AGV network path planning method and system based on multi-agent path planning algorithm with lazy constraint addition

Abstract

The present application provides an AGV network path planning method and system based on a multi-agent path planning algorithm with lazy constraint addition, the method including receiving input data, initial path generation, non-deterministic node extraction, dynamic path optimization, Monte Carlo path generation, dynamic introduction of alternative paths, and path output. By optimizing the existing multi-agent path planning algorithm based on lazy constraint addition, this method can not only effectively plan collision-free multi-AGV paths in complex environments, but also dynamically adjust and optimize paths to ensure that near-optimal solutions are achieved in real-time operation. By dynamically introducing alternative paths, the flexibility and response speed of path optimization are improved; by utilizing spatially dispersed optimization technology, more evenly distributed paths are pre-calculated, reducing conflicts and congestion between AGVs; combined with the Monte Carlo configuration generation method, the best paths are generated and screened multiple times to ensure that the optimal solution is found within a limited time.

Description

Technical Field

[0001] The present application relates to the technical field of automatic guided vehicles, and in particular to an AGV network path planning method and system based on a multi-agent path planning algorithm with lazy constraint addition. Background Art

[0002] With the rapid development of Industry 4.0 and smart manufacturing, automated guided vehicles (AGVs) are increasingly being used in modern warehousing, logistics, and manufacturing environments. AGVs navigate complex industrial environments using pre-set paths or real-time planning, performing material transportation, loading and unloading, and scheduling tasks. In large warehouses or manufacturing plants, multiple AGVs often need to operate simultaneously, and these AGVs must collaborate efficiently and collision-free within limited space and time. This poses a significant challenge to path planning algorithms, particularly when handling multiple AGVs simultaneously. Path conflicts, computational efficiency, and planning quality all need to be effectively addressed.

[0003] Traditional multi-agent path planning (MAPF) algorithms, such as A*, Dijkstra, and conflict-based search (CBS), perform well in small-scale scenarios. However, in large-scale, multi-agent scenarios, graph-based search algorithms like A* and Dijkstra require traversing all possible path combinations, resulting in exponentially increasing computational complexity with the number of agents. This computational complexity often makes it difficult for these algorithms to reach a solution within a reasonable time, significantly degrading their performance. Furthermore, while algorithms like A* and Dijkstra perform well with a single agent, the need to avoid collisions between agents significantly increases planning complexity when dealing with multiple agents. These algorithms lack effective conflict resolution mechanisms, resulting in poor performance in dynamic and high-density environments. For CBS, conflicts become prevalent in large-scale scenarios, resulting in a large number of subproblems generated after each split. These subproblems form a massive search tree. The rapid expansion of the search space significantly degrades the performance of conflict-based search (CBS) algorithms when considering multiple conflicts, and discrete conflict resolution approaches are even more inefficient.

[0004] Therefore, there is an urgent need for an algorithm that can achieve efficient, collision-free, and near-optimal path planning in complex environments. Summary of the Invention

[0005] This application provides an AGV network path planning method and system based on a multi-agent path planning algorithm with lazy constraint addition, which can be used to solve the technical problem that the performance of traditional multi-agent path planning is significantly reduced when considering multiple conflicts.

[0006] This application provides an AGV network path planning method based on a multi-agent path planning algorithm with lazy constraint addition, the method comprising the following steps:

[0007] Step 1: Receive input data: Receive the starting and target positions of multiple automated guided vehicles (AGVs) and convert them into a graphical model of the path planning problem. The graphical model includes multiple vertices and edges.

[0008] Step 2: Initial path generation: A multi-agent path planning algorithm based on lazy constraint addition generates an initial solution by calculating the shortest path for each AGV. The cost of the path is based on the cumulative conversion cost from the starting position to the target position. The cumulative conversion cost is the sum of the number of moves of all agents in the entire path planning when they do not reach the target state. This metric is often used to measure the quality of path planning. A smaller conversion cost means that the agent reaches the target in fewer steps or time, thus achieving a better path planning solution.

