Intelligent warehouse multi-agv cooperative path planning method, device, system and medium

Abstract

The application discloses a smart warehouse multi-AGV cooperative path planning method, device and system and a medium, and relates to the technical field of path planning. The path planning method comprises the following steps: obtaining carrying task information of a target AGV to determine a set of work points to be visited by the AGV. Topological information of a warehouse channel is obtained, and the shortest driving distance between any two work points is calculated to obtain a distance matrix. A traveling salesman problem instance is constructed by taking the work points as nodes and the distance as edge weight. The m nearest neighbor nodes of each node in the distance matrix in the traveling salesman problem instance are obtained, and a local neighborhood matrix of each node is obtained. The local neighborhood matrix is input into a pre-trained anchor point prediction network to predict whether each node is an anchor point, so that all nodes are divided into an anchor point set and a non-anchor point set. Based on the distance matrix, the anchor point set and the non-anchor point set, a planning path is constructed through an anchor point insertion algorithm. The planning path is converted into a carrying path scheduling instruction of the AGV.

Description

Technical Field

[0001] This invention relates to the field of path planning technology, and more specifically, to a method, apparatus, system, and medium for collaborative path planning of multiple AGVs in intelligent warehousing. Background Technology

[0002] In intelligent warehousing systems, Automated Guided Vehicles (AGVs) need to efficiently complete transport tasks between multiple work points to improve warehouse throughput and operational efficiency. In this scenario, the AGV path planning problem is often abstracted as the Traveling Salesman Problem (TSP), which involves finding the shortest sequence of visits given nodes and a distance matrix. However, factors such as complex warehouse layouts, dispersed storage locations, and real-time scheduling requirements make it difficult to guarantee path quality using traditional manual experience or simple rules. This can easily lead to excessively long AGV travel distances, aisle congestion, and unstable task completion times.

[0003] To address the aforementioned issues, existing technologies offer various methods for solving the TSP (Tracking Strategies for Practical Applications). While exact algorithms can obtain globally optimal solutions, they incur enormous computational costs, making them unsuitable for large-scale instances and online scheduling. Heuristic and metaheuristic algorithms (such as nearest neighbor and insertion algorithms) strike a trade-off between efficiency and solution quality, but they rely on manually designed rules and have limited generalization capabilities. In recent years, supervised learning-based methods have been introduced, such as anchor point prediction-enhanced heuristics, which utilize local neighborhood information to train models to predict key nodes, guiding path construction and thus rapidly generating approximate solutions in structurally similar instances.

[0004] However, existing anchor prediction methods still have significant limitations. These methods require generating and labeling a large number of training instances (e.g., TSP25, TSP50, and TSP100 scales), resulting in high costs for data construction and model training. Their network structure primarily relies on fully connected layers, failing to explicitly mine the correlation between local spatial structure and channels, and lacking sufficient representation capabilities for complex geometric layouts. The anchor insertion strategy in the path construction stage is relatively coarse, and the handling of non-anchor points relies on a single incremental criterion, making the solution process sensitive to stochastic decisions, resulting in poor path quality stability and difficulty in adapting to the needs of large-scale, heterogeneous tasks in intelligent warehousing. Summary of the Invention

[0005] The present invention provides a method, apparatus, system and medium for collaborative path planning of multiple AGVs in intelligent warehousing, in order to improve at least one of the above-mentioned technical problems.

[0006] In a first aspect, the present invention provides a method for collaborative path planning of multiple AGVs in intelligent warehousing, which includes steps S1 to S6.

[0007] S1. Obtain the transport task information of the target AGV to determine the set of work points that the AGV needs to visit.

[0008] S2. Obtain the warehouse aisle topology information and calculate the shortest travel distance between any two work points to obtain the distance matrix. Using work points as nodes and distances as edge weights, construct a Traveling Salesman Problem instance.

[0009] S3. Obtain the m nearest neighbors of each node in the distance matrix of the Traveling Salesman Problem instance, and the local neighborhood matrix of each node.

[0010] S4. The local neighborhood matrix is ​​input into a pre-trained anchor prediction network to predict whether each node is an anchor, thereby dividing all nodes into an anchor set and a non-anchor set. The anchor prediction network uses a network structure containing convolutional layers and channel attention layers to extract features and classify the local neighborhood matrix.

[0011] S5. Based on the distance matrix, the set of anchor points and the set of non-anchor points, the planned path is constructed through the anchor point insertion algorithm.

[0012] S6. Convert the planned path into a AGV transport path scheduling instruction.

[0013] Secondly, the present invention provides an intelligent warehouse multi-AGV collaborative path planning device, which includes a task acquisition module, a topology acquisition module, a neighbor matrix module, an anchor point prediction module, an anchor point insertion module, and a path conversion module.

[0014] The task acquisition module is used to obtain the transport task information of the target AGV in order to determine the set of work points that the AGV needs to visit.

[0015] The topology acquisition module is used to obtain warehouse aisle topology information and calculate the shortest travel distance between any two work points to obtain a distance matrix. Using work points as nodes and distances as edge weights, a Traveling Salesman Problem instance is constructed.

[0016] The Neighbor Matrix module is used to obtain the m nearest neighbors of each node in the distance matrix in the Traveling Salesman Problem instance, as well as the local neighborhood matrix of each node.

