Leader-following multi-agent path planning algorithm considering task distribution characteristics

Abstract

The invention discloses a leader-following multi-agent path planning algorithm considering task distribution characteristics, which comprises the steps of self-adaptive task clustering, leader path planning and follower path planning, and is characterized in that tasks are grouped by utilizing a self-adaptive clustering method, and then a leader-following two-layer planning mode is adopted, so that the path of the leader-following multi-agent path planning algorithm is obtained; the influence of aggregation conflicts caused by task distribution aggregation is reduced to the minimum, and compared with a conflict-based search method which depends on a conflict detection and rollback strategy, the conflict solving times can be greatly reduced, and the online calculation efficiency of path planning can be improved; differentiated leader paths are generated among the groups, and a similar following strategy is adopted in the groups, so that the search complexity can be remarkably reduced, and high-concentration real-time agent navigation scheduling is realized in a scene with unbalanced large-scale task distribution; according to the self-adaptive task clustering method, grouping is carried out by comprehensively considering the distance and the vector included angle between the tasks, the distribution characteristics of the tasks can be accurately recognized, and clustering is carried out in a self-adaptive mode.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent warehousing technology, and specifically is a "leader-follower" multi-agent path planning algorithm that takes into account task distribution characteristics. Background Art

[0002] With the continuous growth of e-commerce transaction volume and logistics business, the traditional warehousing model is difficult to handle massive order demands in real time under limited manpower and space conditions. Modern warehousing has gradually turned to the "intelligent warehousing" model based on automation technology. In intelligent warehousing, a large number of intelligent agents (such as automatic guided vehicles, robots, etc.) are usually deployed to transport goods. In order to ensure the overall operational efficiency and safety of the warehouse, it is necessary to plan a set of non-conflicting mobile paths for all intelligent agents and minimize the total path cost. This type of problem is called multi-agent path planning problem. In the field of logistics, especially in large-scale warehouses and distribution centers, the distribution of tasks is usually unbalanced, which may lead to a significant decrease in the efficiency of path planning. The present invention can automatically identify the distribution characteristics of tasks according to the distance and angle between tasks through an adaptive task clustering method, thereby effectively grouping tasks and reducing conflicts between tasks through a leader-follower mode. In this scenario, the algorithm can optimize the navigation scheduling of intelligent agents and ensure that intelligent agents in high-density areas can complete tasks quickly and safely. With the expansion of the scale of intelligent warehousing, the path planning algorithm of the present invention shows good scalability. Whether adding more tasks or dealing with more agents participating in scheduling, the algorithm can maintain high efficiency and stability. Especially in large-scale warehousing environments, it has strong adaptability and can adaptively adjust task clustering and path planning strategies, thereby ensuring the efficient operation of the system in complex environments. The present invention is not only applicable to modern intelligent warehousing, but can also be extended to a wider range of intelligent logistics, unmanned distribution and other fields. In the construction of smart cities, the efficient operation of automated logistics distribution systems will be a key component, and the implementation of this algorithm will help improve the efficiency and safety of these systems.

[0003] The multi-agent path planning problem is modeled as follows: In an unweighted undirected graph G = (V, E), V represents a set of vertices, each vertex v∈V represents a position (x, y) in the graph, E represents a set of edges, and each edge e∈E represents a connection between adjacent vertices. There is a set of agents A = {a 1 ,a 2 ,…,a n}, where agent a i Assigned a task τ i , the task consists of a pick-up location τ i .pcik∈V and a delivery position τ i.del∈V. The multi-agent path planning problem generally uses a discrete time model. In each time step, the agent can move to an adjacent location or stay where it is. The problem requires planning conflict-free paths for all agents to avoid: (1) vertex conflicts, that is, two agents occupy the same vertex at the same time step; (2) edge conflicts, that is, two agents traverse the same edge in opposite directions at the same time step, while minimizing the total path cost.