[0009] Step 3, non-deterministic node extraction: During the path optimization process, the non-deterministic node extraction method breaks the limitations of the traditional depth-first search mechanism. Instead of always selecting the shortest path or a fixed optimal node during the optimization phase, randomness is introduced, and multiple feasible nodes are randomly selected for trial. This operation can prevent the search from falling into local optimality, making the path selection more diversified, thereby reducing the probability of different AGVs conflicting at the same node;

[0010] Step 4, Dynamic Path Optimization: Dynamically optimize the initial path using a spatially dispersed path optimization method. Precomputed spatially dispersed paths guide each agent's path selection. In practice, these precomputed paths are distributed as evenly as possible in space and time. Paths are dynamically adjusted to minimize conflicts with other paths. A hyperparameter, m, is introduced to increase the search space and avoid congested areas. This approach effectively reduces conflicts between AGV paths and optimizes path utilization by introducing non-shortest, but more dispersed, paths.

[0011] Step 5, Monte Carlo path generation: During the planning process, the Monte Carlo method is used to generate multiple candidate paths. (The Monte Carlo method is a numerical calculation method based on random sampling, which is used to solve complex problems that are difficult to analyze or calculate using traditional mathematical methods. It estimates the results by simulating a large number of samples of random events.) The optimal path is selected based on the cost function f(Q) = g(Q) + h(Q). The cost function is the sum of the cost g(Q) to the current point and the estimated cost h(Q) to the target. The cost function takes into account the length of the path, conflicts, and time efficiency.

[0012] Step 6: Dynamically introduce alternative paths: During the path planning process, based on the current solution, select some AGVs that may cause conflicts, such as intersections or narrow paths, and replan them based on the solutions found. Alternatively, local search or other path planning algorithms are used to replan and find alternative paths that avoid conflicts, thereby further reducing path conflicts and optimizing AGV movement efficiency.

[0013] Step 7, path output: Output the final collision-free path set to the AGV control system to ensure that each AGV can drive to the target location according to the planned path.

[0014] Furthermore, the multi-agent path planning algorithm based on lazy constraint addition in the initial path generation in step 2 uses the algorithm characteristics of any time. When there are still unexplored nodes in the open stack or no interruption is encountered, the search and optimization will continue. In each search, the algorithm uses the heuristic function h to evaluate the distance between the current node and the target node, and compares it with the existing feasible solutions to find the potential optimal solution.

[0015] Furthermore, in step 2, the cumulative conversion cost formula is as follows:

[0016]

[0017] Where: Q t and Q t+1 are two consecutive configurations at time steps t and t+1, representing the set of positions of all agents;

[0018] cost e (Q t ,Q t+1 ): indicates that the configuration Q t To configure Q t+1 switching costs;

[0019] represents the number of agents that do not remain at their target position during the transition from configuration X to Y; that is, if agent i has not yet reached its target position g i , then it will contribute to the switching cost.

[0020] Furthermore, in step 3, non-deterministic node extraction is performed by randomly selecting a node in the search tree and selecting a node different from the top node of the current stack with a preset probability (for example, 0.01). There are two selection methods: a "restart" strategy that restarts the search from the starting node and a "random" strategy that randomly selects a node in the stack. By introducing this method of non-deterministic node extraction, the search stagnation phenomenon that occurs during the path optimization process is avoided.

[0021] Furthermore, the dynamic path optimization in step 4 utilizes a spatially dispersed path optimization algorithm to pre-calculate several possible conflict-free shortest paths for each AGV and select a path set that minimizes the total path conflicts, thereby reducing path conflicts and congestion between AGVs.

[0022] Furthermore, in step 5, the Monte Carlo path generation includes:

[0023] Step 51, randomly generating a number of candidate paths multiple times;

[0024] Step 52: Calculate the cost of each candidate path: f(Q) = g(Q) + h(Q). The cost is determined by factors such as path length, collision risk, and time overhead.

[0025] Step 53: Select the path with the minimum cost from the candidate paths as the final path.

[0026] Furthermore, step 6 is introduced, and the alternative path dynamics include:

[0027] Step 61 , based on the current AGV path status, select AGVs that may cause conflicts, such as intersections or narrow paths, and generate local alternative paths;

[0028] Step 62, introducing an alternative path generated by a local optimization algorithm or other path planning algorithm;

[0029] Step 63: Replace the existing path to escape the local optimum and re-explore a better path, thereby improving the quality of the solution.