[0017] The anchor point prediction module is used to input the local neighborhood matrix into a pre-trained anchor point prediction network to predict whether each node is an anchor point, thereby dividing all nodes into an anchor point set and a non-anchor point set. The anchor point prediction network employs a network structure including convolutional layers and channel attention layers to extract features and classify the local neighborhood matrix.

[0018] The anchor insertion module is used to construct a planned path based on the distance matrix, the set of anchor points, and the set of non-anchor points using an anchor insertion algorithm.

[0019] The path conversion module is used to convert the planned path into AGV transport path scheduling instructions.

[0020] Thirdly, the present invention provides an intelligent warehousing multi-AGV collaborative path planning system, which includes a path planning server and at least two AGVs. The intelligent warehousing multi-AGV collaborative path planning system is used to execute the intelligent warehousing multi-AGV collaborative path planning method described in any paragraph of the first aspect.

[0021] Fourthly, the present invention provides a computer-readable storage medium. The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the intelligent warehousing multi-AGV collaborative path planning method described in any paragraph of the first aspect.

[0022] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0023] This invention presents an intelligent warehousing multi-AGV collaborative path planning method based on anchor point enhancement heuristics. It combines the Anchor Point Channel Attention Prediction Network (CEAP) with the Anchor Point Multi-Factor Insertion Algorithm (ARC). By abstracting the AGV transport task into a TSP (Traffic Path Provider) model, it can generate high-quality transport paths even with limited training data. This method represents a significant improvement over existing technologies in terms of training data cost, anchor point prediction accuracy, path construction quality, and algorithm scalability. Attached Figure Description

[0024] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the specific embodiments of the present invention will be briefly introduced below. It should be understood that the following drawings only show some specific embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating a multi-AGV collaborative path planning method for intelligent warehousing.

[0026] Figure 2 It is the structure of an intelligent warehousing AGV handling path planning system.

[0027] Figure 3 This is a network structure diagram of a multi-AGV collaborative path planning method for intelligent warehousing.

[0028] Figure 4 It is the planning result of a classic example of the internationally recognized Traveling Salesman Problem. Detailed Implementation

[0029] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.

[0030] Example 1, please refer to Figures 1 to 4 The first embodiment of the present invention provides a method for collaborative path planning of multiple AGVs in intelligent warehousing, which can be executed by an intelligent warehousing multiple AGV collaborative path planning device (hereinafter referred to as: path planning device). In particular, it is executed by one or more processors in the path planning device to realize S1 to S6.

[0031] S1. Obtain the transport task information of the target AGV to determine the set of work points that the AGV needs to visit.

[0032] S2. Obtain the warehouse aisle topology information and calculate the shortest travel distance between any two work points to obtain the distance matrix. Using work points as nodes and distances as edge weights, construct a Traveling Salesman Problem instance.

[0033] Specifically, in the intelligent warehousing AGV handling path planning scenario, this embodiment regards each to-be-visited storage location or work point as a TSP node, and uses the shortest travel distance between nodes calculated based on the warehouse channel network as the edge weight to construct a Traveling Salesman Problem instance.

[0034] In a simplified implementation, work points can be treated as nodes on a plane. Euclidean distances are calculated based on their coordinates and appropriately discretized, then directly entered into the distance matrix for training and inference of the anchor prediction network. In another implementation, the shortest path length satisfying constraints such as one-way and no-entry restrictions between any two work points can be calculated first on the warehouse aisle topology map, and then this length is written into the distance matrix. The anchor prediction and path construction algorithm in this embodiment is designed for the distance matrix itself and does not limit the specific construction method of the distance matrix.

[0035] S3. Obtain the m nearest neighbors of each node in the distance matrix of the Traveling Salesman Problem instance, and the local neighborhood matrix of each node.

[0036] Specifically, each node's local neighborhood consists of its m nearest neighbors in the full distance matrix, forming an m×m local distance matrix. This local neighborhood matrix serves as the input feature for the anchor prediction network. For example, when m=5, the input feature for each node is a 5×5 matrix. During label generation, for each node, whether it is an anchor is used as the binary classification label.

[0037] In the actual deployment phase, the intelligent warehousing multi-AGV collaborative path planning method of this embodiment runs on the path planning server of the intelligent warehousing system.

[0038] The Warehouse Management System (WMS) sends the currently pending handling tasks (including the starting location, destination location, task priority, etc.) to the route planning server. The handling task information obtained from the WMS includes at least the starting location, destination location, and corresponding work point identifier of the goods to be handled, which are used to construct the set of work points that a single AGV handling task needs to access.

[0039] The Warehouse Control System (WCS) or warehouse layout database provides warehouse aisle topology information (aisles, intersections, restricted areas, etc.) or planar coordinate information of each work point. Information such as warehouse aisle structure, rack aisle location, and restricted areas, read from the WCS or warehouse layout database, is used to construct the warehouse aisle network and calculate the shortest travel distance between each work point.

[0040] Based on the above information, the path planning server selects key work points (such as the starting buffer, several shelf ports, and the ending exit) that each AGV needs to visit sequentially, and calculates the shortest travel distance between any two work points on the warehouse aisle topology, constructing the corresponding distance matrix. Using work points as nodes and distances as edge weights, a Traveling Salesman Problem instance (i.e., an abstract path planning model of the AGV's handling task) is constructed for subsequent anchor point prediction and path construction.

