AGV heuristic path planning method enhanced by conditional diffusion cost field

Abstract

The application discloses an AGV heuristic path planning method enhanced by conditional diffusion cost field, and belongs to the technical field of automatic guided vehicle (AGV) path planning, generative diffusion artificial intelligence and intelligent manufacturing logistics scheduling. A workshop grid map is first constructed, and then a geometric prior cost field with safety constraints is constructed; the workshop map, start and end points, geometric priors and other information are input into a DDIM model as conditions to generate a residual cost field, and the two are weighted and fused to obtain a fused cost field. The fused cost field is introduced into a differentiable A* evaluation function to reduce the invalid search range. During offline training of the model, multi-objective optimization is performed on diffusion generative loss, search area loss, path length loss and safety constraint loss, and in the reasoning stage, few-step deterministic DDIM sampling is adopted to ensure real-time performance. The application takes into account the interpretability of traditional A* and the semantic learning ability of the diffusion model, and improves the AGV path planning efficiency, driving safety and dynamic environment self-adaptive ability.

Description

Technical Field

[0001] This invention relates to the fields of automated guided vehicle (AGV) path planning, generative diffusion artificial intelligence, and intelligent manufacturing logistics scheduling, and particularly to a conditional diffusion cost field-enhanced AGV heuristic path planning method. Background Technology

[0002] A digital printing workshop comprises various functional units, including a paper warehouse, digital printing presses, laminating, cutting, drying, binding, quality inspection, semi-finished product storage, finished product storage, and manual operation areas. Currently, the digital printing industry is characterized by small-batch, multi-variety, and rapid delivery production. The frequency of pallet transfers of paper, ink, semi-finished products, and finished products within the workshop has increased significantly, making AGVs (Automated Guided Vehicles) the core equipment for material handling within the workshop.

[0003] In actual production environments, digital printing workshops have compact equipment layouts and narrow logistics channels. At the same time, there are dynamic interference factors such as temporary material stacking, equipment occupation, personnel interleaving operations, and multiple AGVs cooperating. AGV path planning must simultaneously meet multiple constraints such as path reachability, real-time and efficient planning, driving safety, and dynamic environment adaptability.

[0004] Traditional A* heuristic path planning algorithms are simple in logic, have good path connectivity, and are highly interpretable in engineering. However, node evaluation relies solely on the cumulative path cost G and the heuristic distance H to the destination, which has significant limitations in terms of environmental representation capabilities. It cannot quantitatively distinguish the passage priorities of main logistics channels, manual operation areas, temporary material storage areas, and dangerous areas near equipment. In complex and dense obstacle scenarios, it will expand a large number of invalid search nodes, which will significantly reduce the path planning speed.

[0005] Existing deep learning-based path optimization schemes fall into two categories: one type directly predicts the cost map through a neural network to guide the search, relying on a massive amount of manually labeled trajectory samples, and has poor generalization performance when workshop equipment layout, order processes, and temporary material stockpiling change; the other type of network directly outputs the complete driving path, losing the path connectivity constraints of traditional A*, resulting in weak interpretability and high difficulty in on-site debugging.

[0006] Existing technologies lack an integrated path planning scheme that takes into account geometric safety priors, dynamic workshop logistics semantics, and the stability of traditional heuristic search; they cannot simultaneously solve the technical pain points of low planning efficiency, poor dynamic adaptability, difficulty in ensuring path safety, and insufficient model generalization ability in the digital printing workshop environment. Summary of the Invention

[0007] The purpose of this invention is to propose a conditionally diffused cost field-enhanced AGV heuristic path planning method, which is designed for complex and dynamic logistics scenarios in digital printing workshops. It integrates the geometric prior cost field and the conditional DDIM-generated residual cost field to construct a fused cost field, and embeds the fused cost field into the evaluation function of differentiable A* nodes. This method takes into account path planning efficiency, driving safety, and dynamic environmental adaptability, while fully preserving the engineering interpretability and path connectivity of the A* algorithm.