[0030] This application also provides an AGV network path planning system based on an improved multi-agent path planning algorithm with lazy constraint addition. The system is used to implement the method provided in this application, and the system includes:

[0031] An input data receiving module is used to receive the starting positions and target positions of multiple automated guided vehicles (AGVs) and convert them into a graphical model of a path planning problem, where the graphical model includes multiple vertices and edges.

[0032] The initial path generation module is used in the multi-agent path planning algorithm based on lazy constraint addition. It generates an initial solution by calculating the shortest path for each AGV. The cost of the path is based on the cumulative conversion cost from the starting position to the target position. The cumulative conversion cost is the sum of the number of moves of all agents in the entire path planning when they do not reach the target state. This metric is often used to measure the quality of path planning. A smaller conversion cost means that the agent reaches the target in fewer steps or time, thus achieving a better path planning solution.

[0033] The non-deterministic node extraction module is used to break the limitations of the traditional depth-first search mechanism during the path optimization process through a non-deterministic node extraction method. Instead of always selecting the shortest path or a fixed optimal node during the optimization phase, it introduces randomness and randomly selects multiple feasible nodes for trial. This operation can prevent the search from falling into local optimality, making the path selection more diversified, thereby reducing the probability of conflicts between different AGVs at the same node;

[0034] The dynamic path optimization module dynamically optimizes the initial path. Using a spatially dispersed path optimization method, it precomputes a number of spatially dispersed paths to guide each agent's path selection. In practice, these precomputed paths are distributed as evenly as possible in space and time. Paths are dynamically adjusted to minimize conflicts with other paths. A hyperparameter, m, is introduced to increase the search space and avoid congested areas. This approach effectively reduces conflicts between AGV paths and optimizes path utilization by introducing non-shortest, but more dispersed, paths.

[0035] The Monte Carlo path generation module is used to generate multiple candidate paths using the Monte Carlo method during the planning process and select the optimal path based on the cost function f(Q) = g(Q) + h(Q). The cost function is the sum of the cost g(Q) to the current point and the estimated cost h(Q) to the target. The cost function takes into account the length, conflicts, and time efficiency of the path.

[0036] The dynamic introduction module of alternative paths is used to combine the current solution in the path planning process. Based on the solutions found, some AGVs that may cause conflicts, such as intersections or narrow paths, are selected for re-planning or re-planning using local search or other path planning algorithms to find alternative paths that avoid conflicts, thereby further reducing path conflicts and optimizing the movement efficiency of AGVs.

[0037] The path output module is used to output the final collision-free path set to the AGV control system to ensure that each AGV can drive to the target location according to the planned path.

[0038] Furthermore, the multi-agent path planning algorithm based on lazy constraint addition in the initial path generation module uses the algorithm characteristics of any time. When there are still unexplored nodes in the open stack or no interruption is encountered, the algorithm continues to perform search and optimization. In each search, the algorithm uses the heuristic function h to evaluate the distance between the current node and the target node, and compares it with the existing feasible solutions to find the potential optimal solution.

[0039] Furthermore, the cumulative conversion cost formula is as follows:

[0040]

[0041] Where: Q t and Q t+1 are two consecutive configurations at time steps t and t+1, representing the set of positions of all agents;

[0042] cost e (Q t ,Q t+1 ): indicates that the configuration Q t To configure Q t+1 switching costs;

[0043] represents the number of agents that do not remain at their target position during the transition from configuration X to Y; that is, if agent i has not yet reached its target position g i , then it will contribute to the switching cost.

[0044] The dynamic path optimization module is configured to optimize the paths of multiple AGVs through parallel computing. The optimization process includes parallel computing candidate paths, evaluating the cost of each path, and selecting the global optimal path.

[0045] The system can dynamically adjust the planned path according to changes in the actual environment, and use real-time generated alternative paths and optimization algorithms to ensure efficient collaborative work of multiple AGVs and conflict-free paths.