[0041] S4. The local neighborhood matrix is ​​input into a pre-trained anchor prediction network to predict whether each node is an anchor, thereby dividing all nodes into an anchor set and a non-anchor set. The anchor prediction network uses a network structure containing convolutional layers and channel attention layers to extract features and classify the local neighborhood matrix.

[0042] The Channel-Enhanced AnchorPrediction Network (CEAP) in this embodiment takes the local neighborhood distance matrix of the node as input and adopts an overall structure of "convolution + channel attention + global pooling + multi-layer fully connected".

[0043] Specifically, such as Figure 3As shown, the m×m local neighborhood distance submatrix (m=5 in this embodiment) corresponding to each node is input into the network. First, it is reshaped into an m×m×1 single-channel feature map through a reshape layer. Then, it passes through two convolutional blocks, each consisting of a 2D convolutional layer (Conv2D), a LeakyReLU activation layer, a channel attention layer (ECALayer), and a layer normalization layer, used to extract and reweight the local spatial structure and inter-channel correlation features. After the two convolutional blocks, a global average pooling layer (GlobalAveragePooling2D) is introduced to aggregate the spatial dimensions, compressing the m×m×C feature map into a C-dimensional channel vector. This channel vector then passes through a 128-dimensional fully connected layer and several 256-dimensional fully connected layers. Each fully connected layer is followed by a LeakyReLU activation layer, a layer normalization layer, and a dropout layer to enhance the network's non-linear expressive power and suppress overfitting. Finally, an output neuron provides the predicted probability of whether the current node is an anchor point. Preferably, the output layer uses the sigmoid activation function.

[0044] The channel attention layer performs global average pooling on the input feature map, then learns the correlation between channels through one-dimensional convolution to generate channel attention weights, which are then used to further weight the feature map. In the channel attention layer, the convolution output feature map is first processed... Perform global average pooling along the spatial dimension to obtain the channel average vector (channel description vector). This vector is then subjected to a one-dimensional convolution and a sigmoid activation to generate the attention coefficients for each channel. Finally, the attention weights are broadcast back to the original spatial dimension by channel, and the input features are weighted channel by channel to obtain the weighted feature map.

[0045] .

[0046] In the formula This is the weighted feature map. and These represent the horizontal and vertical coordinates of the feature map, respectively. For channel indexing. For the first Attention coefficients for each channel. This is a feature map.

[0047] Through this operation, the network can adaptively highlight channel features that are more sensitive to anchor point discrimination and suppress redundant channels.

[0048] After completing two convolutional layers and channel weighting, a global average pooling is performed on the feature map to compress the spatial dimension into a channel vector, which is then used for classification in subsequent fully connected layers.

[0049] .

[0050] In the formula For the first Aggregate representation of channels. and The height and width of the feature map. This is the weighted feature map. and These represent the horizontal and vertical coordinates of the feature map, respectively. For channel indexing.

[0051] After global pooling, the features are compressed into channel vectors, and then anchored for classification using several fully connected (Dense) layers. The final output layer uses the Sigmoid activation function to achieve binary classification prediction.

[0052] Compared to traditional methods that input the neighborhood matrix into the MLP, CEAP explicitly models the local spatial structure through convolution and attention mechanisms, improving the accuracy and generalization ability of anchor point prediction.

[0053] This embodiment introduces the Anchor Point Channel Attention Prediction Network (CEAP), replacing the large-scale training of the original method, which relied on 1 million TSP25 sets, 500,000 TSP50 sets, and 250,000 TSP100 sets totaling approximately 75 million feature samples, with a small-scale training using only 50,000 TSP50 sets to generate instances and approximately 2.5 million feature samples. Under the conditions of similar number of training rounds and the same total training time, relying on convolution + channel attention to explicitly model the correlation between local geometry and channels, it still maintains a high anchor point prediction accuracy, precision, and recall rate, thereby substantially reducing the resource consumption of training data construction and model training.

[0054] During the prediction phase, the Channel-Enhanced AnchorPrediction Network (CEAP) predicts the anchor probability for all nodes of a given Traveling Salesman Problem (TSP) instance, and divides the nodes into anchor sets based on a threshold. Non-anchor set The entire path is then constructed using the Anchor Regret-Cover heuristic algorithm.

[0055] S5. Based on the distance matrix, anchor point set, and non-anchor point set, an AGV planning path (i.e., the node access sequence of the transport path) is constructed through an anchor point insertion algorithm. The node access sequence defines the order in which the AGV should reach the work points at the higher-level task level. The specific travel path and obstacle avoidance control between work points are calculated and executed by the WCS and the onboard navigation control system, combined with the warehouse passage topology and the environmental perception results from the LiDAR / obstacle avoidance algorithm. This invention does not limit the specific implementation of this part.

[0056] Preferably, step S5 includes sub-steps S51 to S54.

[0057] S51. Path initialization: Select the pair of nodes that are furthest apart from the set of anchor points or all nodes as the initial path.

[0058] To avoid the path initially concentrating in a local area, the anchor point insertion algorithm first checks the set of anchor points. The system selects the pair of nodes furthest apart as the starting skeleton. If the number of anchor points is too small, it degenerates into using the entire set of nodes. Select from the options. Assume... or The farthest point pair is then determined through the following optimization problem.

