Capacity constraint vehicle path optimization method based on visual language model

Abstract

The invention discloses a capacity constraint vehicle path optimization method based on a visual language model, and relates to the technical field of vehicle path optimization image recognition processing, and the method comprises the steps: constructing vehicle path information into image input, and employing RGB three channels to encode vehicle residual capacity, customer demand intensity and path load density; jointly inputting the image and a natural language prompt into a pre-training visual language model, and identifying a to-be-adjusted region node set; and constructing a local CVRP sub-problem according to the regional node set, reserving an unmodified path segment as a capacity sensing super node, optimizing the sub-problem through an asynchronous parallel solver, and only backfilling a sub-solution with an improved target value to a main path. According to the method, combined reasoning of images and languages is achieved, the to-be-optimized area can be recognized without depending on special training, dynamic adjustment and updating of the path structure are achieved through sub-problem reconstruction and a parallel solving mechanism, and therefore the multi-modal path intelligent optimization effect which is difficult to achieve in the prior art is achieved.

Description

Technical Field

[0001] This invention relates to the field of image recognition and processing technology for vehicle routing optimization, specifically a capacity-constrained vehicle routing optimization method based on a visual language model. Background Technology

[0002] The Capacitated Vehicle Routing Problem (CVRP), a classic combinatorial optimization problem in logistics scheduling and route optimization, has long been an important research subject in operations research, artificial intelligence, and intelligent transportation. Traditional solution methods include exact algorithms and heuristic methods. In recent years, deep learning and neural network models have been gradually introduced to improve solution efficiency and generalization ability. Furthermore, with the development of multimodal artificial intelligence, visual language models, due to their powerful expressive capabilities in semantic understanding and visual perception, have achieved significant results in areas such as image and text retrieval and visual question answering, and are increasingly being explored for collaborative optimization of structured tasks.

[0003] While some academic work has attempted to incorporate deep learning methods into the CVRP solution process, such as using graph neural networks to model path structures or generating vehicle access sequences through sequence modeling, these methods still primarily rely on single-modal information input and lack the ability to comprehensively express spatial layout and constraint semantics. They also struggle to fully utilize the complex interactions between vehicle capacity, node requirements, and path load. Furthermore, some methods that attempt to incorporate image or visualization information are largely limited to path visualization and have not yet achieved embedded applications of path deconstruction at the information encoding level. In addition, existing methods generally employ static path solving mechanisms, lacking dynamic adjustment strategies oriented towards local structures, making it difficult to cope with complex delivery scenarios characterized by high real-time demands and varied structures.

[0004] More importantly, there is currently no effective way to unify the modeling and reasoning of delivery rules, vehicle attributes and constraints in natural language form with route images. The lack of cross-understanding ability between image information and language semantics leads to the system exhibiting significant generalization ability when processing complex rule instructions. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by this invention is that existing capacity-constrained vehicle path optimization methods lack the ability to jointly model multimodal information, cannot perform dynamic region identification and local path adjustment based on complex constraints, have a static and fixed path solution process, and have poor generalization ability. The invention also addresses the problem of how to use visual language models to achieve image-level path region identification and dynamic optimization reconstruction under multiple constraints.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a capacity-constrained vehicle path optimization method based on a visual language model, comprising: constructing vehicle path information as an image input, and using RGB three-channel encoding to determine the remaining vehicle capacity, customer demand intensity, and path load density; jointly inputting the image and natural language prompts into a pre-trained visual language model to identify the set of regional nodes to be adjusted; constructing a local CVRP sub-problem based on the set of regional nodes, retaining unmodified path segments as capacity-aware supernodes, optimizing the sub-problem through an asynchronous parallel solver, and backfilling only the sub-solutions with improved target values ​​to the main path.

[0008] As a preferred embodiment of the capacity-constrained vehicle path optimization method based on visual language model described in this invention, the image input includes: extracting customer node coordinates, path connection relationships, vehicle current location and remaining capacity information in the delivery network, and performing normalization processing to generate an image structure of equal size; nodes in the image are marked with pixels, paths are constructed with connecting lines, and warehouse anchor point areas are added to the image edges.