[0046] This application aims to improve the efficiency and optimization performance of AGV path planning in complex dynamic environments, and to solve the real-time, scalability, and near-optimality issues in AGV network path planning. Based on the research on multi-agent path planning (MAPF) problems, this invention proposes an improved multi-agent path planning algorithm based on lazy constraint addition (Lazy Constraint Addition for MAPF), and combines multiple optimization techniques to improve the quality of the initial solution and the convergence speed of path planning, which is suitable for path planning of large-scale AGV clusters.

[0047] The multi-agent path planning algorithm based on lazy constraint addition is a search-based algorithm that, through the lazy constraint addition mechanism, gradually improves the quality of the solution after finding the initial solution, and ultimately converges to the optimal solution. The present invention makes multiple improvements based on this, including the use of non-deterministic search node extraction, the effective use of pre-computed decentralized paths, a Monte Carlo-style successor generation method, and the dynamic introduction of alternative solutions. These improvements enable the improved multi-agent path planning algorithm based on lazy constraint addition to efficiently plan near-optimal paths under large-scale, multi-constraint conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1This is a flow chart of the AGV path planning method of the present invention;

[0049] Figure 2 This is one of the schematic diagrams of the AGV path during the path planning process of the present invention;

[0050] Figure 3 This is the second schematic diagram of the AGV path during the path planning process of the present invention. DETAILED DESCRIPTION

[0051] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0052] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0053] Example 1

[0054] In multi-AGV mission scenarios, each AGV typically needs to perform multiple, continuous transport tasks in a complex environment. This paper refines each AGV's tasks by improving the multi-agent path planning algorithm based on lazy constraint addition. Based on the importance and time sequence of tasks, the complex tasks are broken down into multiple subtasks. Each subtask has a clear starting point and destination, representing the process of the AGV moving from the starting point to the destination within a specific time period.

[0055] After receiving task information from multiple AGVs, the system first uses the task decomposition module to split the overall task of each AGV into multiple subtasks. Each subtask corresponds to an independent path planning request, ensuring the flexibility and accuracy of the task. The specific process is as follows:

[0056] S1. Initialization path planning: The system assigns an initial path to each AGV subtask and uses pre-calculated dispersed paths to ensure that the initial path has good spatial distribution and reduce conflicts between AGVs.

[0057] S2. Initialize global constraints: including avoiding collisions between AGVs, maintaining a safe distance, and avoiding obstacles on the path.

[0058] S3. Set the key parameters of the improved multi-agent path planning algorithm based on lazy constraint addition, such as the adjustment coefficient of the heuristic function, the relaxation factor of the initial path, etc., to ensure that the algorithm can run efficiently in complex environments.

[0059] By decomposing tasks and initializing path planning, the system provides a foundation for subsequent dynamic adjustments and conflict resolution. This mechanism ensures that AGVs can effectively complete their multiple transport tasks and significantly reduces the probability of path conflicts.

[0060] Example 2

[0061] In practical applications, the working environment of AGVs is often highly dynamic. For example, a new transport task may be assigned to an AGV while it is executing an existing task. In addition, unexpected events on the path (such as the sudden appearance of an obstacle or a temporary AGV failure) may also affect the AGV's path planning and task execution. Therefore, this invention proposes a flexible path adjustment mechanism that can respond to these changes in real time and ensure the smooth completion of the AGV's tasks. The specific implementation process is as follows:

[0062] S1. Real-time Task Reassignment: When a new task enters the system, it prioritizes evaluating the completion status of existing tasks and the AGV's current location. For AGVs nearing completion, the system directly assigns them a new task and quickly generates an optimized path from the task's end point to the new task's starting point based on their current location. For AGVs that have not yet completed a task, the system determines the degree of conflict between the new task and the current path and, based on the conflict, decides whether to adjust the existing path.

[0063] S2. Dynamic Path Response: The system features dynamic local path adjustment. If a path conflict or environmental change (such as a new obstacle) is detected during mission execution, the system will not replan the entire path, but will only adjust the affected path. This real-time optimization of the partial path improves system response speed while reducing unnecessary computational overhead, allowing the AGV to quickly circumvent obstacles or replan its path to continue its mission.