[0059] .

[0060] In the formula and These are the pairs of nodes that are furthest apart. This represents the value of the independent variable that maximizes the objective function. For nodes With nodes The distance. The set of candidate nodes (i.e., the set of anchor points) or the set of all nodes ).

[0061] Then initialize the path as follows: This initialization method ensures that the initial path is "expanded" as much as possible in space, leaving more geometric space for subsequent insertions and helping to reduce the overall path length.

[0062] S52, Anchor Insertion Phase: For each anchor point in the anchor point set that has not yet been inserted into the path, calculate the normalized incremental cost, regret value, and coverage resulting from its insertion into each position of the current path. Construct a comprehensive score based on the normalized incremental cost, the regret value, and the coverage, and insert anchor points sequentially into the optimal positions of the current path according to the comprehensive score, until all anchor points have been inserted.

[0063] Specifically, for nodes where a path has not yet been inserted... In the current path In all slot positions, the anchor insertion algorithm calculates the node Insert into a pair of adjacent nodes The incremental cost incurred between them. Let the insertion before be... After insertion is and The incremental cost is as follows.

[0064] .

[0065] In the formula This is the incremental cost. For nodes With nodes The distance. For nodes With nodes The distance. For nodes With nodes The distance.

[0066] For each node Find the minimum increment and the second smallest increment among all slots, and denote them as: minimum increment and the second smallest increment .

[0067] To characterize the magnitude of the increment relative to the current path scale, the average side length of the current path is defined. .

[0068] .

[0069] In the formula For path. The nodes in the set of nodes of the path and nodes The distance.

[0070] Based on this, the node is obtained. normalized increment .

[0071] To reflect the "sensitivity of insertion position," a "regret value" is introduced. The concept of "suboptimal slot" is used to describe the additional losses that occur when a suboptimal slot is chosen instead of the optimal one.

[0072] .

[0073] In the formula For nodes The regret value. This indicates taking the larger value.

[0074] In addition, to characterize the "supplementation" of a node to the current path in space, the minimum distance from the node to all nodes in the current path is defined as the "coverage".

[0075] .

[0076] In the formula For nodes Coverage. This indicates taking the minimum value. For path. For nodes With nodes The distance.

[0077] Normalize the three quantities to 0 and 1 respectively, and construct a comprehensive scoring function.

[0078] .

[0079] In the formula for Node comprehensive score. Normalized values ​​from 0 to 1 Normalized values ​​from 0 to 1 Normalized values ​​from 0 to 1 for The weighting coefficients. for The weighting coefficients. for The weighting coefficients.

[0080] In this embodiment, the following is taken .

[0081] In each round, the Top-valued nodes are selected from the "unvisited anchor set". In the candidate set consisting of anchor points, the overall score is... The highest anchor point and its optimal insertion position are inserted into the current path until all anchor points are inserted. Through the combined effect of three factors—incremental cost, regret value, and coverage—anchor points not only tend to be added to the path with smaller increments, but also take into account the balance and stability of the overall geometric structure of the path.

[0082] Top- This refers to the "candidate anchor point screening scale parameter" in the anchor point insertion phase. Its core meaning is: in each round, from the "set of anchor points without inserted paths," select the top-ranked anchor points based on their overall score. Anchor points are selected to form a candidate anchor point set, and then the anchor point with the highest comprehensive score is selected from the candidate set and inserted into the current path. This represents a "quantity threshold," for example, if That is, the top 3 anchor points are selected as candidates in each round.

[0083] S53. Non-anchor insertion stage: For each remaining node in the non-anchor set, a candidate node set is obtained by filtering based on its coverage to the current path. For each node in the candidate node set, the normalized incremental cost generated by inserting it into the optimal position of the current path is calculated, and based on the normalized incremental cost, non-anchor nodes are inserted into the current path sequentially until all nodes are inserted, thus obtaining the initial path.

[0084] For the remaining unvisited non-anchor nodes To balance computational complexity and path quality, this embodiment adopts a "Δ_norm-only + candidate pruning" strategy.

[0085] First, based on the minimum distance from the node to the current path. Sort all non-anchor points and select the top-m nodes to form a candidate set. This avoids attempting to insert all remaining nodes, thus reducing computational load.

[0086] In the candidate set In, for each node Calculate its minimum increment in the current path. and according to the average side length Normalization yields the non-anchor point normalized increment.

[0087] .

[0088] In the formula For nodes The normalized increment. For nodes The minimum increment. This represents the average side length of the current path.

[0089] The comprehensive scoring function in the non-anchor phase no longer calculates regret value and coverage normalization.

[0090] .

[0091] As a scoring criterion, from Selected from The largest node and its corresponding optimal slot position are then inserted into the current path. By combining "Cover pre-pruning + Δ_norm-only scoring", the computational complexity of each insertion round is significantly reduced while ensuring path quality.

[0092] S54. Path optimization: Perform 2-opt local search optimization on the initial path to obtain the optimized node access sequence as the planned path.

