A double-layer optimization path planning method for multi-robot task cooperation

Abstract

This invention discloses a two-layer optimized path planning method for multi-robot task collaboration. The method includes establishing a two-layer optimization structure, with a lower layer for path planning and an upper layer for task scheduling. An objective function is constructed using the total path cost of all robot task nodes, and minimizing this objective function is the optimization goal. In the path planning layer, an improved A* algorithm is used to plan paths between nodes, obtaining the shortest feasible path between any two nodes. In the task scheduling layer, an improved artificial bee colony algorithm is used to optimize the task node ranking factor based on the node path cost matrix, obtaining the optimal task code. The optimal task code obtained from the task scheduling layer is decoded, and the execution paths between robot nodes are determined based on the path planning layer. This invention effectively solves the complex multi-robot task allocation problem.

Description

Technical Field

[0001] This invention relates to a two-layer optimized path planning method for multi-robot task collaboration, belonging to the field of mobile robot path planning technology. Background Technology

[0002] With the increasing demand in fields such as intelligent manufacturing, warehousing and logistics, inspection and monitoring, and emergency operations, multi-robot collaborative systems have been widely used in various scenarios. Multi-robot systems not only improve work efficiency but also reduce human intervention, offering unparalleled advantages, especially when handling large-scale, complex tasks. However, as the number of tasks increases and constraints become more complex, multi-robot systems must simultaneously address multiple coupled problems, such as task decomposition, path planning, and execution coordination.

[0003] Currently, path planning for multi-robot collaborative tasks is generally considered an NP-hard problem, with its complexity increasing exponentially with the number of robots and the scale of the task. Traditional path planning methods typically optimize for the shortest path or minimum energy consumption, but these methods often fail to provide effective solutions when faced with complex environments (such as dynamic obstacles, terrain undulations, and task priorities). Furthermore, many existing studies separate task scheduling from path planning, leading to a lack of global consistency. Independent optimization of path planning and task scheduling often results in conflicts between tasks, path redundancy, and load imbalances, thus affecting overall efficiency. This problem is particularly pronounced in large-scale, multi-tasking, and dynamic environments, where existing algorithms cannot adapt to environmental changes in real time, further limiting their application in complex tasks.

[0004] Furthermore, the first and second inventors of this application, as corresponding author and first author respectively, published a method for multi-area collision-free path planning and efficient task scheduling optimization for autonomous agricultural robots in the paper "Yang L, Li P, Wang T, et al. Multi-area collision-free path planning and efficient task scheduling optimization for autonomous agricultural robots [J]. Scientific Reports, 2024, 14(1):18347.", for single-robot multi-area collision-free path planning and efficient task scheduling optimization for autonomous agricultural robots. However, as the number of tasks and robots increases, the single-robot path planning and task scheduling methods are difficult to extend to complex scenarios of multi-robot collaborative operations. Based on the field background and existing research foundation, this invention is proposed. Summary of the Invention

[0005] To address the problems of strong path-scheduling coupling, complex obstacle constraints, and the tendency of traditional methods to get trapped in local optima when multi-robot systems perform multiple tasks in complex environments, this invention provides a two-layer optimized path planning method for multi-robot task collaboration.

[0006] The technical solution of this invention is:

[0007] According to a first aspect of the present invention, a two-layer optimized path planning method for multi-robot task collaboration is provided, comprising:

[0008] S1: Construct a grid environment model that includes a free area, an obstacle area, and a set of task nodes, and set the number of robots, the initial position of the robots, and the position of the task nodes;

[0009] S2: Establish a two-layer optimization structure, with the lower layer being the path planning layer and the upper layer being the task scheduling layer. Construct an objective function based on the total path cost of all robot task nodes and use minimizing the objective function as the optimization objective.

[0010] S3: In the path planning layer, the improved A* algorithm is used to plan the path between each node to obtain the shortest feasible path between any two nodes; the node path cost matrix is ​​constructed based on the shortest feasible path between any two nodes; the improved A* algorithm, which is based on the traditional A* algorithm, includes a neighborhood selection strategy guided by the target orientation and a heuristic function weight dynamic adjustment mechanism.

[0011] S4: In the task scheduling layer, an improved artificial bee colony algorithm is used to optimize the task node sorting factor based on the node path cost matrix to obtain the optimal task code; wherein, the improved artificial bee colony algorithm, which is based on the traditional artificial bee colony algorithm, includes a local-global hybrid search strategy to update the honey source, an adaptive acceleration coefficient, an adaptive abandonment strategy, and an adaptive degree weighting mechanism.

[0012] S5: Decode the optimal task code obtained by the task scheduling layer, and determine the execution path between the nodes of each robot according to the path planning layer.

