programming model of path planning parameters and robot arm motion parameters is established, and the branch-and-bound method is used to solve the optimal motion strategy. The variables of the mixed integer programming model include the discrete coordinates of the path nodes (integer variables) and the robot arm joint angles (continuous variables). The objective function is defined as a linear combination of the path length and motion energy consumption, that is, ,in is the total path length, is the integrated energy consumption of the joint torque, and is a weight coefficient (which can be set based on the optimization goals of packing efficiency and energy consumption). Constraints include the robot arm's joint angle limits, velocity limits, and acceleration limits. When solving the problem using the branch-and-bound method, the objective function is first relaxed, ignoring the integer constraints to obtain a continuous relaxation problem, and the lower bound is obtained by solving it. The integer variables are then branched into different subproblems, and the lower bounds and feasible solutions of each subproblem are recursively solved. Branches with lower bounds higher than the current optimal solution are pruned until an optimal solution that satisfies the integer constraints is found.

[0110] The optimal motion strategy is spatially aligned with the optimized path to generate a path plan that optimizes packing efficiency. This spatial alignment is achieved through coordinate transformation, mapping the optimal solution in the robot's joint space to Cartesian space, matching it with the coordinate points of the optimized path, and adjusting the trajectory's timing parameters to ensure motion synchronization. The resulting path plan includes the optimized trajectory coordinates, joint angle sequence, and timing plan, ensuring that the robot completes the packing operation using the shortest path and with the lowest energy consumption while meeting physical constraints.

[0111] This embodiment utilizes techniques such as tensor operations, dynamic programming, voxel modeling, and mixed integer programming to achieve a complete process, from path weighting to optimized trajectory generation, anomaly detection, and strategy optimization. Each step is based on data processing and algorithmic logic, avoiding the introduction of hypothetical experimental effect descriptions. This approach complies with the technical details of the specific implementation methods specified in the patent application, and utilizes multi-layered validation mechanisms (such as intersection detection and constraint solving) to ensure path feasibility and optimized efficiency.

[0112] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0113] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A packing path planning method based on deep learning, characterized in that: include: Collect the size data, weight data and geometric feature data of the objects to be packed, and construct the three-dimensional coordinate parameters of the packing space; A graph convolutional network is used to perform spatial topological modeling on geometric feature data, generate a geometric topological graph, and calculate the path density coefficient based on the spatial density changes between nodes in the topological graph; The Laplacian matrix of the path density coefficients of different binned areas is spectrally decomposed to extract the spatial principal component vectors. The principal component vectors are iteratively optimized using a genetic algorithm to generate the path planning weights for each area. The reference path nodes are dynamically updated according to the path planning weights to generate an optimized path, and the motion trajectory of the robotic arm is controlled based on the optimized path.

2. A packing path planning method based on deep learning according to claim 1, characterized in that: The spatial topology modeling includes: Convert the geometric feature data into a three-dimensional space tensor and eliminate the dimension difference through normalization; Multi-layer graph convolution layers are used to aggregate neighborhood features of the three-dimensional spatial tensor to generate an adjacency matrix between nodes; The topological connection of the adjacency matrix is ​​optimized through a dynamic routing mechanism, and a geometric topological graph is output.

3. The packing path planning method based on deep learning according to claim 1, characterized in that: The path density coefficient calculation includes: Select a reference path node in any boxed area and calculate the spatial density distribution of its n neighboring nodes; The ratio of the standard deviation of spatial density to the Euclidean distance of the reference node is used as the local density measure; Calculate the path residual between the reference node and the associated node, and take its absolute value as the global density; The harmonic mean of the local density and the global density is taken as the path density coefficient.

4. The packing path planning method based on deep learning according to claim 1, characterized in that: The spatial principal component vector extraction includes: Construct the graph Laplacian matrix of path density coefficients of different binned areas, and perform spectral clustering decomposition on it to obtain eigenvectors; The eigenvectors whose cumulative contribution rates corresponding to the eigenvalues ​​exceed the set threshold are selected to form the principal component subspace; Project the graph Laplacian matrix onto the principal component subspace to obtain the principal component vector of the reduced-dimensional space.

5. The packing path planning method based on deep learning according to claim 1, characterized in that: The path planning weight generation includes: Normalize the spatial principal component vector and calculate its cosine similarity with the historical optimal path vector; The similarity is input into the genetic algorithm population and the initial weights are generated through selection, crossover and mutation operations; The initial weights are locally optimized through an adaptive sliding window, and the path planning weights of each area are output.

6. The packing path planning method based on deep learning according to claim 1, characterized in that: The generating optimization path comprises: Multiply the path planning weight by the reference path node tensor to obtain the node update parameter; Calculate the deviation between the node update parameters and the reference path, and use the dynamic programming algorithm to correct the deviation; Connect the corrected node sequences into a continuous trajectory to generate an optimized path.

7. The packing path planning method based on deep learning according to claim 1, characterized in that: Also includes: Construct a three-dimensional packing trajectory space based on the optimized path and extract the intersection points and conflict areas in the trajectory; When the number of intersections exceeds the set threshold or the area of ​​the conflict zone is larger than the limit, it is determined to be a path abnormality state and a trajectory adjustment instruction is generated.

8. The method for packing path planning based on deep learning according to claim 7, characterized in that: The three-dimensional packing trajectory space construction includes: Map the optimized path to a three-dimensional coordinate system according to the motion sequence to generate a trajectory point cloud set; The point cloud reconstruction algorithm is used to perform three-dimensional voxel modeling on the trajectory point cloud and calculate the surface curvature of the voxel model; The curvature distribution is enhanced by the non-maximum suppression algorithm to eliminate noise interference.

9. The method for packing path planning based on deep learning according to claim 1, characterized in that: Also includes: A mixed integer programming model of path planning parameters and robot arm motion parameters is established, and the branch and bound method is used to solve the optimal motion strategy. Spatially align the optimal motion strategy with the optimized path to generate a path plan with the best packing efficiency; The mixed integer programming model solution includes: The objective function is defined as a linear combination of path length and motion energy consumption, and the constraint condition is the limit of the robot arm joint angle; Perform relaxation transformation on the objective function and decompose it into path parameter subproblem and motion parameter subproblem; The two sub-problems are solved alternately and iteratively until convergence, and the optimal motion strategy that satisfies the constraints is output.

10. A packing path planning system based on deep learning, characterized in that: include: 3D data acquisition module, used to obtain the size, weight and geometric feature data of the objects to be packed and construct a 3D space coordinate matrix; The topology modeling module uses a graph convolutional network to aggregate neighborhood features of the three-dimensional spatial coordinate matrix and generates a geometric topology graph through a dynamic routing mechanism; The density coefficient calculation module generates local density and global density according to the spatial density distribution of nodes in the geometric topology graph, and calculates their harmonic mean as the path density coefficient; The spectral decomposition module constructs a graph Laplacian matrix for the path density coefficients of different binned areas and extracts the spatial principal component vectors through spectral clustering decomposition; The weight optimization module uses a genetic algorithm to iteratively optimize the spatial principal component vector to generate the path planning weights for each area; The path generation module dynamically updates the reference path nodes based on the path planning weights, corrects the path deviation through the dynamic planning algorithm, and generates the optimized path; The motion control module converts the optimized path into control instructions for the robotic arm and drives the actuator to complete the packing operation through a real-time communication protocol.

Images