[0093] After inserting all anchor points and non-anchor points, the anchor insertion algorithm performs a 2-opt local search on the constructed initial path. The basic idea is: for any two non-adjacent edges on the path, construct a new path by swapping their connections; if the total length of the new path is shorter, accept the swap. This process is repeated until there are no further edge swaps that can improve the solution. The 2-opt optimization does not change the anchor point structure framework provided by CEAP, but rather serves as a post-processing step to further compress the path length, making the final solution closer to the optimal solution.

[0094] Figure 4 It is the planning result of a classic example of the internationally recognized Traveling Salesman Problem. Figure 4 KroA100, KroC100, and KroD100 are three classic examples from the internationally recognized Traveling Salesman Problem (TSP) Standard Test Library (TSPLIB).

[0095] This embodiment introduces the Anchor Insertion Algorithm (ARC), which replaces the simple strategy of "randomly selecting anchor points + single increment criterion" in the original method with a combined process of "initializing the skeleton with the farthest point pair + normalized increment Δ_norm + regret value + three-factor scoring of coverage + Δ_norm-only non-anchor insertion + 2-opt local search". This significantly improves the global structural quality of the path while ensuring controllable computational complexity.

[0096] On 15 typical TSPLIB instances, the improved anchor insertion algorithm (ARC) in this embodiment achieves an average relative gap of approximately 5.41% in path length. This is significantly lower than the average relative gap of approximately 12.86% for the original anchor framework and approximately 21.31% for the comparative graph pointer network method, effectively reducing the distance to the optimal solution. The TSPLIB (TSPLibrary) instances refer to a publicly available, standardized library of Traveling Salesman Problem (TSP) instances. It is a standard problem set / benchmark set for TSP problems. The gap refers to the relative difference between the approximate solution generated by the algorithm (i.e., the AGV transport path length planned by this embodiment or the comparative method) and the global optimal solution to the problem (the known shortest path length).

[0097] This embodiment decouples "local neighborhood anchor prediction" from "multi-factor insertion heuristic," forming a path planning framework that is easy to extend and transfer. On the one hand, the CEAP network relies only on an m×m local distance submatrix as input, naturally possessing the ability to generalize across scales and distributions. On the other hand, ARC makes insertion decisions based solely on the distance matrix and anchor labels, easily adapting to different post-processing operators and constraints. Experiments show that, even with training on a small-scale TSP50 network, this embodiment can be extended to various classic TSPLIB instances and randomly generated TSP100-scale instances, demonstrating good scalability and engineering application potential.

[0098] It should be noted that the path planning module, which consists of the Channel-Enhanced AnchorPrediction Network (CEAP) and the Anchor Regret–Coverheuristic algorithm (ARC), can perform inference by calling the offline trained CEAP network weight file, or it can periodically trigger the retraining process based on newly collected warehouse task data to adapt to changes in warehouse layout or task distribution.

[0099] S6. Convert the planned path into a AGV transport path scheduling instruction.

[0100] Specifically, the node access sequence output by the path planning module is converted into an AGV transport path scheduling table, and the scheduling table is sent to the onboard controller of the corresponding AGV via WCS. The preceding steps solve the upper-level access sequence; the lower-level path trajectory requires the onboard controller, such as LiDAR / obstacle avoidance algorithm, to calculate the specific driving trajectory between each two nodes. The specific driving trajectory can be generated using existing trajectory algorithms, and this invention does not impose any specific limitations on it.

[0101] Specifically, during the online scheduling phase, after CEAP+ARC completes the path construction and 2-opt local optimization for a specific AGV transport task, a transport path node access sequence represented by multiple work points is obtained. The path planning server encapsulates the work point sequence into a transport path scheduling table for the corresponding AGV and sends it to the WCS as a high-level path planning result. The WCS, based on the warehouse aisle topology and the current state of the AGV, refines the travel process between adjacent work point pairs into a series of low-level navigation instructions (including specific aisle segment sequences, speed / acceleration constraints, and obstacle avoidance behaviors), and sends them to the corresponding AGV's onboard controller for execution. The AGV's local obstacle avoidance, speed control, and trajectory tracking control are existing AGV navigation control technologies, and this invention does not limit their specific implementation methods.

[0102] To address the single-trip path planning problem of AGV handling tasks in intelligent warehousing scenarios, this embodiment presents an intelligent warehousing multi-AGV collaborative path planning method based on anchor point enhancement heuristics. Given a warehouse aisle topology and multiple target storage locations, this method aims to generate handling routes for AGVs with shorter travel distances and more rational path structures. This embodiment upgrades the original framework of "flattened local neighborhood + pure multilayer perceptron + simple anchor point insertion heuristic" to a CEAP anchor point prediction network based on "convolution + ECA channel attention + global pooling," and tightly couples it with the multi-factor scoring ARC anchor point insertion algorithm. This systematically alleviates the bottlenecks of existing methods in terms of high data costs, insufficient spatial structure utilization, and unstable path quality.

[0103] Specifically, the following CEAP+ARC framework is applied to the abstracted TSP instance for path construction: The overall framework consists of an anchor channel attention prediction network (CEAP, Channel-Enhanced Anchor Prediction Network) and an anchor multi-factor score insertion algorithm (ARC, Anchor Regret–Cover heuristic):

[0104] On the model side, the local m×m neighborhood distance matrix of each node is treated as a two-dimensional single-channel "local distance image." Local spatial patterns are extracted through two convolutional blocks, and an ECA channel attention module is introduced after each convolutional block to adaptively weight the channel features. Global average pooling then compresses the spatial information into a global channel vector, which, combined with a multi-layer fully connected network, achieves binary classification prediction of "whether it is an anchor point." During training, only 50,000 TSP50 instances (approximately 2.5 million features and labels) are used, and with the early stopping mechanism, an anchor point prediction model with good generalization ability on large-scale, heterogeneous instances can be obtained.