[0009] As a preferred embodiment of the capacity-constrained vehicle route optimization method based on a visual language model described in this invention, the RGB three-channel encoding includes: embedding the remaining vehicle capacity ratio in the R channel, the customer demand intensity in the G channel, and the unit load density or driving cost of the route segment in the B channel into the pixel values, and linearly mapping the values ​​to the color space; performing weighted color fusion processing on the route intersection area, marking overloaded route segments with border graphics, and highlighting warehouse nodes with specific colors to visualize and identify the capacity and demand status.

[0010] As a preferred embodiment of the capacity-constrained vehicle routing optimization method based on a visual language model according to the present invention, the natural language prompts include text instructions describing delivery rules, vehicle attributes, task priorities, or service time windows, which, together with the image input, constitute the prompt conditions of the visual language model; the visual language model matches key path segments in the image with language-indicated constraints through a cross-attention mechanism, and outputs several suggested sub-region sets containing customer node IDs or image region boxes; the sets serve as candidate target regions for path reconstruction.

[0011] As a preferred embodiment of the capacity-constrained vehicle path optimization method based on visual language model described in this invention, the construction of local CVRP sub-problems includes: extracting the relevant nodes and path segments for each output sub-region to construct a CVRP subgraph; compressing the remaining paths into capacity-aware supernodes to store the corresponding load and structure summary; if the sub-problem involves multiple path intersections or region overlaps, it is split and filtered according to node density, task weight, and capacity conflict degree; the sub-problem structure records the backfill index in the original global solution.

[0012] As a preferred embodiment of the capacity-constrained vehicle path optimization method based on visual language model described in this invention, the optimization sub-problem includes: distributing the sub-problem to multiple solver modules through a master-slave asynchronous task queue, and independently performing optimization within a time limit T using a local search algorithm or reinforcement learning model.

[0013] As a preferred embodiment of the capacity-constrained vehicle path optimization method based on visual language model described in this invention, the backfilling includes: whenever a subproblem is solved, immediately comparing the objective function with the current main path solution; if the global index is improved, the original path segment is replaced and fixed and updated; otherwise, the process is withdrawn; the process is repeated until no improvement is achieved after K iterations or the total running time limit is reached.

[0014] Another objective of this invention is to provide a capacity-constrained vehicle routing optimization system based on a visual language model.

[0015] As a preferred embodiment of the capacity-constrained vehicle routing optimization system based on a visual language model according to the present invention, it includes an encoding module, an identification module, and an optimization backfilling module. The encoding module is used to construct vehicle routing information as an image input and uses RGB three-channel encoding to determine the remaining vehicle capacity, customer demand intensity, and route load density. The identification module is used to jointly input the image and natural language prompts into a pre-trained visual language model to identify the set of regional nodes to be adjusted. The optimization backfilling module is used to construct a local CVRP sub-problem based on the set of regional nodes, retain unmodified path segments as capacity-aware supernodes, optimize the sub-problem using an asynchronous parallel solver, and backfill only the sub-solutions with improved target values ​​to the main path.

[0016] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement a capacity-constrained vehicle path optimization method based on a visual language model.

[0017] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a capacity-constrained vehicle routing optimization method based on a visual language model.

[0018] The beneficial effects of the present invention are as follows: The capacity-constrained vehicle path optimization method based on visual language model provided by the present invention introduces a pre-trained visual language model, which realizes joint reasoning of image and language for the first time. It can identify the region to be optimized without relying on special training, and realizes dynamic adjustment and updating of path structure through sub-problem reconstruction and parallel solution mechanism, thereby achieving multimodal path intelligent optimization effect that is difficult to achieve in the prior art. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is an overall flowchart of a capacity-constrained vehicle routing optimization method based on a visual language model, provided in Embodiment 1 of the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0022] Example 1, referring to Figure 1 As an embodiment of the present invention, a capacity-constrained vehicle routing optimization method based on a visual language model is provided, comprising: S1: Construct vehicle route information as image input and use RGB three-channel encoding to determine vehicle remaining capacity, customer demand intensity, and route load density.

[0023] Furthermore, the image input includes extracting customer node coordinates, path connection relationships, vehicle current location and remaining capacity information in the delivery network, and performing normalization processing to generate an image structure of the same size; Nodes in the image are marked with pixels, paths are constructed with lines, and warehouse anchor point areas are added to the image edges.