[0004] Existing methods for solving multi-agent path planning problems are divided into optimal algorithms and suboptimal algorithms. The optimal multi-agent path planning algorithm performs a global search in the joint search space. Commonly used algorithms are: (1) A*-based algorithms, which directly use the A* algorithm in the joint search space to perform collaborative search for multiple agents, and (2) conflict-based search algorithms, which use a two-layer search framework to search. In the high-level search, the constraint tree is maintained and conflicts are resolved by adding constraints to conflicting agent pairs. In the low-level search, paths that satisfy the constraints are planned for the agents based on the given constraints. The optimal multi-agent path planning algorithm is generally used to solve relatively simple multi-agent path planning problems. As the complexity of the problem increases, it is difficult for the optimal algorithm to find a solution within an acceptable time. The suboptimal multi-agent path planning algorithm sacrifices a certain degree of optimality in exchange for improved solution speed and efficiency. Compared with the optimal algorithm, it can solve more complex problems. The most widely used suboptimal multi-agent path planning methods are: (1) not using the best-first principle in the conflict-based search algorithm but introducing a heuristic function for node expansion, and (2) assigning a unique priority to each agent and planning paths for the agents in order of priority to avoid conflicts.

[0005] Existing multi-agent path planning algorithms usually assume that the picking and delivery locations of agents are randomly and evenly distributed. However, in today's actual scenarios such as e-commerce warehouses and logistics centers, in order to facilitate picking and assembly management, goods of the same or similar categories are often stored on adjacent shelves, which leads to the clustering of tasks generated in the warehouse. Faced with such a scenario of concentrated task distribution, existing multi-agent methods have the following shortcomings: (1) Existing methods only rely on general conflict resolution mechanisms and lack the ability to perform adaptive planning based on task distribution; (2) The concentrated distribution of tasks will cause high-frequency encounters between agents in local areas, resulting in a large number of clustered conflicts. Existing methods often have to perform a large number of conflict detection and resolution, resulting in a significant decrease in the overall path planning efficiency.

[0006] In view of this, the present invention is proposed to help solve the problems existing in the existing methods. Compared with the traditional path planning method based on conflict detection and fallback strategy, the two-layer planning mode of the present invention significantly reduces the number of conflict resolutions and improves the real-time performance of path planning. This method can effectively cope with large-scale intelligent systems, especially in real-time path updates in dynamic environments, thereby ensuring the stable operation of warehouses under various load conditions and meeting the stringent requirements of minute-level dynamic scheduling in peak scenarios such as the "Double Eleven" e-commerce. Summary of the invention

[0007] The purpose of the present invention is to provide a "leader-follower" multi-agent path planning algorithm that takes into account the distribution characteristics of tasks in order to solve the above problems, and to use the distribution characteristics of tasks in space to accelerate the solution of multi-agent path planning problems. This algorithm first groups tasks with similar pickup and delivery locations, and then uses a two-level planning model to generate leader paths and follower paths respectively, so as to better deal with local high-density conflicts and reduce repeated detours and waiting, thereby solving the problems mentioned in the background technology.

[0008] In order to solve the above problems, the present invention provides a technical solution:

[0009] A "leader-follower" multi-agent path planning algorithm considering the characteristics of task distribution includes the following steps:

[0010] S1, adaptive task clustering: group the task set T according to task similarity and determine a leader task for each group;

[0011] S2, leader path planning: All leader tasks are assigned priorities through a priority determination method based on degree of freedom perception, and the path planning method with minimal time and space overlap is used to plan the paths in priority order to obtain differentiated leader paths;

[0012] S3, follower path planning: by identifying the topological features of the leader path in each group, a set of waypoints is established, and then the A* algorithm based on waypoint guidance is used to plan paths for group members to obtain follower paths similar to the group leader path.

[0013] Preferably, the adaptive task clustering includes: task similarity calculation and adaptive clustering strategy;

[0014] The task similarity calculation calculates the task similarity by considering the distance and vector angle between two tasks;

[0015] The adaptive clustering strategy is used to group the tasks of all agents according to task similarity and determine a leader task for each group.