[0013] Furthermore, the path planning layer in S2 constructs the node path cost matrix using the distance calculated from the shortest feasible path between any two nodes. , ;in, Represents a node With nodes The distance between them , A value of 0 indicates a warehouse node; , Pick This indicates a task node; Indicates the total number of task nodes;

[0014] The task scheduling layer in S2 adopts a dual-structure task coding method. ;in, This represents the first step in the improved artificial bee colony algorithm. Individual bees in a colony; For the number of robots; A sorting factor for task nodes, used to indicate the access order of task nodes; Breakpoint factor, used to divide the task node sorting factors into... There are several sub-paths, each corresponding to a robot execution path; the breakpoint factor satisfies... .

[0015] Furthermore, the objective function in S2 is: The constraints of the objective function are:

[0016] ;

[0017] in, This is the robot's serial number; For binary decision variables, if the robot From the node Move to node ,but Otherwise, it is 0; , A value of 0 indicates a warehouse node; , Pick This indicates a task node; To represent nodes With nodes The distance between them; This represents the total number of task nodes. This refers to the number of robots.

[0018] Furthermore, in S3, a neighborhood filtering strategy based on target orientation guidance is adopted: five of the most valuable expansion directions are dynamically selected from eight neighborhood directions. Specifically, the main guiding direction and the four directions closest to the main guiding direction are determined as expansion directions by the relative direction between the current grid point and the target node.

[0019] Furthermore, the heuristic function weight dynamic adjustment mechanism in S3 is as follows: based on the distance scaling factor... The dynamic adjustment mechanism adjusts the heuristic cost. Weights in the heuristic function; where, This is the adjustment coefficient; The current grid point under the sub-path To the target node The distance; Indicates the starting node in the current raster path under the sub-path. To the target node The distance.

[0020] The local-global hybrid search strategy in S4 updates the honey source expression as follows:

[0021] ;

[0022] in, To improve the artificial bee colony algorithm in the first After the update of the local-global hybrid search strategy in the second iteration Individual bees in a colony To improve the artificial bee colony algorithm in the first Before the first iteration update Individual bees in a colony To improve the artificial bee colony algorithm in the first Before the first iteration update Individual bees in a colony ; Indicates the current size of the bee colony; For the mixed search balance factor, The random factor is used for a local-global hybrid search strategy.

[0023] Furthermore, the adaptive acceleration coefficient in S4 is defined as follows: ;in, and These represent the initial and final acceleration coefficients, respectively; the adaptive acceleration coefficients... Introduce the step size term from the traditional leading bee update formula; This represents the current iteration number. This represents the maximum number of iterations.

[0024] Furthermore, the adaptive abandonment strategy in S4 is as follows:

[0025] ;

[0026] in, To improve the artificial bee colony algorithm in the first After the adaptive abandonment strategy is updated in the nth iteration Individual bees in a colony To improve the artificial bee colony algorithm in the first After the update of the local-global hybrid search strategy in the second iteration Individual bees in a colony To adaptively abandon the random factor of the strategy, and Control large-amplitude and small-amplitude disturbances separately. For the first The number of times an individual bee in a colony has not been improved. The preset threshold for abandonment, The first in the current bee colony The objective function value of each individual This represents the objective function value of the best individual in the entire population during the current iteration.

[0027] Furthermore, the adaptive degree weighting mechanism in S4 is as follows: by introducing a weighting factor. Constructing individual weighted fitness Weighted fitness is defined as ;in, , It is the first in the entire population The objective function value of each individual This represents the average of the objective function for all individuals. These are weighting coefficients. To prevent tiny constants with a denominator of zero, This represents the objective function value of the best individual in the entire population during the current iteration.

[0028] Furthermore, the breakpoint factor of an individual in the improved artificial bee colony algorithm is calculated from the number of tasks performed by each robot: Let the first... The number of tasks for each robot is ,but: Among them, the former Number of tasks for each robot exist Randomly generated within the range, the number of tasks for the last robot is determined by... Sure, , .

[0029] Furthermore, the decoding method in S5 is as follows: ;in, , , The 0 in the code represents a warehouse node; Represents robots Decoding path, Indicates the task node sorting factor for robot 1; Represents robots Task node sorting factor; This represents the task node sorting factor for robot M.

[0030] According to a second aspect of the present invention, a two-layer optimized path planning system for multi-robot task collaboration is provided, comprising modules of the two-layer optimized path planning method for multi-robot task collaboration described in any one of the preceding claims.

[0031] The beneficial effects of this invention are:

[0032] This invention provides a two-layer optimized path planning method for multi-robot task collaboration, which has the following advantages compared to existing technologies:

[0033] (1) This invention adopts a two-layer optimization structure, which organically coordinates the path planning layer and the task scheduling layer: the lower layer constructs the node path cost matrix through the improved A* algorithm to provide accurate path cost input for the upper layer; the upper layer optimizes task allocation through the improved artificial bee colony algorithm. This structure avoids the problems of task conflict, path redundancy and load imbalance caused by independent optimization of path planning and task scheduling in traditional methods, and ensures global consistency.