[0008] To achieve the above objectives, this invention provides a conditional diffusion cost field-enhanced AGV heuristic path planning method, comprising the following steps: Step S1: Model the digital printing workshop environment as OK Columnar 2D raster map, construct raster occupancy matrix Collect the starting point of the AGV for a single transfer task. Target workstation endpoint Generate the starting point one-hot matrix One-hot matrix at the endpoint ; Step S2: Calculate obstacle distance transformation map based on grid map Endpoint Normalized Heuristic Distance Plot A geometric prior cost field is constructed by combining preset weights. ; Step S3: Integrate the grid occupancy matrix Starting point unique hot matrix One-hot matrix at the endpoint Heuristic distance graph Obstacle distance transformation diagram Geometric prior cost field As input to the conditional DDIM model, conditional information is used to generate the residual cost field. ; Step S4: Apply the geometric prior cost field With residual cost field Linear fusion yields the fusion cost field. ; Step S5: Merge the cost field A node evaluation function with embedded differentiable A* algorithm is used to guide the differentiable A* algorithm to complete node expansion search based on the fusion cost field, and output the planned driving path for material transfer of AGV in digital printing workshop.

[0009] Preferably, in step S2, the formula for calculating the geometric prior cost field is: ; In the formula: Weighting for proximity penalty to obstacles; Weights are assigned to guide the direction of the endpoint; The penalty coefficient is set for the obstacle area. To prevent extremely small constants with a denominator of 0; Obstacle distance transformation diagram , The set of all obstacle nodes within the grid map; heuristic distance graph For grid To the destination The normalized Euclidean distance, Manhattan distance, or eight-neighbor distance.

[0010] Preferably, in step S3, the conditional DDIM model is divided into an offline training phase and a real-time inference phase; The offline training phase involves a forward noise addition process: labeling the real residual cost field. Adding Gaussian noise yields different diffusion time steps Noisy residual field , The cumulative noise attenuation coefficient is the diffusion coefficient. Standard two-dimensional Gaussian noise; Noisy residual field diffusion time step The concatenated conditional tensor Input noise prediction network, output predicted noise .

[0011] Preferably, during the offline training phase, a multi-objective joint total loss function is constructed to optimize the parameters of the conditional DDIM model. The total loss function is: ; In the formula: Loss is generated for diffusion; This is the loss due to path planning. To constrain losses for safety reasons; Weights for path planning loss; Weight the loss for safety constraints; Diffusion generates loss ; Path planning loss , To mark the standard compliance path matrix, For search area loss, For path length loss, These are the weights of the two sub-losses, respectively. Safety constraint loss , A preset safety grid distance threshold is set between the AGV and obstacles; when the distance between the path grid point and the obstacle is less than... At that time, the value of the safety constraint loss increases, forcing the model to generate a cost field with a high safety level.

[0012] Preferably, in step S3, the real-time inference stage uses few-step deterministic DDIM denoising sampling to generate the residual cost field. Single-step clean residual field estimation formula: ; go through The final residual cost field is output after iterative denoising. , The sampling steps are preset, and there is no random noise input during the sampling process, which meets the real-time low-latency planning requirements of AGV.

[0013] Preferably, in step S4, the formula for calculating the fusion cost field is: ; In the formula: The residual modulation coefficient is used to adjust the influence of the residual cost field generated by DDIM on the overall fusion cost field. The fusion cost field also includes the geometric safety constraints of the raster map and the semantic correction information of the digital printing workshop process transfer, main logistics channel, and dynamic obstacles.

[0014] Preferably, in step S5, the differentiable A* algorithm updates the node comprehensive evaluation function as follows: ; In the formula: From the starting point to the current grid node The cumulative path cost; For nodes The heuristic distance to the endpoint; For nodes The value of the fusion cost field; the safe main logistics channel nodes with lower fusion cost field value will be prioritized for expansion by the algorithm.

[0015] Preferably, in step S5, the nodes to be expanded are selected using a matrix-differentiable method, and the node selection matrix is ​​calculated using the following formula: ; In the formula: This is the global cumulative path cost matrix; This is an open candidate node 0-1 matrix; Temperature coefficient; For element-wise multiplication; It is the inner product of matrices; To find the maximum value, convert the position into a one-hot matrix and output the expanded node for this round; iterate and update the search matrix until the target endpoint is reached.