[0105] On the algorithm side, ARC first calls the CEAP network to predict anchor points for all nodes of a given TSP instance, dividing the nodes into anchor point sets and non-anchor point sets.

[0106] Then, during the path construction process, the pair of nodes furthest apart from the anchor point set or all nodes is selected as the starting point of the skeleton. Then, in the anchor point stage, a comprehensive score is constructed based on three factors: incremental normalization Δ_norm, regret value, and path coverage. The candidate anchor points are sorted and the optimal insertion position is selected.

[0107] In the non-anchor stage, the Top-m candidates are first obtained by coverage pruning, then the insertion node and slot position are selected according to the Δ_norm-only criterion, and finally the path is further improved by 2-opt local search.

[0108] The solution in this embodiment completes model training using only 50,000 TSP50 instances (approximately 2.5 million feature-label pairs), significantly reducing the reliance on millions of training instances and tens of millions of features. This reduces the time and resource overhead for training data construction and model training (the original method required 1 million TSP25 instances, 500,000 TSP50 instances, and 250,000 TSP100 instances, generating a total of 75 million features and labels).

[0109] The training data construction and label generation are as follows: First, 50,000 TSP50 instances are randomly generated using gurobi to simulate the handling tasks that AGVs may encounter under different warehouse layouts and storage location distributions, constructing a small-scale dataset. For each instance, the optimal path is solved using an exact algorithm, and then each node is labeled as either an "anchor point" or not. To generate training data, the local neighborhood information of each node is used as input features. The local neighborhood of each node consists of its m nearest neighbors in the full distance matrix, forming an m×m local distance matrix. For example, when m=5, the input features for each node are a 5×5 matrix. During label generation, for each node, whether it is an anchor point or not is used as a binary classification label.

[0110] By introducing convolution and channel attention mechanisms in the local m×m neighborhood, the ability to model the correlation between local geometry and channels is enhanced, improving the anchor prediction accuracy and generalization ability to large-scale, heterogeneous instances while maintaining a small training set.

[0111] By employing a multi-factor scoring strategy of "incremental cost Δ_norm + regret value + coverage" for anchor points during the path construction phase, and using coverage pruning and Δ_norm-only insertion rules for non-anchor points, combined with 2-opt local search, the sensitivity of the path generation process to random selection is reduced, and the relative gap with the global optimum is significantly narrowed.

[0112] Experiments demonstrate that, on 15 typical TSPLIB instances, with a significant reduction in training costs, the intelligent warehousing multi-AGV collaborative path planning method in this embodiment further reduces the average relative gap between the generated path cost and the optimal path cost from approximately 21.31% (GPN) and 12.86% (Anchor Insertion) to approximately 5.41% (CEAP), achieving a comprehensive improvement in data efficiency, solution accuracy, and algorithm robustness.

[0113] The method in this embodiment has an average relative gap of approximately 5.41% relative to the globally optimal path, which is significantly better than the original Anchor Insertion method's approximately 12.86% and the Graph Pointer Network (GPN) method's approximately 21.31%, verifying the comprehensive advantages of this embodiment in terms of data efficiency, anchor prediction performance, and path solution quality.

[0114] Through the above improvements, this embodiment, by abstracting the intelligent warehousing AGV transport path planning problem into TSP, can provide AGVs with higher quality transport path solutions while keeping training costs under control, thereby reducing AGV travel distance and task completion time, and improving the overall throughput capacity of the warehouse.

[0115] This embodiment presents an intelligent warehousing multi-AGV collaborative path planning method based on anchor point enhancement heuristics. It combines the Anchor Point Channel Attention Prediction Network (CEAP) with the Anchor Point Multi-Factor Insertion Algorithm (ARC). By abstracting the AGV transport task into a TSP (Traffic Path Provider) model, it can generate high-quality transport paths even with limited training data. Significant improvements have been achieved compared to existing technologies in terms of training data cost, anchor point prediction accuracy, path construction quality, and algorithm scalability.

[0116] By introducing a channel attention convolutional network in the anchor prediction stage and a multi-factor anchor scoring and staged insertion mechanism in the path construction stage, this method significantly improves the accuracy and robustness of the approximate solution of TSP while significantly compressing the training data scale and training cost. This approach tightly combines "local neighborhood anchor learning" with "multi-factor insertion heuristics," retaining the advantages of supervised learning in pattern extraction while avoiding the difficulties in training and limited generalization of global end-to-end learning on large-scale TSPs.

[0117] Example 2: The present invention provides an intelligent warehouse multi-AGV collaborative path planning device, which includes a task acquisition module, a topology acquisition module, a neighbor matrix module, an anchor point prediction module, an anchor point insertion module, and a path conversion module.

[0118] The task acquisition module is used to obtain the transport task information of the target AGV in order to determine the set of work points that the AGV needs to visit.