[0034] (2) This invention makes two key improvements to the traditional A* algorithm: First, a neighborhood selection strategy based on target orientation guides the selection of 5 of the most valuable expansion directions from 8 neighborhood directions, discarding 3 directions that deviate from the target direction; Second, a dynamic adjustment mechanism for heuristic function weights, which introduces a distance scaling factor into the heuristic function. This invention achieves accelerated long-distance advancement in the early stages of the search and enhanced local optimization accuracy as the search approaches the target. Experimental results show that, compared to the traditional eight-neighborhood A* algorithm, the improved A* algorithm of this invention shortens the path length by approximately 5.2%, reduces the number of traversed nodes by approximately 91.2%, and reduces the number of turns by 50%, significantly improving search efficiency while maintaining path quality.

[0035] (3) This invention systematically improves the traditional artificial bee colony algorithm by introducing a local-global hybrid search strategy, adaptive acceleration coefficient, adaptive abandonment strategy, and adaptive degree weighting mechanism, forming a hierarchical mechanism of "global exploration - diversity maintenance - cooperative search - selection adjustment". Ablation experiments show that the synergistic effect of each improved module continuously improves the algorithm performance, reducing the average path cost from 594.86 in the traditional artificial bee colony algorithm to 384.47 in the complete model, an optimization of 35.37%, effectively improving the solution capability and stability of complex multi-robot task allocation problems.

[0036] (4) The present invention demonstrates good scalability in large-scale task allocation experiments. In a high-load scenario with 70 task nodes and 20 robots, the average path length of the method of the present invention is 2420.46 m, which is significantly better than the comparison algorithms such as PSO, WDO, SOA, ABC, AGDO, and DOA, and is still about 2.8% lower than the suboptimal algorithm EO. As the number of robots increases from 2 to 20, the method of the present invention always maintains the lowest path length and the change is stable. The Wilcoxon test results show that the performance improvement is statistically significant, verifying the effectiveness and robustness of the method in multi-robot collaborative scenarios of different scales. Attached Figure Description

[0037] Figure 1 The algorithm flowchart provided for the embodiments of the present invention;

[0038] Figure 2 The terrain environment provided for embodiments of the present invention;

[0039] Figure 3 This is a schematic diagram illustrating the encoding and decoding methods provided in an embodiment of the present invention;

[0040] Figure 4 This invention provides an improved neighborhood selection strategy for the A* algorithm.

[0041] Figure 5 Comparative experiments on the improved A* algorithm path planning provided in the embodiments of the present invention;

[0042] Figure 6 Ablation experiments of the improved artificial bee colony algorithm provided in this embodiment of the invention;

[0043] Figure 7 Results of a large-scale task allocation experiment - path planning layer provided in this embodiment of the invention;

[0044] Figure 8 Results of a large-scale task allocation experiment - task scheduling layer provided in this embodiment of the invention;

[0045] Figure 9 The curves showing the path length variation with the number of robots under different algorithms provided in the embodiments of the present invention;

[0046] Figure 10 The average path length, average time consumption, and statistical significance under different algorithms provided in the embodiments of the present invention are presented. Detailed Implementation

[0047] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0048] Example 1: As Figures 1-10 As shown, a two-layer optimized path planning method for multi-robot task collaboration includes:

[0049] S1: Construct a grid environment model containing a free area, an obstacle area, and a set of task nodes, and set the number of robots, their initial positions, and the positions of the task nodes; wherein, the initial position of all robots is a warehouse node. For example, as shown... Figure 2 As shown, crop areas and non-crop areas are used as obstacle areas, and blank areas are used as free areas.

[0050] S2: Establish a two-layer optimization structure, with the lower layer being the path planning layer and the upper layer being the task scheduling layer. Construct an objective function based on the total path cost of all robot task nodes and use minimizing the objective function as the optimization objective.

[0051] S3: In the path planning layer, the improved A* algorithm is used to plan the paths between each node to obtain the shortest feasible path between any two nodes; the node path cost matrix is ​​constructed based on the shortest feasible path between any two nodes; the improved A* algorithm, which is based on the traditional A* algorithm, includes a neighborhood selection strategy guided by the target orientation and a heuristic function weight dynamic adjustment mechanism.

[0052] S4: In the task scheduling layer, an improved artificial bee colony algorithm is used to optimize the task node ranking factor based on the node path cost matrix to obtain the optimal task code; wherein, the improved artificial bee colony algorithm, which is based on the traditional artificial bee colony algorithm, includes a local-global hybrid search strategy to update the honey source, an adaptive acceleration coefficient, an adaptive abandonment strategy, and an adaptive degree weighting mechanism.

[0053] S5: Decode the optimal task code obtained by the task scheduling layer, determine the execution path between the nodes of each robot according to the path planning layer, and output the path length.