[0016] Therefore, the AGV heuristic path planning method with conditional diffusion cost field enhancement described above has the following advantages: (1) This invention proposes a two-layer cost field architecture that integrates geometric prior and DDIM residual. The geometric prior provides stable and quantifiable safety constraints, while the DDIM residual learns semantic information such as process transfer, main channel preference, and dynamic obstacles in the digital printing workshop. Compared with a single neural network that directly generates the cost map, the training stability is stronger, and the generalization ability is significantly improved after changes in workshop layout and logistics tasks. (2) Introducing the fusion cost field into the A* node evaluation function, while retaining the advantages of traditional A* path connectivity, engineering interpretability and easy deployment, guides the search to focus on the effective logistics channel area, greatly reduces invalid extended nodes, and improves the speed of path planning and dynamic replanning in the multi-material transfer scenario of digital printing workshop. (3) DDIM adopts a few-step deterministic denoising sampling, which eliminates random noise sampling time and meets the low latency requirement of AGV real-time scheduling. Unlike the deep learning scheme that directly outputs the path from end to end, this invention relies on a standard heuristic search framework, which allows for on-site manual adjustment of the cost field weight parameters, making debugging convenient. (4) The model adopts multi-objective joint training with diffusion loss, search efficiency loss, path length loss and safety distance loss. The model simultaneously optimizes the three major indicators of planning speed, path length and driving safety, adapts to the complex working environment of digital printing workshop with dense equipment, mixed human and machine traffic and narrow passages, and effectively reduces the risk of AGV scratching equipment and material collision.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of an AGV heuristic path planning method with conditional diffusion cost field enhancement in an embodiment of the present invention. Figure 2 This is a schematic diagram of the conditional DDIM training process in an embodiment of the present invention; Figure 3 This is a schematic diagram of the conditional DDIM inference process in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. 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.

[0020] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0021] Example like Figures 1-3 As shown in the figure, this embodiment proposes a conditional diffusion cost field-enhanced AGV heuristic path planning method, the specific steps of which are as follows: I. Workshop Environment Grid Modeling: The entire digital printing workshop is abstracted as a two-dimensional raster map with R rows and C columns. The raster node set is defined as follows: ; Constructing a grid occupancy matrix : Raster value The area representing obstacles includes the printing main unit, lamination and cutting fixing equipment, walls, safety fences, impassable material stacks, and temporary closed passages; This represents the area that the AGV can traverse.

[0022] Partitioning node set: Obstacle node set ; set of walkable nodes .

[0023] Input starting point for a single path planning task Target workstation endpoint Encode the start and end points into one-hot matrices respectively. , Basic environment input tensor: ; In the formula: This represents the operation of splicing tensor channel dimensions.

[0024] II. Geometric Prior Cost Field Build: The geometric prior cost field is calculated based on fixed map geometric rules, providing a stable and safe constraint basis for guidance. It includes three parts: obstacle distance transformation term, endpoint heuristic distance term, and obstacle penalty term.

[0025] (1) Obstacle distance transformation diagram Arbitrary grid Euclidean distance to the nearest obstacle: ; A higher value indicates a greater distance from obstacles and a higher level of safety; a lower value indicates proximity to equipment, material piles, or fences and a higher risk of obstruction.

[0026] (2) Normalized heuristic distance graph Calculate the grid To the destination The normalized distance can be any of the Manhattan distance, eight-neighbor distance, or Euclidean distance, used to provide search direction guidance toward the target workstation.

[0027] (3) Formula for calculating the geometric prior cost field: ; In the formula: The penalty weight for proximity to obstacles is increased, amplifying the cost of areas close to obstacles; Weights are assigned to guide the endpoint direction, constraining the search to converge toward the target workstation; A strong penalty coefficient is applied to obstacle areas, imposing an extremely high cost on impassable grids; It is a minimal constant, to prevent The denominator is 0 to avoid calculation overflow.

[0028] This cost field is generated solely from fixed geometric information on the map, making it fully interpretable and ensuring basic driving safety constraints.

[0029] III. Conditional DDIM generation of residual cost field : Geometric priors can only express fixed spatial constraints and cannot learn complex semantic information such as process transfer preferences in digital printing workshops, priority of main logistics channels, temporary dynamic material stacking, and manual operation avoidance. This invention uses conditional DDIM to generate residual cost fields, which locally correct the geometric prior cost fields and do not directly generate complete paths or complete cost maps, thus reducing the difficulty of model training.