[0024] It should also be noted that a preferred scheme for constructing vehicle routing information as image input specifically includes, Extracting key input data for the delivery network from the current CVRP problem instance mainly includes: the spatial coordinates of each customer node. Demand Spatial coordinates of warehouse nodes The current location and remaining capacity of each vehicle The order in which vehicles are accessed in the current solution path To ensure consistent representation of CVRP problems across different scales or coordinate systems in the image, all coordinates need to be normalized. Let the maximum horizontal and vertical coordinates of all nodes be... , Then the normalized coordinates are: Then, the normalized coordinates are mapped to the image pixel space. Let the image resolution be... , Image height, Given the image width, the pixel coordinates of each node in the image are: In the image, each customer node is drawn as a solid circle at its corresponding pixel position, with an initial pixel value of zero. The path connections connect the nodes sequentially according to the vehicle access order, and the edges are drawn as pixel lines to form the entire vehicle path structure diagram.

[0025] To enhance the visual model's ability to spatially recognize warehouse nodes, warehouse anchor point regions are set at the image edges (such as the top left corner or the center of the image). For example, an anchor point is set at the top left corner of the image. The red rectangular marker blocks of pixels are used to represent the global location anchor point of the warehouse, while the original coordinates of the warehouse are also drawn as standard nodes in the image.

[0026] It should be noted that RGB three-channel encoding includes linearly mapping the values ​​to the color space by embedding the vehicle remaining capacity ratio in the R channel, the customer demand intensity in the G channel, and the path segment unit load density or driving cost in the B channel into the pixel values ​​respectively. Weighted color fusion processing is applied to the intersection areas of paths, overloaded path segments are marked with border graphics, and warehouse nodes are highlighted with specific colors to visualize and identify capacity and demand status.

[0027] It should also be noted that after the image structure is established, it undergoes RGB three-channel capacity information encoding processing. Specifically, this includes: R-lane, or red lane: the remaining capacity of vehicles traversing each path segment. Normalize to [0,1], then linearly map to integer values ​​of [0,255], and use them as the red values ​​for the path segment pixels; G-channel, or green channel: customer node demand. After normalization, it is mapped to the green channel; Channel B, or the blue channel: calculates the unit load density or unit distance cost between two nodes in a path segment. And map it to the blue channel: in This is the maximum value for the unit load density across all path segments, ensuring that all blue values ​​are within the standard color space. Represents a node arrive To enhance the model's ability to recognize intersection areas in images, for overlapping path segments or areas where multiple paths intersect, a weighted average of pixel values ​​is used for color synthesis. in, The merged pixel color value indicates the final displayed color of the intersection area. Indicates the first The color value of a certain channel of a path. By assigning corresponding weight values ​​(which can be allocated based on path traffic or importance), this mechanism avoids color overlay and saturation, maintaining image continuity. Furthermore, for path segments experiencing overload in the solution (i.e., a segment of the path where the actual delivery load exceeds the vehicle capacity limit), a specific colored border (such as yellow or purple) is added to the image for graphic marking; simultaneously, warehouse nodes are highlighted with a preset color (such as pure red or pure white) to ensure their distinguishability and recognizability in the image. The final generated image is a three-channel image. The color image, in which each channel embeds multi-dimensional features such as spatial structure, vehicle state and path load, serves as the image modality part of the subsequent visual language model input.

[0028] It should also be noted that by constructing traditional CVRP path data as visual image input and embedding multi-dimensional delivery constraint information using RGB three-channel encoding, unified modeling of path structure and capacity constraints is achieved, significantly improving the model's perceptibility and interpretability of path conflicts. Through this image construction method, node coordinates, demand, vehicle remaining capacity, and path load status in the delivery network are all embedded in the image space in pixel form, making complex constraint relationships spatially concrete and providing an efficient and information-rich input interface for multimodal models. Compared to traditional path vector encoding methods, this method achieves simultaneous encoding of spatial topology and state variables, effectively solving problems such as information dimension separation and weak path correlation, enhancing the downstream identification module's ability to locate conflict areas, and possessing stronger structural adaptability and scalability.

[0029] S2: Input the image and natural language prompts together into the pre-trained visual language model to identify the set of region nodes to be adjusted.