[0054] Furthermore, the path planning layer in S2 constructs a node path cost matrix using the distance calculated from the shortest feasible path between any two nodes (where the nodes are task nodes and warehouse nodes). , The dimension of the matrix is );in, Represents a node With nodes The distance between them (Euclidean distance), , Represents a warehouse node; , The path cost matrix represents the task node (1 represents task node 1, and so on); the path cost matrix serves as the input parameter for the optimization calculation of the task scheduling layer. This indicates the total number of task nodes.

[0055] Furthermore, the task scheduling layer in S2 adopts a dual-structure task coding method. ;in, This represents the first step in the improved artificial bee colony algorithm. Individual bees in a colony; This represents the total number of task nodes. For the number of robots; Sorting factors for task nodes (e.g.) For task nodes The sorting factor (and others similarly) is used to indicate the access order of task nodes; Breakpoint factor, used to divide the task node sorting factors into... There are several sub-paths, each corresponding to a robot execution path; the breakpoint factor satisfies... .

[0056] For example, such as Figure 3 As shown, the number of robots is set to 3, and the number of task nodes is 10; the initial encoding of the task scheduling layer is... After one iteration of the improved artificial bee colony algorithm, the obtained population code is: ; =2 (indicates that a breakpoint will be set after the sorting factor of the second task node). =5 (indicates a breakpoint after the 5th task node sorting factor), forming 3 sub-paths based on the breakpoint factor, namely: sub-path Explanation: This means that robot 2 travels from task node 1 (the starting node of the grid path) through task node 3 to task node 4 (the target node of the grid path), and considers the path between any two adjacent task nodes as a grid path (i.e., ]、[ Two grid paths).

[0057] Furthermore, the objective function in S2 is: The constraints of the objective function are:

[0058] ;

[0059] in, This is the robot's serial number; For binary decision variables, if the robot From the node Move to node ,but Otherwise, it is 0; , Represents a warehouse node. , Indicates a task node; To represent nodes With nodes The distance between them. Equation (1) is the unique access constraint for the target point, Equation (2) is the robot participation constraint, Equation (3) is the path closure constraint, and Equation (4) is the obstacle avoidance constraint, ensuring that the path does not cross obstacles.

[0060] The traditional A* algorithm's neighborhood selection strategy uses eight neighborhood directions as expansion directions. However, considering the possibility of ineffective expansion in these eight directions, this invention improves the neighborhood selection strategy as follows:

[0061] The S3 section employs a neighborhood filtering strategy guided by the target orientation: it dynamically selects five of the most valuable expansion directions from eight neighborhood directions, discarding three directions that deviate from the target direction to reduce invalid expansions and improve search efficiency and directionality. Specifically, it determines the main guiding direction and the four directions closest to the main guiding direction as expansion directions based on the relative direction between the current grid point and the target node, thereby optimizing search focusing capabilities.

[0062] For example, such as Figure 4 As shown, the main guiding direction is determined by the current grid point and the target node, which is the direction indicated by the dashed arrow in the figure, i.e., direction 1. The four directions closest to the main guiding direction are direction 2, direction 3, direction 8, and direction 7 (i.e., among directions 2-8, the Euclidean distance between the grid points corresponding to directions 2, 3, 8, and 7 and the target node is smaller than that between directions 4, 5, and 6). Figure 4 When the main guiding direction is any one of the directions from direction 1 to direction 8, the current grid point constructed by the right half of the grid determines the four directions that are closest to the main guiding direction. Then, based on the main guiding direction and the four directions that are closest to the main guiding direction, the five directions that are extended are retained (e.g., when the main guiding direction is direction 2, the five retained directions are directions 1, 2, 3, 4, and 8). The other three directions are discarded.

[0063] Furthermore, the heuristic function weight dynamic adjustment mechanism in S3 is as follows: the heuristic cost is adjusted through a dynamic adjustment mechanism based on the distance ratio. The weights in the heuristic function accelerate the advancement of distant grid points in the early stages of the search and enhance the accuracy of local optimization as the target approaches. The improved heuristic function. ;in, The adjustment coefficient is 1 in this embodiment of the invention. The current grid point under the sub-path To the target node The distance; Indicates the starting node in the current raster path under the sub-path. To the target node The distance; Indicates starting from the node To the current grid point The cumulative path length.

[0064] The improved artificial bee colony algorithm's population initialization is as follows: the task node sorting factor for each individual in the artificial bee colony algorithm is a randomly generated permutation of 1 to N; the breakpoint factor is obtained by accumulating the number of tasks for each robot, where the first... The number of tasks per robot Random value within ( , The number of tasks for the last robot is made up by the remaining number of tasks, and then converted into the cumulative breakpoint positions.

[0065] For example: Let the total number of task nodes be... =10, Number of robots =3, then = 2, = 10. When initializing a certain entity, first generate a task sorting factor, for example [4,7,2,9,1,5,8,3,6,10]; then randomly generate the task quantity for the first two robots, let... =3, If all are within the range [2, 10], then Finally, calculate the breakpoint factor. = =3, = + =7, resulting in the individual code [4,7,2,9,1,5,8,3,6,10; 3,7].