[0030] (1) DDIM Multi-Source Conditional Input Tensor Construction: Integrating all workshop space, task, and geometric prior information as generation conditions, the original condition set is as follows: ; After normalization, the components are concatenated along the channel dimension to form the network input conditional tensor: .

[0031] (2) DDIM forward noise addition process (training phase): Let the label of the true residual cost field be... This represents the cost of logistic semantic correction that needs to be added in addition to geometric priors; diffusion time step Lower-band residual field: ; In the formula: This is the cumulative noise attenuation coefficient. It is standard two-dimensional Gaussian noise.

[0032] (3) Noise prediction network: The noisy residual field Time step Conditional tensor Input noise prediction network, output predicted noise: .

[0033] (4) DDIM inverse deterministic denoising inference (real-time planning stage): adopts few-step deterministic DDIM sampling, eliminating the need for random sampling and meeting the real-time planning requirements of AGV; single-step clean residual field estimation: ; go through After iterative denoising, the final residual cost field is output. .

[0034] IV. Generation of the Fusion Cost Field: The geometric prior cost field and the DDIM residual cost field are linearly fused to obtain the final fused cost field used for A* search guidance. : ; The residual modulation coefficient controls the influence of the deep learning semantic correction term on the overall cost field and can be dynamically adjusted according to the degree of workshop congestion.

[0035] V. Differentiable A* path search enhanced by fusion cost field: (1) Improvement of traditional A* defects: standard A* evaluation function Includes only path cumulative cost Endpoint Inspiration Distance The lack of semantic constraints related to workshop safety and logistics means that this invention will integrate the cost field into the evaluation function and update the comprehensive cost value of the nodes. ; For grid nodes The value of integration in the market includes main logistics channels and safe, open areas. Lower values ​​make it easier to be prioritized in search results.

[0036] (2) Matrix-based differentiable A* modeling: The search state is stored in matrix form: open candidate node matrix , The node to be expanded is represented; the closed node matrix has been traversed. , This indicates that the node has completed its neighborhood expansion; This is the global cumulative path cost matrix.

[0037] The node selection matrix is ​​based on softmax differentiability selection of nodes to be expanded. ; In the formula: For the temperature coefficient, select the smoothness of the control node; Multiply the matrix element by element; This refers to matrix inner product operations. To convert the position of the maximum value in the matrix into a one-hot matrix, output the current expanded node.

[0038] The algorithm iteratively expands the node's neighborhood and updates... Matrix, until the target endpoint is reached. Output the complete AGV travel path.

[0039] VI. Multi-objective joint loss function (for offline model training): To adapt the residual cost field generated by DDIM to the path planning task, the network parameters are simultaneously optimized by jointly applying diffusion generation loss, search region loss, path length loss, and safety constraint loss. The total loss function is: ; In the formula: The total weight of the path planning loss; The weighting of loss is constrained to ensure driving safety; the three optimization objectives of planning speed, path length, and safety distance are balanced by adjusting the weights.

[0040] Itemized loss description: (1) Diffusion generation loss Constrained DDIM: Accurately Reproducing the True Residual Cost Field Tags: .

[0041] (2) Path planning loss It is composed of a weighted average of search region loss and path length loss. ; To mark the standard compliance path matrix; To compensate for the search area loss, invalid and redundant nodes visited during the search process are penalized; the smaller the value, the higher the planning efficiency. The path length loss is constrained to minimize the total length of the output path; These are the two loss weights, respectively.

[0042] (3) Safety constraint loss Force the path to maintain a preset safe distance from obstacles. The loss increases when the path point is less than a threshold distance from the obstacle: .

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A heuristic path planning method for AGVs with conditional diffusion cost field enhancement, characterized in that, Includes the following steps: Step S1: Model the digital printing workshop environment as OK Columnar 2D raster map, construct raster occupancy matrix Collect the starting point of the AGV for a single transfer task. Target workstation endpoint Generate the starting point one-hot matrix One-hot matrix at the endpoint ; Step S2: Calculate obstacle distance transformation map based on grid map Endpoint Normalized Heuristic Distance Plot A geometric prior cost field is constructed by combining preset weights. ; Step S3: Integrate the grid occupancy matrix Starting point unique hot matrix One-hot matrix at the endpoint Heuristic distance graph Obstacle distance transformation diagram Geometric prior cost field As input to the conditional DDIM model, conditional information is used to generate the residual cost field. ; Step S4: Apply the geometric prior cost field With residual cost field Linear fusion yields the fusion cost field. ; Step S5: Merge the cost field A node evaluation function with embedded differentiable A* algorithm is used to guide the differentiable A* algorithm to complete node expansion search based on the fusion cost field, and output the planned driving path for material transfer of AGV in digital printing workshop.