[0030] Furthermore, natural language prompts include text instructions describing delivery rules, vehicle attributes, task priorities, or service time windows, which, together with image input, form the prompting conditions of the visual language model. The visual language model matches key path segments in an image with language referential constraints through a cross-attention mechanism, and outputs a set of suggested sub-regions containing client node IDs or image region boxes. The set serves as a candidate target region for path reconstruction.

[0031] It should be noted that a preferred scheme for identifying the set of regional nodes to be adjusted specifically includes, firstly, constructing multimodal input pairs. ,in: This represents the RGB three-channel image generated in step S1, which embeds the vehicle's remaining capacity, customer demand intensity, and load density per unit of the path segment. Natural language prompts describe the delivery strategy rules, task requirements, or vehicle attribute limitations in the current CVRP instance. The structure can use templated sentences, with specific examples as follows: "Please identify areas in the route where there is overload"; "Prioritize routes that meet high-demand nodes (demand greater than 80%)"; "Identify route segments where vehicle capacity is less than 20%"; "Prioritize scheduling nodes closer to the warehouse within the service time window." Natural language prompts can be automatically generated by the system or manually input by the scheduling system, supporting automatic parameter variable filling and template standardization.

[0032] Pre-trained visual language model (VLM) receives input Subsequently, the internal structure is based on a graph-text cross-attention mechanism, and its workflow includes: For images Perform patch segmentation and embedding to generate image representation vector sequences. , Indicates the first in the image Visual embedding vectors at each location; Natural language prompts Encoded as a sequence representation ,in Represents the vector of each word; During the joint encoding phase, the model uses graph-text cross-attention to calculate the following interaction relationships: in, For image With language Attention value, These are the linear transformation matrices for the query and the key, respectively. To scale the dimensions, For transpose operation, The function is a normalization function that ensures the attention weights satisfy a probability distribution.

[0033] Cross-attention output is used to determine which regions in the image have a high degree of matching with the language cues, and marks the "regions to be optimized" on the image as region boxes or sets of node coordinates.

[0034] The output is a set of several candidate regions, denoted as: Each of them It can be in one of the following two forms: Client node set: This indicates the path node number that should be replanned; Image sub-region: rectangular region box , indicating areas of dense conflict identified in the image.

[0035] To improve the accuracy and mutual exclusivity of candidate regions, the system performs a set of overlapping regions. Non-maximum suppression algorithm is applied for deduplication, retaining duplicates with an intersection-union ratio below a threshold. The high-confidence region. The final output region set. As the input target region for constructing local CVRP subproblems, the selected region is ensured to be representative, have capacity for conflict, or possess optimization potential. Through this step, the visual language model can accurately identify potential problem regions in the current vehicle path solution based on multimodal input without requiring specific training for the CVRP task, significantly improving the overall path optimization efficiency and the system's adaptability.

[0036] It should also be noted that by jointly inputting image input and natural language prompts into a pre-trained visual language model, automatic identification of regions to be optimized in the path is achieved. This mechanism utilizes cross-attention to capture the matching relationship between high-risk path segments in the image and semantic rules, enabling dynamic identification of potential conflict regions such as "overloaded paths," "high-demand regions," or "nodes with low remaining capacity" without task-specific training. Compared to the limitations of traditional path optimization methods that rely on manual rules or specific model training, this invention can handle different types of scheduling instructions without altering the model structure, greatly improving the system's versatility and real-time response capability, and providing a flexible, efficient, and scalable intelligent interface for multi-constraint intelligent scheduling systems.

[0037] S3: Construct a local CVRP subproblem based on the set of regional nodes, retain unmodified path segments as capacity-aware supernodes, optimize the subproblem through an asynchronous parallel solver, and only backfill the sub-solutions with improved target values ​​to the main path.

[0038] Furthermore, constructing local CVRP sub-problems involves extracting the relevant nodes and path segments for each output sub-region to construct a CVRP subgraph, and compressing the remaining paths into capacity-aware supernodes to store the corresponding load and structure summary. If a subproblem involves multiple paths or overlapping regions, it is split and filtered based on node density, task weight, and capacity conflict. The subproblem structure records the backfill index in the original global solution.

[0039] It should also be noted that a preferred scheme for constructing local CVRP subproblems specifically includes dividing each region A local CVRP subproblem is constructed, independently optimized using a parallel solver, and valid sub-solutions are backfilled into the global path solution according to rules. Subproblem construction and supernode compression are performed on each candidate region. Extract the set of customer nodes contained in this region. and the corresponding path segment connections. , forming a subgraph ; will not belong to any path segment Unified compression into capacity-aware supernodes Each supernode Save the following attributes: source path segment number, current total load. Remaining capacity estimate Start and end positions and path weights.