[0119] The topology acquisition module is used to obtain warehouse aisle topology information and calculate the shortest travel distance between any two work points to obtain a distance matrix. Using work points as nodes and distances as edge weights, a Traveling Salesman Problem instance is constructed.

[0120] The Neighbor Matrix module is used to obtain the m nearest neighbors of each node in the distance matrix in the Traveling Salesman Problem instance, as well as the local neighborhood matrix of each node.

[0121] The anchor point prediction module is used to input the local neighborhood matrix into a pre-trained anchor point prediction network to predict whether each node is an anchor point, thereby dividing all nodes into an anchor point set and a non-anchor point set. The anchor point prediction network employs a network structure including convolutional layers and channel attention layers to extract features and classify the local neighborhood matrix.

[0122] The anchor insertion module is used to construct a planned path based on the distance matrix, the set of anchor points, and the set of non-anchor points using an anchor insertion algorithm.

[0123] The path conversion module is used to convert the planned path into AGV transport path scheduling instructions.

[0124] In embodiment three, the present invention provides an intelligent warehouse multi-AGV collaborative path planning system, which includes a path planning server and at least two AGVs. The intelligent warehouse multi-AGV collaborative path planning system is used to execute the intelligent warehouse multi-AGV collaborative path planning method described in any paragraph of embodiment one.

[0125] Example 4: This invention provides a computer-readable storage medium. The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform any section of the intelligent warehousing multi-AGV collaborative path planning method described in Example 1.

[0126] It is understood that the path planning device can be an electronic device with computing power, such as a portable laptop computer, desktop computer, server, smartphone, or tablet computer.

[0127] Obviously, the embodiments described above are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0128] In the several embodiments provided in this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus and method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0129] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0130] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks. It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0131] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention are also intended to include the plural forms unless the context clearly indicates otherwise.

[0132] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0133] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0134] The terms "first" and "second" used in the embodiments are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first" and "second" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0135] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for collaborative path planning of multiple AGVs in intelligent warehousing, characterized in that, Include: S1. Obtain the transport task information of the target AGV to determine the set of work points that the AGV needs to visit; S2. Obtain the warehouse aisle topology information and calculate the shortest travel distance between any two work points to obtain the distance matrix; using work points as nodes and distances as edge weights, construct a Traveling Salesman Problem instance; S3. Obtain the m nearest neighbors of each node in the distance matrix in the Traveling Salesman Problem instance, and the local neighborhood matrix of each node; S4. Input the local neighborhood matrix into a pre-trained anchor prediction network to predict whether each node is an anchor, thereby dividing all nodes into an anchor set and a non-anchor set; wherein, the anchor prediction network uses a network structure containing convolutional layers and channel attention layers to extract features and classify the local neighborhood matrix. S5. Based on the distance matrix, the anchor point set, and the non-anchor point set, construct the planned path using the anchor point insertion algorithm; wherein, step S5 includes sub-steps S51 to S54; S51. Select the pair of nodes with the farthest distance from the anchor point set or all nodes as the initial path; S52. For each anchor point in the anchor point set that has not yet been inserted into the path, calculate the normalized incremental cost, regret value, and coverage generated by inserting it into each position of the current path; construct a comprehensive score based on the normalized incremental cost, the regret value, and the coverage, and insert the anchor points sequentially according to the comprehensive score. S53. For each remaining node in the non-anchor set, filter it according to its coverage to the current path to obtain a candidate node set; for each node in the candidate node set, calculate the normalized incremental cost generated by inserting it into the optimal position of the current path, and insert non-anchors into the current path in sequence according to the normalized incremental cost until all nodes are inserted to obtain the initial path; S54. Perform 2-opt local search optimization on the initial path to obtain the optimized node access sequence as the planned path; S6. Convert the planned path into a AGV transport path scheduling instruction.

2. The intelligent warehousing multi-AGV collaborative path planning method according to claim 1, characterized in that, The anchor point prediction network takes the local neighborhood distance matrix of the node as input and adopts a structure of convolution + channel attention + global pooling + multi-layer fully connected layers. First, the shape is reshaped into a single-channel feature map of m×m×1 through a shape reshaping layer; Subsequently, the local spatial structure and inter-channel correlation features are extracted and reweighted through two convolutional blocks; Then, a global average pooling layer is introduced to aggregate the spatial dimensions, compressing the m×m×C feature map into a C-dimensional channel vector; The channel vectors then pass through multiple fully connected layers in sequence to enhance the nonlinear expressive power of the network and suppress overfitting; Finally, an output neuron provides the predicted probability of whether the current node is an anchor point.

3. The intelligent warehousing multi-AGV collaborative path planning method according to claim 2, characterized in that, The convolutional block consists of a two-dimensional convolutional layer, a LeakyReLU activation layer, a channel attention layer, and a layer normalization layer; Channel attention layer: First, global average pooling is performed on the convolution output feature map according to the spatial dimension to obtain the channel average vector; The average vector is then subjected to one-dimensional convolution and sigmoid activation to generate attention coefficients for each channel; finally, the attention weights are broadcast back to the original spatial dimension by channel, and the input features are weighted channel by channel to obtain the weighted feature map. ； In the formula This is the weighted feature map; and These are the horizontal and vertical coordinates of the feature map, respectively. For channel indexing; For the first Attention coefficient of each channel; This is a feature map.