[0016] Preferably, the task similarity calculation step is:

[0017] The two tasks τ i and τ j The distance between them is denoted as d(τ i ,τ j ), which is calculated as follows:

[0018] d(τ i ,τ j ) = dist(τ i .pick,τ j .pick)+dist(τ i .del,τ j .del)

[0019] Among them, the function dist(v 1 ,v 2 ), used to calculate the Euclidean distance between two vertices;

[0020] The two tasks τ i and τ j The angle between them is denoted as θ(τ i ,τ j ), which is calculated as follows:

[0021]

[0022] Task similarity is calculated through comprehensive calculation:

[0023]

[0024] Preferably, the specific steps of the adaptive clustering strategy are as follows:

[0025] A1, initialize the number of groups N;

[0026] A2. Adaptively select the initial group center with low clustering degree: First, randomly select a task and add it to the group center set centerSet. Then, for each remaining task, calculate its minimum distance to all members in centerSet. The calculation formula is as follows:

[0027]

[0028] Add the task with the largest minDist as the new group center to centerSet, and repeat adding new group centers to centerSet until centerSet contains N group centers;

[0029] A3. All tasks are assigned to the group with the closest group center, and the new group center is recalculated based on the mean of all members in the group. The calculation formula is as follows:

[0030]

[0031] where gp j represents the jth task group, c j Indicates gp j The group center of

[0032] Repeat the above process. When the sum of the distances between all new group centers and the previous generation group centers is less than the preset threshold δ, the grouping ends.

[0033] A4. Calculate the minimum distance between all group centers in each group. The calculation formula is as follows:

[0034]

[0035] A5. Increase the number of groups N by 1 and repeat steps A2-A4 until the average distance of all tasks to their respective group centers is greater than minCenterDist. Then, N stops increasing, thereby adaptively determining the number of groups and designating the group center of each group as the leader task of the group.

[0036] Preferably, the leader path planning includes: a priority determination method based on degree of freedom perception and a path planning method for minimizing space-time overlap;

[0037] The priority determination method based on degree of freedom perception: estimates the degree of freedom according to the distribution of task pickup locations and delivery locations, and assigns priorities according to the degree of freedom of the tasks;

[0038] The method for minimizing space-time overlap path planning: based on the determined priorities, a leader path that minimizes space-time overlap is generated for the leader task in order of priority.

[0039] Preferably, the steps of the priority determination method based on degree of freedom perception are as follows:

[0040] B1. For task τ i , according to its pick-up position τ i .pick and delivery locationτ i .pick distribution to approximate τ i The degree of freedom (τ i ), the calculation formula is as follows:

[0041]

[0042] Among them, Δx and Δy are τ i.pick and τ i .del's horizontal and vertical coordinate differences;

[0043] B2. Assign priorities based on the degree of freedom of tasks, and assign higher priorities to tasks with lower degrees of freedom.

[0044] Preferably, the path planning method for minimizing spatiotemporal overlap comprises the following specific steps:

[0045] C1. Select the leader task to be planned in order of priority. cur , according to the group gp cur The space-time overlap radius K is calculated based on the number of tasks contained in , and the calculation formula is as follows:

[0046]

[0047] C2, using the A* algorithm to minimize the time and space overlap is l cur Perform path planning to minimize the spatiotemporal overlap with higher priority paths by adding a penalty term to the evaluation function of the A* algorithm.

[0048] Preferably, in the follower path planning stage in step S3, a path similar to the leader's path is planned for group members in each group by learning the topological features of the leader's path. The specific steps are as follows:

[0049] S31. Prioritize all task groups according to the priority order of the leader's tasks;

[0050] S32, selecting a task group gp to be planned according to the priority order, and adding the starting position, the end position and all the turning positions in the leader path of the group to the waypoint set wp of the group in order;

[0051] S33, assigning priorities to members in gp through a priority determination method based on degree of freedom perception, and using the waypoint-guided A* algorithm to plan for group members according to priority: in each expansion of the waypoint-guided A* algorithm, the follower (i.e., the group member) will sequentially select a waypoint from wp that is closer to the end point of the leader's path than the current position as a secondary target, and when expanding nodes with the same cost value, the node closer to the secondary target is considered first;

[0052] S34. Repeat steps S32-S34 until planning is completed for all task groups.