2. The AGV heuristic path planning method with conditional diffusion cost field enhancement according to claim 1, characterized in that: In step S2, the formula for calculating the geometric prior cost field is: ； In the formula: Weighting for proximity penalty to obstacles; Weights are assigned to guide the direction of the endpoint; The penalty coefficient is set for the obstacle area. To prevent extremely small constants with a denominator of 0; Obstacle distance transformation diagram , The set of all obstacle nodes within the grid map; heuristic distance graph For grid To the destination The normalized Euclidean distance, Manhattan distance, or eight-neighbor distance.

3. The AGV heuristic path planning method with conditional diffusion cost field enhancement according to claim 1, characterized in that: In step S3, the conditional DDIM model is divided into an offline training phase and a real-time inference phase. The offline training phase executes a forward noise addition process: labeling the true residual cost field. Adding Gaussian noise yields different diffusion time steps Noisy residual field , The cumulative noise attenuation coefficient is the diffusion coefficient. Standard two-dimensional Gaussian noise; Noisy residual field diffusion time step The concatenated conditional tensor Input noise prediction network, output predicted noise .

4. The AGV heuristic path planning method with conditional diffusion cost field enhancement according to claim 3, characterized in that: During the offline training phase, a multi-objective joint total loss function is constructed to optimize the parameters of the conditional DDIM model. The total loss function is as follows: ； In the formula: Loss is generated for diffusion; This is the loss due to path planning. To constrain losses for safety reasons; Weights for path planning loss; Weight the loss for safety constraints; Diffusion generates loss ; Path planning loss , To mark the standard compliance path matrix, For search area loss, For path length loss, These are the weights of the two sub-losses, respectively. Safety constraint loss , A preset safety grid distance threshold is set between the AGV and obstacles; when the distance between the path grid point and the obstacle is less than... At that time, the value of the safety constraint loss increases, forcing the model to generate a cost field with a high safety level.

5. The AGV heuristic path planning method with conditional diffusion cost field enhancement according to claim 3, characterized in that: In step S3, the real-time inference stage uses few-step deterministic DDIM denoising sampling to generate the residual cost field. ; Formula for single-step clean residual field estimation: ； go through The final residual cost field is output after iterative denoising. , The sampling steps are preset, and there is no random noise input during the sampling process, which meets the real-time low-latency planning requirements of AGV.

6. The AGV heuristic path planning method with conditional diffusion cost field enhancement according to claim 1, characterized in that: In step S4, the formula for calculating the fusion cost field is as follows: ； In the formula: The residual modulation coefficient is used to adjust the influence of the residual cost field generated by DDIM on the overall fusion cost field. The fusion cost field also includes the geometric safety constraints of the raster map and the semantic correction information of the digital printing workshop process transfer, main logistics channel, and dynamic obstacles.

7. The AGV heuristic path planning method with conditional diffusion cost field enhancement according to claim 1, characterized in that: In step S5, the differentiable A* algorithm updates the node comprehensive evaluation function as follows: ； In the formula: From the starting point to the current grid node The cumulative path cost; For nodes The heuristic distance to the endpoint; For nodes The value of the fusion cost field; the safe main logistics channel nodes with lower fusion cost field value will be prioritized for expansion by the algorithm.

8. The AGV heuristic path planning method with conditional diffusion cost field enhancement according to claim 7, characterized in that: In step S5, the nodes to be expanded are selected using a matrix-differentiable method. The formula for calculating the node selection matrix is ​​as follows: ； In the formula: This is the global cumulative path cost matrix; This is an open candidate node 0-1 matrix; Temperature coefficient; For element-wise multiplication; It is the inner product of matrices; To find the maximum value, convert the position into a one-hot matrix and output the expanded node for this round; iterate and update the search matrix until the target endpoint is reached.

Images