[0040] The subproblem structure formed after compression is as follows: in This represents the connection relationship between subgraphs and supernodes, used to ensure path connectivity and consistency of backfill locations. To avoid path overlap or solution conflicts between subproblems, the system cross-validates all candidate regions. If any two regions exist... and satisfy: in If a set overlap threshold is set, such as 0.3, the system prioritizes conflicting regions based on indicators such as node density, customer demand weight, and current path conflict level, retaining only high-priority regions and filtering out the others. For each retained subproblem, its "path segment start and end index" in the original solution structure is recorded as a backfill anchor point to ensure that the optimized result can seamlessly replace the corresponding segment in the original solution.

[0041] It should be noted that the optimization of sub-problems includes distributing sub-problems to multiple solver modules through a master-slave asynchronous task queue, and performing optimization independently within the time limit using a local search algorithm or reinforcement learning model.

[0042] It should also be noted that a preferred solution for optimizing subproblems specifically includes, after the asynchronous parallel optimization solver has completed the construction of all subproblems, the system sets the subproblem set together. The solutions are distributed asynchronously to the parallel solver for optimization. Each solution task is packaged into a task unit. Its scheduling and execution mechanism is as follows: The system uses an asynchronous task queue with a master-slave structure. The master thread is responsible for distributing and scheduling subproblems, and each slave thread is bound to a local solver. The solver can be a local search algorithm, a metaheuristic method, or a lightweight reinforcement learning model; Each task is bound to a time limit during scheduling. To prevent prolonged computation from impacting overall performance, the task performs the following optimization objectives: in This indicates the penalty for exceeding vehicle capacity in a route segment. As a regulating factor, Represents a node arrive The distance between them For sub-paths, after the task is completed, candidate optimized sub-path solutions are output. .

[0043] It should also be noted that backfilling includes, whenever a subproblem is solved, immediately comparing the objective function with the current main path solution; if the global metrics are improved, the original path segment is replaced and the update is fixed; otherwise, it is withdrawn. The process is repeated until no improvement is achieved after K iterations or the maximum total runtime is reached.

[0044] It should also be noted that a preferred scheme for backfilling the sub-solution to the main path to improve the target value specifically includes, whenever a task Upon completion, immediately execute the following backfill judgment logic: calculate the global path target value after sub-solution replacement. , and the target value of the current main path solution Compare, if the following conditions are met: in For the minimum significant improvement threshold, such as 1e-3, then... Replace the corresponding position in the main path and update it permanently; if this condition is not met, cancel the sub-solution and leave the main path unchanged. The entire process continues iteratively until one of the following termination conditions is met: K consecutive optimizations failed to reduce the objective function, such as when K=10; Total running time has reached the set limit. .

[0045] It should also be noted that the local subproblem construction and capacity-aware supernode compression mechanism proposed in this invention achieves locally controllable optimization of the path structure. By constructing local CVRP subgraphs for identified candidate regions and introducing supernodes to retain the load information of the original path segments, the optimization process achieves efficient local structure replacement while ensuring global connectivity and feasibility. Furthermore, by combining an asynchronous parallel solver with an objective value improvement backfill mechanism, it ensures that optimized sub-solutions are only replaced and applied when they meet the condition of global performance improvement, effectively avoiding the problem of "getting worse with each improvement" in traditional local optimization. This mechanism possesses high real-time performance and convergence control capabilities, enabling the path optimization process to have adaptive, verifiable, and backtrackable characteristics, significantly improving the scheduling stability and optimization accuracy of the system in complex dynamic environments.

[0046] Example 2, an embodiment of the present invention, provides a capacity-constrained vehicle routing optimization system based on a visual language model, including an encoding module, a recognition module, and an optimization backfilling module.

[0047] The encoding module is used to construct vehicle route information as image input and uses RGB three-channel encoding to encode vehicle remaining capacity, customer demand intensity, and route load density; the recognition module is used to input the image and natural language prompts into a pre-trained visual language model to identify the set of regional nodes to be adjusted; the optimization and backfilling module is used to construct local CVRP sub-problems based on the set of regional nodes, retain unmodified path segments as capacity-aware supernodes, optimize the sub-problems through an asynchronous parallel solver, and backfill only the sub-solutions with improved target values ​​to the main path.