[0066] Furthermore, the local-global hybrid search strategy in S4 updates the nectar source as follows: Traditional artificial bee colony algorithms rely on differences in nectar sources for location updates. As iterations progress, population differences decrease and step sizes weaken, making it prone to getting trapped in local optima. Therefore, a local-global hybrid search strategy is introduced, initially employing a global search. Later, the focus shifted to local search. For detailed development. Defined as:

[0067] ;

[0068] in, To improve the artificial bee colony algorithm in the first After the update of the local-global hybrid search strategy in the second iteration Individual bees in a colony To improve the artificial bee colony algorithm in the first Before the first iteration update Individual bees in a colony To improve the artificial bee colony algorithm in the first Before the first iteration update Individual bees in a colony ; This indicates the current size of the bee colony; for the first... Each individual remains unchanged; For the mixed search balance factor, For the local-global hybrid search strategy, a random factor is used. (In the embodiments of the present invention, (Randomly select values ​​within a preset range). , This represents the current iteration number. The maximum number of iterations. In this embodiment of the invention, the early and late stages are divided by the maximum number of iterations, with the first iteration being the second iteration. As the initial stage, the remaining iterations are used as the later stage. In this embodiment of the invention, the maximum number of iterations is set to 500, the total population size is 50 (i.e., 50 individuals), and there are three types of bee colonies: leader bees, follower bees, and scout bees. The proportions of leader bees and follower bees are a% and b% respectively (e.g., the size of the leader bee colony is 50% of the total population size, and the size of the follower bee colony is 25% of the total population size, rounded down). The remainder are scout bees.

[0069] Furthermore, the adaptive acceleration coefficient in S4 is defined as follows: In traditional artificial bee colony algorithms, the search step size of individual bees is controlled by a fixed acceleration coefficient, making it difficult to balance the initial global exploration with the later fine-grained search. Therefore, an adaptive acceleration coefficient mechanism is introduced, which dynamically adjusts the perturbation intensity through a linear decay function, defined as... ;in, and These represent the initial and final acceleration coefficients, respectively (in the embodiments of the present invention). Take 1 and (Take a value of 0.1). This adaptive acceleration coefficient will eventually be introduced into the traditional leading bee update formula. This reduces the risk of getting trapped in local optima; among which, Indicates the disturbance factor. .

[0070] Furthermore, the adaptive abandonment strategy in S4 is as follows: Traditional artificial bee colony algorithms use a fixed threshold for nectar source abandonment. Therefore, an adaptive perturbation update mechanism based on historical performance is proposed, which dynamically adjusts the abandonment strategy according to the difference in the objective function value, defined as... ;in, To improve the artificial bee colony algorithm in the first After the adaptive abandonment strategy is updated in the nth iteration Individual bees in a colony For the adaptive abandonment strategy random factor, (In the embodiments of the present invention, (Randomly selected values ​​within a preset range) and Large and small amplitude disturbances are controlled separately (in embodiments of the invention). Take 4. Take 1), For the first The number of times an individual bee in a colony has not been improved. The preset abandonment threshold is set (in this embodiment of the invention, the abandonment threshold is 10). The first in the current bee colony The objective function value of each individual This represents the objective function value of the best individual in the entire population during the current iteration.

[0071] Furthermore, the adaptive weighting mechanism in S4 is as follows: In the traditional artificial bee colony algorithm, to avoid the decline in search ability due to the decrease in population differences as iterations progress, a weighting factor is introduced. Constructing individual weighted fitness This allows individuals with better performance to have a higher probability of being selected; the proposed weighted fitness is defined as... ;in, , It is the first in the entire population The objective function value of each individual This represents the average objective function value (i.e., average cost) for all individuals. It is a weighting coefficient (in this embodiment of the invention, the weighting coefficient is 1). A small constant to prevent the denominator from being zero (taken as 10 in this embodiment of the invention).-12 In this embodiment of the invention, when the iteration terminates, the minimum... The corresponding individual is used as the optimal task code.

[0072] Furthermore, the breakpoint factor of an individual in the improved artificial bee colony algorithm is calculated from the number of tasks performed by each robot: Let the first... The number of tasks for each robot is ,but: Among them, the former Number of tasks for each robot exist Randomly generated within the range, the number of tasks for the last robot is determined by... Sure, , ;in, This represents the total number of task nodes. This refers to the number of robots.

[0073] Furthermore, the decoding method in S5 is as follows: ;in, , , The 0 in the code represents the warehouse code; Represents robots Decoding path, Indicates the task node sorting factor for robot 1; Represents robots Task node sorting factor; Z M Represents the task node sorting factor for robot M; assume optimal task encoding. Through breakpoint factor =2、 =5 forms 3 sub-paths, namely: Then Z1 corresponds to Z2 corresponds to Z3 corresponds to .