[0064] S3. Task Level and Path Priority Adjustment: When the system detects certain high-priority tasks (such as urgent material delivery), it adjusts the path planning priority based on the importance of the task. AGVs carrying out important tasks will be given priority for shorter, less risky paths, while AGVs carrying out other non-urgent tasks will need to adjust their paths based on the new priority to avoid conflicts with high-priority AGVs. This task-level prioritization mechanism ensures that high-priority tasks can be completed promptly and efficiently when multiple tasks are running in parallel.

[0065] This application aims to solve the problem of path conflicts when large-scale and multiple AGVs are running at the same time, and to significantly improve the efficiency of path planning and the quality of solutions. By optimizing the existing multi-agent path planning algorithm based on lazy constraint addition, this method can not only effectively plan collision-free multi-AGV paths in complex environments (such as large warehouses and manufacturing plants), but also dynamically adjust and optimize the paths to ensure that near-optimal solutions are achieved in real-time operation. The key technical improvements of the system include: improving the flexibility and response speed of path optimization by dynamically introducing alternative paths; using spatially dispersed optimization technology to pre-calculate more evenly distributed paths, reducing conflicts and congestion between AGVs; combining the Monte Carlo configuration generation method to generate and screen the best paths multiple times to ensure that the optimal solution is found within a limited time. This system can maintain efficient path planning and execution capabilities when dealing with up to hundreds of AGVs operating at the same time, achieving real-time, collision-free, and near-optimal operation results.

[0066] The above-described embodiments of the present application do not constitute a limitation on the scope of protection of the present application.

Claims

1. An AGV network path planning method based on a multi-agent path planning algorithm with lazy constraint addition is characterized in that: The method comprises the following steps: Step 1: Receive input data: Receive the starting and target positions of multiple automated guided vehicles (AGVs) and convert them into a graphical model of the path planning problem. The graphical model includes multiple vertices and edges. Step 2: Initial Path Generation: A multi-agent path planning algorithm based on lazy constraint addition generates an initial solution by calculating the shortest path for each AGV. The cost of the path is based on the cumulative conversion cost from the starting position to the target position. The cumulative conversion cost is the sum of the number of moves of all agents that did not reach the target state during the entire path planning. Step 3, non-deterministic node extraction: In the optimization phase, instead of always selecting the shortest path or a fixed optimal node, randomness is introduced to randomly select multiple feasible nodes for trial; Step 4, Dynamic Path Optimization: Dynamically optimize the initial path using a spatially dispersed path optimization method. Precomputed spatially dispersed paths are used to guide each agent's path selection. The precomputed paths are distributed as evenly as possible in space and time. The paths are dynamically adjusted to minimize conflicts with other paths. A hyperparameter m is introduced to increase the search space and avoid congested areas. Step 5, Monte Carlo path generation: Generate multiple candidate paths using the Monte Carlo method and select the optimal path based on the cost function f(Q) = g(Q) + h(Q), where the cost function is the sum of the cost g(Q) to the current point and the estimated cost h(Q) to the target. Step 6: Dynamically introduce alternative paths: During the path planning process, based on the current solution, select some AGVs that may cause conflicts, such as intersections or narrow paths, and replan them based on the solutions found. Alternatively, use local search or other path planning algorithms to replan and find alternative paths that avoid conflicts. Step 7, path output: Output the final collision-free path set to the AGV control system to ensure that each AGV can drive to the target location according to the planned path.

2. The path planning method according to claim 1, characterized in that: The multi-agent path planning algorithm based on lazy constraint addition in the initial path generation step in step 2 uses the algorithm feature of any time. When there are still unexplored nodes in the open stack or no interruption is encountered, the algorithm continues to perform search and optimization. In each search, the algorithm uses the heuristic function h to evaluate the distance between the current node and the target node.

3. The path planning method according to claim 1, wherein: In step 2, the cumulative conversion cost formula is as follows: Where: Q t and Q t+1 are two consecutive configurations at time steps t and t+1, representing the set of positions of all agents; cost e (Q t ,Q t+1 ): indicates that the configuration Q t To configure Q t+1 switching costs; represents the number of agents that do not remain at their target position during the transition from configuration X to Y; that is, if agent i has not yet reached its target position g i , then it will contribute to the switching cost.