Claims

1. A capacity-constrained vehicle routing optimization method based on a visual language model, characterized in that, include: Vehicle route information is constructed as an image input, and RGB three-channel encoding is used to determine vehicle remaining capacity, customer demand intensity, and route load density. The image and natural language prompts are jointly input into a pre-trained visual language model to identify the set of region nodes to be adjusted; Construct a local CVRP subproblem based on the set of regional nodes, retain unmodified path segments as capacity-aware supernodes, optimize the subproblem using an asynchronous parallel solver, and only backfill sub-solutions with improved target values ​​to the main path.

2. The capacity-constrained vehicle routing optimization method based on a visual language model as described in claim 1, characterized in that: The image input includes, Extract customer node coordinates, path connection relationships, vehicle current location and remaining capacity information from the delivery network, and perform normalization processing to generate an image structure of the same size; Nodes in the image are marked with pixels, paths are constructed with lines, and warehouse anchor point areas are added to the image edges.

3. The capacity-constrained vehicle routing optimization method based on a visual language model as described in claim 2, characterized in that: The RGB three-channel encoding includes, By embedding the remaining vehicle capacity ratio in the R channel, the customer demand intensity in the G channel, and the path segment unit load density or driving cost in the B channel into the pixel values, the values ​​are linearly mapped to the color space. Weighted color fusion processing is applied to the intersection areas of paths, overloaded path segments are marked with border graphics, and warehouse nodes are highlighted with specific colors to visualize and identify capacity and demand status.

4. The capacity-constrained vehicle routing optimization method based on a visual language model as described in claim 3, characterized in that: The natural language prompts include, Text instructions describing delivery rules, vehicle attributes, task priorities, or service time windows, combined with image input, form the prompting conditions of the visual language model; The visual language model matches key path segments in an image with language referential constraints through a cross-attention mechanism, and outputs a set of suggested sub-regions containing client node IDs or image region boxes. The set serves as a candidate target region for path reconstruction.

5. The capacity-constrained vehicle routing optimization method based on a visual language model as described in claim 4, characterized in that: The construction of local CVRP subproblems includes, For each output sub-region, extract the relevant nodes and path segments to construct a CVRP subgraph, and compress the remaining paths into capacity-aware supernodes to save the corresponding load and structure summary. If a subproblem involves multiple paths or overlapping regions, it is split and filtered based on node density, task weight, and capacity conflict. The subproblem structure records the backfill index in the original global solution.

6. The capacity-constrained vehicle routing optimization method based on a visual language model as described in claim 5, characterized in that: The optimization sub-problems include, Subproblems are distributed to multiple solver modules through a master-slave asynchronous task queue, and optimization is performed independently within the time limit using local search algorithms or reinforcement learning models.

7. The capacity-constrained vehicle routing optimization method based on a visual language model as described in claim 6, characterized in that: The backfilling includes, Whenever a subproblem is solved, the objective function is immediately compared with the current main path solution. If the global performance is improved, the original path segment is replaced and the update is fixed; otherwise, the solution is withdrawn. The process is repeated until no improvement is achieved after K iterations or the maximum total runtime is reached.

8. A capacity-constrained vehicle routing optimization system based on a visual language model, employing the capacity-constrained vehicle routing optimization method based on a visual language model as described in any one of claims 1 to 7, characterized in that: Includes an encoding module, an identification module, and an optimized backfilling module; The encoding module is used to construct vehicle route information as an image input and uses RGB three-channel encoding to determine vehicle remaining capacity, customer demand intensity, and route load density. The recognition module is used to input the image and natural language prompts into a pre-trained visual language model to identify the set of region nodes to be adjusted; The optimization backfilling module is used to construct local CVRP subproblems based on the set of regional nodes, retain unmodified path segments as capacity-aware supernodes, optimize the subproblems through an asynchronous parallel solver, and backfill only the sub-solutions with improved target values ​​to the main path.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the capacity-constrained vehicle path optimization method based on a visual language model as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the capacity-constrained vehicle path optimization method based on a visual language model as described in any one of claims 1 to 7.

Images