[0074] Based on the above-described method of the present invention, the following is an optional experimental verification process for the present invention:

[0075] Experiment 1: Performance Comparison Experiment of Improved A* Algorithm for Path Planning

[0076] To verify the effectiveness of the improved A* algorithm proposed in this invention in path planning, it was tested on a typical agricultural raster map (such as...). Figure 2A comparative experiment was conducted in the map shown. The map size was 104×112 grids, and the robot's starting point was set as the warehouse node (Depot). The traditional four-neighborhood A* algorithm (4-A*), the traditional eight-neighborhood A* algorithm (8-A*), and the improved A* algorithm of this invention were compared. The improved A* algorithm of this invention adopts a neighborhood selection strategy guided by the target orientation, dynamically selecting the 5 most valuable expansion directions from 8 neighborhood directions, discarding the 3 directions that deviate from the target direction to reduce invalid expansion; at the same time, a heuristic function weight dynamic adjustment mechanism is introduced (the adjustment coefficient is 1). 4-A* and 8-A* search based on 4-neighborhood and 8-neighborhood respectively, and the heuristic function adopts the traditional method, i.e. .

[0077] Experimental results are as follows Figure 5 As shown in Table 1. Figure 5 Table 1 shows the path planning results of 4-A*, 8-A*, and the algorithm of this invention, respectively. The gray grid represents the search points visited during the expansion process of each algorithm. As shown in Table 1, the path lengths of 4-A*, 8-A*, and the algorithm of this invention are 120.0000 meters, 97.1543 meters, and 92.0928 meters, respectively; the number of grid points traversed are 2174, 1268, and 111, respectively; and the number of turns are 7, 6, and 3, respectively. Compared with 8-A*, the algorithm of this invention shortens the path length by approximately 5.2%, reduces the number of traversed nodes by approximately 91.2%, and reduces the number of turns by 50%. The results show that the improved A* algorithm proposed in this invention significantly compresses the search space, reduces path irregularity, and reduces computational resource consumption while ensuring path quality. It can effectively support the real-time path planning requirements in complex agricultural environments and provides an efficient and reliable underlying path planning foundation for the construction of node path cost matrices in subsequent multi-robot task scheduling.

[0078] Table 1. Experimental Results of Path Planning

[0079]

[0080] Experiment 2: Ablation Experiment of Improved Artificial Bee Colony Algorithm

[0081] To systematically evaluate the effectiveness of each improved module in the proposed improved artificial bee colony algorithm, an ablation experiment was designed for comparative analysis. Using the traditional artificial bee colony algorithm (ABC) as a baseline, and following a layer-by-layer enhancement approach from global search to local optimization, adaptive acceleration coefficient (M1), adaptive abandonment strategy (M2), local-global hybrid search strategy (M3), and adaptive degree weighting mechanism (M4) were sequentially introduced to construct different algorithm variants. Experimental parameters were set as follows: 4 robots, 40 task nodes (using the VRP standard test set C101), and all algorithms were run independently 30 times. The total path cost was used as the evaluation index, and statistical analysis was performed using mean ± standard deviation, optimal value, worst value, and Wilcoxon rank-sum test. Specific module combinations and their effects are shown in Table 2.

[0082] Table 2 Ablation Module

[0083]

[0084] As shown in Table 3, the algorithm performance continuously improves with the gradual introduction of the improved modules. Compared with the original ABC algorithm, the average path cost significantly decreased from 594.86 to 505.25 after introducing M1 (adaptive acceleration coefficient), representing a relative optimization of 15.1%. Furthermore, the Wilcoxon test p-value was less than 0.05, indicating that this improvement is statistically significant and is the core improvement module. However, after introducing M2 (adaptive abandonment strategy) on top of M1, the average cost rebounded to 568.82, and the performance did not improve but instead decreased. Figure 6 (b) The low contribution of M2 in the Improve dimension of the attribution matrix indicates that M2 mainly plays a role in maintaining population diversity and restoring stagnation, rather than directly optimizing solution quality. When M3 (a local-global hybrid search strategy) is further introduced, the algorithm performance is significantly improved, with the average path cost decreasing to 404.28, representing a 32.0% improvement over the original ABC approach. This indicates that M1 and M3 form an effective synergy between global exploration and local development, which is a key factor in the performance improvement. Finally, after introducing M4 (adaptive weighting mechanism), the average cost is further optimized to 384.47, representing a 4.9% improvement over ABC-M1-M2-M3 and a 35.37% improvement over the original ABC approach. The convergence curve is smoother, combined with… Figure 6 (b) shows that M4 has a relatively high proportion in the Calls dimension, indicating that M4 mainly improves search efficiency indirectly by strengthening the selection mechanism, thus playing an auxiliary enhancement role.