[0053] The beneficial effects of the present invention are:

[0054] 1. For multi-agent path planning scenarios with concentrated task distribution, we use an adaptive clustering method to first group tasks reasonably, and then use a two-layer planning model of "leader-follower" to minimize the impact of high-density conflicts caused by task distribution clustering. Compared with the conflict-based search method that relies on conflict detection and fallback, this method can significantly reduce the number of conflict resolutions and improve the overall efficiency of path planning;

[0055] 2. By generating differentiated leader paths between groups and adopting a similar following strategy within the group, the search complexity can be significantly reduced, thereby achieving real-time intelligent agent navigation planning for highly concentrated distributed tasks in large-scale scenarios;

[0056] 3. The adaptive task clustering method comprehensively considers the distance and vector angle between tasks for grouping, which can accurately identify the distribution characteristics of tasks and adaptively cluster them. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] For ease of explanation, the present invention is described in detail with reference to the following specific implementations and the accompanying drawings.

[0058] Figure 1 Flowchart of the “leader-follower” path planning algorithm based on task similarity grouping.

[0059] Figure 2 Example diagram of leader path planning visualization.

[0060] Figure 3 An example diagram of a follower's path planning visualization. DETAILED DESCRIPTION

[0061] like Figure 1 As shown, this specific implementation adopts the following technical solutions:

[0062] Example:

[0063] A "leader-follower" multi-agent path planning algorithm considering the characteristics of task distribution includes the following steps:

[0064] S1. Adaptive task clustering: The task set T is grouped according to task similarity and a leader task is determined for each group. The distance and angle difference between tasks are comprehensively considered for adaptive grouping. It can accurately classify high-density and similar-path tasks in the same area or adjacent areas into the same group.

[0065] S2. Leader path planning: All leader tasks are assigned priorities through a priority determination method based on degree of freedom perception, and the path planning method with minimal time-space overlap is used to plan the paths in order of priority to obtain differentiated leader paths. When planning leader tasks, time-space penalty terms are added to the A* search process to minimize the mutual occupation of time and space between leader paths.

[0066] S3, follower path planning: by identifying the topological features of the leader's path in each group, a waypoint set is established, and then the A* algorithm based on waypoint guidance is used to plan paths for group members, and a follower path similar to the leader's path of the group is obtained. By identifying the topological features of the leader's path, a waypoint set is constructed to guide the remaining followers in the group to learn the leader's path in a local range. This method reduces conflicts within the group while also retaining a certain amount of local detour space, and can improve the feasibility and flexibility of the path while ensuring the similarity with the leader's path.

[0067] Specific:

[0068] Adaptive task clustering includes task similarity calculation and adaptive clustering strategy. Task similarity calculation calculates task similarity by considering the distance and vector angle between two tasks. The adaptive clustering strategy is used to group the tasks of all agents according to task similarity and determine a leader task for each group. When implementing, the two tasks τ i and τ j The distance between them is denoted as d(τ i ,τ j ), which is calculated as follows: d(τ i ,τ j ) = dist(τ i .pick,τ j .pick)+dist(τ i .del,τ j .del),

[0069] Among them, the function dist(v 1 ,v 2 ), used to calculate the Euclidean distance between two vertices;

[0070] The two tasks τ i and τ j The angle between them is denoted as θ(τ i ,τ j ), which is calculated as follows:

[0071]

[0072] Task similarity is calculated through comprehensive calculation:

[0073]

[0074] Initialize the number of groups N; Adaptively select the initial group center with low aggregation: First, randomly select a task and add it to the group center set centerSet, then, for each remaining task, calculate its minimum distance to all members in centerSet;

[0075] The calculation formula is: Add the task with the largest minDist as the new group center to centerSet, and repeat adding new group centers to centerSet until centerSet contains N group centers; assign all tasks to the group with the closest group center, and recalculate the new group center based on the mean of all members in the group. The calculation formula is: where gp j represents the jth task group, c j Indicates gp j The group center of the group is repeated. When the sum of the distances between all new group centers and the previous generation group center is less than the preset threshold δ, the grouping ends. The minimum distance between all group centers in each group is calculated. The calculation formula is: Add 1 to the number of groups N and repeat the above steps until the average distance of all tasks to their respective group centers is greater than minCenterDist, then N stops increasing, thereby adaptively determining the number of groups and designating the group center of each group as the leader task of the group (i.e., the pickup position and delivery position of the leader task are the average of the pickup position and delivery position of all tasks in the task group, respectively);