4. The path planning method according to claim 1, wherein: In step 3, non-deterministic node extraction is performed by randomly selecting a node in the search tree and selecting a node different from the top node of the current stack with a preset probability. There are two selection methods: the "restart" strategy that restarts the search from the starting node and the "random" strategy that randomly selects a node in the stack.

5. The path planning method according to claim 1, wherein: The dynamic path optimization in step 4 utilizes the spatially dispersed path optimization algorithm to pre-calculate several possible conflict-free shortest paths for each AGV and select the path set that minimizes the total path conflicts.

6. The path planning method according to claim 1, characterized in that: Step 5, Monte Carlo path generation includes: Step 51, randomly generating a number of candidate paths multiple times; Step 52, calculate the cost of each candidate path f(Q) = g(Q) + h(Q); Step 53: Select the path with the minimum cost from the candidate paths as the final path.

7. The path planning method according to claim 1, characterized in that: Introducing step 6, the alternative path dynamics include: Step 61 , based on the current AGV path status, select AGVs that may cause conflicts, such as intersections or narrow paths, and generate local alternative paths; Step 62, introducing an alternative path generated by a local optimization algorithm or other path planning algorithm; Step 63: Replace the existing path, escape from the local optimum, and re-explore a better path.

8. An AGV network path planning system based on a multi-agent algorithm with lazy constraint addition, the system being used to implement the method of any one of claims 1 to 7, characterized in that: The system comprises: An input data receiving module is used to receive the starting positions and target positions of multiple automated guided vehicles (AGVs) and convert them into a graphical model of a path planning problem, where the graphical model includes multiple vertices and edges. The initial path generation module is used for the multi-agent path planning algorithm based on lazy constraint addition. It generates an initial solution by calculating the shortest path for each AGV. The cost of the path is based on the cumulative conversion cost from the starting position to the target position. The cumulative conversion cost is the sum of the number of moves of all agents that did not reach the target state during the entire path planning. The non-deterministic node extraction module is used to introduce randomness during the optimization phase, rather than always selecting the shortest path or a fixed optimal node. Instead, it randomly selects multiple feasible nodes for trial. The dynamic path optimization module dynamically optimizes the initial path. It uses a spatially dispersed path optimization method to precompute a number of spatially dispersed paths to guide each agent's path selection. The precomputed paths are distributed as evenly as possible in space and time. The paths are dynamically adjusted to minimize conflicts with other paths. A hyperparameter m is introduced to increase the search space and avoid congested areas. The Monte Carlo path generation module is used to generate multiple candidate paths using the Monte Carlo method and select the optimal path based on the cost function f(Q) = g(Q) + h(Q), where the cost function is the sum of the cost g(Q) to the current point and the estimated cost h(Q) to the target point. The dynamic introduction module of alternative paths is used to combine the current solution in the path planning process and select some AGVs that may cause conflicts, such as intersections or narrow paths, based on the solutions found, to replan or use local search or other path planning algorithms to find alternative paths that avoid conflicts; The path output module is used to output the final collision-free path set to the AGV control system to ensure that each AGV can drive to the target location according to the planned path.

9. The system according to claim 8, characterized in that The multi-agent path planning algorithm based on lazy constraint addition in the initial path generation module uses the algorithm feature of any time. When there are still unexplored nodes in the open stack or no interruption is encountered, the algorithm continues to perform search and optimization. In each search, the algorithm uses the heuristic function h to evaluate the distance between the current node and the target node.

10. The system according to claim 8, wherein: The cumulative conversion cost formula is as follows: Where: Q t and Q t+1 are two consecutive configurations at time steps t and t+1, representing the set of positions of all agents; cost e (Q t ,Q t+1 ): indicates that the configuration Q t To configure Q t+1 switching costs; represents the number of agents that do not remain at their target position during the transition from configuration X to Y; that is, if agent i has not yet reached its target position g i , then it will contribute to the switching cost.

Images