[0085] comprehensive Figure 6 As shown in Table 3, the improved modules exhibit a hierarchical mechanism of action in the algorithm: "global exploration (M1) → diversity maintenance (M2) → cooperative search (M3) → selection regulation (M4)". Figure 6 (a) The path allocation results show that the complete model has a more balanced and compact path allocation, with a significant reduction in redundant paths. The convergence curve shows that the complete model continues to decrease in the later stages, avoiding premature convergence. In the Wilcoxon rank-sum test, the p-value of ABC-M1-M2-M3 relative to the complete model is 0.688 (greater than 0.05), and the result is marked as "≈", indicating that the difference between the two is not significant. That is, the introduction of M4 did not bring a statistically significant improvement, consistent with the evaluation of "auxiliary enhancement effect". The p-values ​​of the other algorithms relative to the complete model are all much less than 0.05, and the result is marked as "+", indicating that the performance difference is significant. In summary, the improved artificial bee colony algorithm proposed in this invention effectively improves the solution capability and stability of complex multi-robot task allocation problems through the synergistic effect of each module.

[0086] Table 3 Ablation Experiment Results

[0087]

[0088] Note: In the Result, "+", "-", and "≈" indicate that the main algorithm has a significant advantage, no significant advantage, and no significant difference compared to the comparison algorithm, respectively; N / A indicates that it is not applicable.

[0089] Experiment 3: Comparative Experiment on Large-Scale Task Allocation Using a Two-Layer Optimization Path Planning Method

[0090] To further evaluate the scalability of the proposed two-layer collaborative optimization method, a large-scale task allocation experiment was conducted. Selected Figure 2 The map shown has 70 task nodes, and the number of robots is increased from 2 to 20 to create a high-load multi-robot collaborative scenario. Figure 7 As shown in (a). The experimental procedure is as follows: First, in the path planning layer, the improved A* algorithm of this invention is used to pre-plan the paths between each node, calculate the shortest feasible path length between any two task nodes, and construct the node path cost matrix. The results are shown in (a). Figure 7 As shown in (b) and (c); secondly, at the task scheduling layer, the aforementioned path cost matrix is ​​used as input parameters, and the improved artificial bee colony algorithm of this invention is used to establish a multi-traveling salesman problem and a sequence constraint model for task allocation optimization. To verify the algorithm performance, Particle Swarm Optimization (PSO), Wind Driven Optimization (WDO), Seagull Optimization (SOA), Balanced Optimizer (EO), Standard Artificial Bee Colony (ABC), Adam Gradient Descent Optimization (AGDO), and Jackal Optimization (DOA) were selected as comparison algorithms. Each group of experiments was run independently 30 times to ensure the reliability of the results. The experimental results are shown in Tables 4 and 5. Figures 8-10 As shown.

[0091] As shown in Tables 4 and 5, taking a high-load scenario with 20 robots as an example, the proposed two-layer collaborative optimization method achieves the best performance. Its average path length is 2420.4615 m, significantly better than PSO (2769.4436 m), WDO (2954.6939 m), SOA (2868.9654 m), ABC (3183.1494 m), AGDO (3282.4243 m), and DOA (2959.7996 m), reducing the path length by approximately 12.6%, 18.1%, 15.6%, 23.9%, 26.2%, and 18.2%, respectively; compared to the suboptimal algorithm EO (2491.4219 m), it still reduces the path length by approximately 2.8%. In terms of optimal performance, the proposed method achieves a mean squared error of 2262.4987 m, outperforming all comparable algorithms. In terms of worst-case performance, the proposed method achieves a mean squared error of 2595.2661 m, significantly outperforming EO (2916.0558 m) and ABC (3251.4531 m). In terms of standard deviation, the proposed method achieves a mean squared error of 76.3814, significantly lower than PSO (135.5481) and EO (211.0921), indicating better robustness in multiple runs. Wilcoxon rank-sum tests show that, except for EO (p=0.0897), all other algorithms are significantly different from the proposed method (p<0.05), verifying the statistical significance of the performance improvement.

[0092] comprehensive Figures 8-10 The experimental results show that, under the condition of a unified path cost matrix, the performance differences among the algorithms mainly stem from their task allocation optimization capabilities. Figure 8 It is evident that comparative algorithms generally suffer from uneven task allocation and path redundancy in large-scale scenarios, while the method of this invention generates a more balanced task division, effectively reducing the overall path cost. Figure 9 The curves showing the change of path length with the number of robots under different algorithms are presented. As the number of robots increases from 2 to 20, the path length of each algorithm shows a decreasing trend, while the method of this invention (ours) always maintains the lowest level and the change is more stable, demonstrating better scalability. Figure 10 Further evidence shows that the method of this invention comprehensively outperforms the comparative algorithms in terms of average path length. Although the average time is slightly higher, it obtains a better solution, demonstrating a more comprehensive search capability. This advantage is mainly attributed to the improved global search and adaptive mechanism of the artificial bee colony algorithm, which maintains the diversity of solutions and avoids premature convergence in the high-dimensional task space, thereby improving stability and robustness. In summary, under the same path modeling conditions, the proposed two-layer optimization method has stronger global optimization capability and better scalability at the task allocation optimization level, verifying its effectiveness and scalability in multi-robot collaborative scenarios.