[0076] According to the task groups divided by the adaptive task clustering method and the leader tasks generated for each group, a leader path with low spatiotemporal overlap is planned for each group in the leader path planning stage. i , according to its pick-up position τ i .pick and delivery locationτ i .pick distribution to approximate τ i The degree of freedom (τ i ), the calculation formula is: Among them, Δx and Δy are τ i .pick and τ i.del’s horizontal and vertical coordinates; assign priorities to agents based on the degrees of freedom of the task, and assign higher priorities to agents with low degrees of freedom tasks; since there may be multiple optimal paths between the pickup location and the delivery location of a task, the number of these optimal paths is affected by the spatial distribution of obstacles, pickup locations, and delivery locations. The number of optimal paths is the degrees of freedom of the task. The higher the degrees of freedom, the more likely the agent performing the task is to bypass the conflict at a lower cost. Based on the above theory, the method for determining the priority of degrees of freedom perception is to generate a leader path that minimizes spatiotemporal overlap for the leader task in order of priority according to the determined priorities. Figure 2 An example diagram of leader path planning visualization is shown;

[0077] Select the leader task to be planned in order of priority l cur , according to the group gp cur The space-time overlap radius K is calculated based on the number of tasks contained in , and the calculation formula is: Use the A* algorithm to minimize the space-time overlap to get l cur Perform path planning by adding a penalty term to the evaluation function of the A* algorithm to minimize the spatiotemporal overlap with higher priority paths: Specifically, for the spatiotemporal vertex to be expanded (v cur ,t), detect with v cur The occupancy of the vertex with K as the center and K as the radius by the higher priority path within the time range [tK, t+K], (v cur ,t) is occupied by 1, and the penalty value of other space-time vertices is set to 1. cur ,t) If the spatial dimension or time dimension differs by one unit, the penalty value is halved;

[0078] In each group, by learning the topological features of the leader's path, a path similar to the leader's path is planned for the group members. All task groups are prioritized according to the priority order of the leader's task, and the task group gp to be planned is selected in order of priority. The starting position, the end position and all the turning positions in the leader's path of the group are added to the waypoint set wp of the group in sequence. The members in gp are assigned priorities through a priority determination method based on degree of freedom perception. The waypoint-guided A* algorithm is used to plan for the group members according to priority: in each expansion of the waypoint-guided A* algorithm, the follower (i.e., the group member) will select a waypoint from wp that is closer to the end position of the leader's path than the current position as the secondary goal in sequence. When expanding nodes with the same cost value, the nodes closer to the secondary goal are considered first. The above steps are repeated until planning is completed for all task groups. Figure 3 Shows an example visualization of the follower path planning.

[0079] While embodiments of the present invention have been shown and described, it will be apparent to those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A "leader-follower" multi-agent path planning algorithm considering the characteristics of task distribution, characterized in that: The steps include: S1, adaptive task clustering: group the task set T according to task similarity and determine a leader task for each group; S2, leader path planning: All leader tasks are assigned priorities through a priority determination method based on degree of freedom perception, and the path planning method with minimal time and space overlap is used to plan the paths in priority order to obtain differentiated leader paths; S3, follower path planning: by identifying the topological features of the leader path in each group, a set of waypoints is established, and then the A* algorithm based on waypoint guidance is used to plan paths for group members to obtain follower paths similar to the group leader path.

2. The "leader-follower" multi-agent path planning algorithm considering the task distribution characteristics according to claim 1 is characterized in that: The adaptive task clustering includes: task similarity calculation and adaptive clustering strategy; The task similarity calculation calculates the task similarity by considering the distance and vector angle between two tasks; The adaptive clustering strategy is used to group the tasks of all agents according to task similarity and determine a leader task for each group.