4. The intelligent warehousing multi-AGV collaborative path planning method according to claim 2, characterized in that, Global average pooling is suitable for compressing spatial dimensions into channel vectors for subsequent fully connected layer classification; ； In the formula For the first Aggregate representation of channels; and The height and width of the feature map; This is the weighted feature map; and These are the horizontal and vertical coordinates of the feature map, respectively. For channel indexing; Multiple fully connected layers include one 128-dimensional fully connected layer and at least two 256-dimensional fully connected layers; each fully connected layer is followed by a LeakyReLU activation layer, a layer normalization layer, and a random deactivation layer. The output neuron uses the Sigmoid activation function for binary classification prediction.

5. The intelligent warehousing multi-AGV collaborative path planning method according to claim 1, characterized in that, The specific operation of S52 is as follows; For nodes where no path has been inserted yet In all slot positions on the current path, the anchor insertion algorithm calculates the node... Insert into a pair of adjacent nodes The incremental cost incurred between them; let the insertion before be... After insertion is and The incremental cost is as follows; ； In the formula Incremental cost; For nodes With nodes The distance; For nodes With nodes The distance; For nodes With nodes The distance; For each node Find the minimum and second-smallest increments among all slots, and denote them as the minimum increments. and the second smallest increment ; Get Nodes normalized increment ; In the formula This represents the average side length of the current path. The concept of regret value is introduced to describe the additional loss caused by choosing the second-best slot instead of the optimal one. In the formula For nodes Regret value; This indicates taking the larger value; Define nodes The minimum distance to all nodes on the current path is used as the coverage. In the formula For nodes Coverage; This indicates taking the minimum value; For path; For nodes With nodes The distance; Normalize the normalized increment, regret value, and coverage to 0 to 1, and construct a comprehensive scoring function; ； In the formula for Node comprehensive score; Normalized values ​​from 0 to 1 ; Normalized values ​​from 0 to 1 ; Normalized values ​​from 0 to 1 ; for Weighting coefficients; for Weighting coefficients; for Weighting coefficients; In each round, the Top-valued anchors are selected from the set of unvisited anchors. In the candidate set consisting of anchor points, the overall score is... Find the highest anchor point and its optimal insertion position, and insert it into the current path until all anchor points are inserted.

6. The intelligent warehousing multi-AGV collaborative path planning method according to claim 1, characterized in that, The specific operation of S53 is as follows; For the remaining unvisited non-anchor nodes A strategy of "Δ_norm-only + candidate pruning" is adopted; First, based on the minimum distance from the node to the current path. Sort all non-anchor points and select the top-m nodes to form a candidate set. ; In the candidate set In, for each node Calculate its minimum increment in the current path. and according to the average side length Normalization yields the non-anchor point normalized increment; In the formula For nodes The normalized increment; For nodes The minimum increment; This represents the average side length of the current path. Calculate the comprehensive scoring function for the non-anchor stage; In the formula node The overall score.

7. A smart warehousing multi-AGV collaborative path planning device, characterized in that, Include: The task acquisition module is used to acquire the material handling task information of the target AGV in order to determine the set of work points that the AGV needs to visit. The topology acquisition module is used to obtain warehouse aisle topology information and calculate the shortest travel distance between any two work points to obtain a distance matrix; using work points as nodes and distances as edge weights, it constructs a Traveling Salesman Problem instance; The Neighbor Matrix module is used to obtain the m nearest neighbors of each node in the distance matrix in the Traveling Salesman Problem instance, and the local neighborhood matrix of each node; An anchor prediction module is used to input the local neighborhood matrix into a pre-trained anchor prediction network to predict whether each node is an anchor, thereby dividing all nodes into an anchor set and a non-anchor set; wherein, the anchor prediction network adopts a network structure containing convolutional layers and channel attention layers to extract features and classify the local neighborhood matrix. An anchor insertion module is used to construct a planned path based on a distance matrix, an anchor set, and a non-anchor set using an anchor insertion algorithm. Specifically, the module performs the following steps: First, it selects the pair of nodes furthest apart from the anchor set or all nodes as the initial path. For each anchor in the anchor set that has not yet been inserted into the path, it calculates the normalized incremental cost, regret value, and coverage resulting from its insertion into each position of the current path. Based on the normalized incremental cost, regret value, and coverage, it constructs a comprehensive score and, according to the comprehensive score, sequentially inserts anchors into the optimal positions of the current path until all anchors are inserted. Second, for each remaining node in the non-anchor set, it filters for candidate nodes based on their coverage to the current path. Third, for each node in the candidate node set, it calculates the normalized incremental cost resulting from its insertion into the optimal position of the current path and, based on this normalized incremental cost, sequentially inserts non-anchor nodes into the current path until all nodes are inserted, thus obtaining the initial path. Finally, it performs a 2-opt local search optimization on the initial path to obtain the optimized node access sequence as the planned path. The path conversion module is used to convert the planned path into AGV transport path scheduling instructions.

8. A smart warehousing multi-AGV collaborative path planning system, characterized in that, It includes a path planning server and at least two AGVs; the intelligent warehouse multi-AGV collaborative path planning system is used to execute an intelligent warehouse multi-AGV collaborative path planning method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform a smart warehouse multi-AGV collaborative path planning method as described in any one of claims 1 to 6.

Images