3. The "leader-follower" multi-agent path planning algorithm considering the task distribution characteristics according to claim 2 is characterized in that: The task similarity calculation steps are: The two tasks τ i and τ j The distance between them is denoted as d(τ i ,τ j ), which is calculated as follows: d(t i ,t j )=dist(τ i .pick,t j .pick)+dist(τ i .del,t j .del) Among them, the function dist(v1,v2) is used to calculate the Euclidean distance between two vertices; The two tasks τ i and τ j The angle between them is denoted as θ(τ i ,τ j ), which is calculated as follows: Task similarity is calculated through comprehensive calculation:

4. The "leader-follower" multi-agent path planning algorithm considering the task distribution characteristics according to claim 2 is characterized in that: The specific steps of the adaptive clustering strategy are as follows: A1, initialize the number of groups N; A2. Adaptively select the initial group center with low clustering degree: First, randomly select a task and add it to the group center set centerSet. Then, for each remaining task, calculate its minimum distance to all members in centerSet. The calculation formula is as follows: Add the task with the largest minDist as the new group center to centerSet, and repeat adding new group centers to centerSet until centerSet contains N group centers; A3. All tasks are assigned to the group with the closest group center, and the new group center is recalculated based on the mean of all members in the group. The calculation formula is as follows: where gp j represents the jth task group, c j Indicates gp j The group center of Repeat the above process. When the sum of the distances between all new group centers and the previous generation group centers is less than the preset threshold δ, the grouping ends. A4. Calculate the minimum distance between all group centers in each group. The calculation formula is as follows: A5. Increase the number of groups N by 1 and repeat steps A2-A4 until the average distance of all tasks to their respective group centers is greater than minCenterDist. Then, N stops increasing, thereby adaptively determining the number of groups and designating the group center of each group as the leader task of the group.

5. The "leader-follower" multi-agent path planning algorithm considering task distribution characteristics according to claim 1 is characterized in that: The leader path planning includes: a priority determination method based on degree of freedom perception and a path planning method for minimizing space-time overlap; The priority determination method based on degree of freedom perception: estimates the degree of freedom according to the distribution of task pickup locations and task delivery locations, and assigns priorities according to the degree of freedom of the tasks; The method for minimizing space-time overlap path planning: based on the determined priorities, a leader path that minimizes space-time overlap is generated for the leader task in order of priority.

6. The "leader-follower" multi-agent path planning algorithm considering the task distribution characteristics according to claim 5 is characterized in that: The steps of the priority determination method based on degree of freedom perception are as follows: B1. For task τ i , according to its pick-up position τ i .pick and delivery locationτ i .pick distribution to approximate τ i The degree of freedom (τ i ), the calculation formula is as follows: Among them, Δx and Δy are τ i .pick and τ i .del's horizontal and vertical coordinate differences; B2. Assign priorities based on the degree of freedom of the task, and assign higher priorities to tasks with low degrees of freedom.

7. The "leader-follower" multi-agent path planning algorithm considering the task distribution characteristics according to claim 5 is characterized in that: The specific steps of the spatiotemporal overlap minimization path planning method are as follows: C1. Select the leader task to be planned in order of priority. cur , according to the group gp cur The space-time overlap radius K is calculated based on the number of tasks contained in , and the calculation formula is as follows: C2, using the A* algorithm to minimize the time and space overlap is l cur Perform path planning to minimize the spatiotemporal overlap with higher priority paths by adding a penalty term to the evaluation function of the A* algorithm.

8. The "leader-follower" multi-agent path planning algorithm considering task distribution characteristics according to claim 1 is characterized in that: In the follower path planning stage in step S3, a path similar to the leader's path is planned for the group members in each group by learning the topological features of the leader's path. The specific steps are as follows: S31. Prioritize all task groups according to the priority order of the leader's tasks; S32, selecting a task group gp to be planned in order of priority, and adding the pickup position, delivery position and all turning positions in the leader path of the group to the waypoint set wp of the group in order; S33, assigning priorities to members in the gp by a priority determination method based on degree of freedom perception, and planning for group members according to priority using an A* algorithm based on waypoint guidance; S34. Repeat steps S32-S34 until planning is completed for all task groups.

Images