[0093] Table 4. Results of Large-Scale Task Allocation Experiment

[0094]

[0095] Table 5 Results of Large-Scale Task Allocation Experiment II

[0096]

[0097] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for multi-robot task coordination oriented double-layer optimization path planning, characterized in that, include: S1: Construct a grid environment model that includes a free area, an obstacle area, and a set of task nodes, and set the number of robots, the initial position of the robots, and the position of the task nodes; S2: Establish a two-layer optimization structure, with the lower layer being the path planning layer and the upper layer being the task scheduling layer. Construct an objective function based on the total path cost of all robot task nodes and use minimizing the objective function as the optimization objective. S3: In the path planning layer, the improved A* algorithm is used to plan the path between each node to obtain the shortest feasible path between any two nodes; the node path cost matrix is ​​constructed based on the shortest feasible path between any two nodes; the improved A* algorithm, which is based on the traditional A* algorithm, includes a neighborhood selection strategy guided by the target orientation and a heuristic function weight dynamic adjustment mechanism. S4: In the task scheduling layer, an improved artificial bee colony algorithm is used to optimize the task node sorting factor based on the node path cost matrix to obtain the optimal task code; wherein, the improved artificial bee colony algorithm, which is based on the traditional artificial bee colony algorithm, includes a local-global hybrid search strategy to update the honey source, an adaptive acceleration coefficient, an adaptive abandonment strategy, and an adaptive degree weighting mechanism. S5: Decode the optimal task code obtained by the task scheduling layer, and determine the execution path between the nodes of each robot according to the path planning layer.

2. The two-layer optimized path planning method for multi-robot task collaboration according to claim 1, characterized in that, The path planning layer in S2 adopts the distance of the shortest feasible path between any two nodes to construct a node path cost matrix , ; wherein, denotes the distance between node and node , , takes 0 to represent a warehouse node; , takes to represent a task node; denotes the total number of task nodes; The task scheduling layer in S2 adopts a double-structure task coding mode ; wherein, represents the i-th bee colony individual in the improved artificial bee colony algorithm; is the number of robots; is a task node ordering factor, used to represent the access order of the task node; is a breakpoint factor, used to divide the task node ordering factor into sub-paths, each of which corresponds to a robot execution path; the breakpoint factor satisfies .​ 3. The two-layer optimized path planning method for multi-robot task collaboration according to claim 1, characterized in that, The objective function in S2 is: The constraints of the objective function are: ； in, This is the robot's serial number; For binary decision variables, if the robot From the node Move to node ,but Otherwise, it is 0; , A value of 0 indicates a warehouse node; , Pick This indicates a task node; To represent nodes With nodes The distance between them; This represents the total number of task nodes. This refers to the number of robots.

4. The two-layer optimized path planning method for multi-robot task collaboration according to claim 1, characterized in that, The S3 adopts a neighborhood filtering strategy based on target orientation guidance: five of the most valuable expansion directions are dynamically selected from eight neighborhood directions. Specifically, the main guiding direction and the four directions closest to the main guiding direction are determined as expansion directions by the relative direction between the current grid point and the target node. The heuristic function weight dynamic adjustment mechanism in S3 is as follows: based on the distance scaling factor... The dynamic adjustment mechanism adjusts the heuristic cost. Weights in the heuristic function; where, This is the adjustment coefficient; The current grid point under the sub-path To the target node The distance; Indicates the starting node in the current raster path under the sub-path. To the target node The distance.

5. The two-layer optimized path planning method for multi-robot task collaboration according to claim 1, characterized in that, The local-global hybrid search strategy in S4 updates the honey source expression as follows: ； in, To improve the artificial bee colony algorithm in the first After the update of the local-global hybrid search strategy in the second iteration Individual bees in a colony To improve the artificial bee colony algorithm in the first Before the first iteration update Individual bees in a colony To improve the artificial bee colony algorithm in the first Before the first iteration update Individual bees in a colony ; Indicates the current size of the bee colony; For the mixed search balance factor, The random factor is used for a local-global hybrid search strategy.

6. The two-layer optimized path planning method for multi-robot task collaboration according to claim 1, characterized in that, The adaptive acceleration coefficient in S4 is defined as follows: ;in, and These represent the initial and final acceleration coefficients, respectively; the adaptive acceleration coefficients... Introduce the step size term from the traditional leading bee update formula; This represents the current iteration number. This represents the maximum number of iterations.

7. The two-layer optimized path planning method for multi-robot task collaboration according to claim 1, characterized in that, The adaptive abandonment strategy in S4 is as follows: ； in, To improve the artificial bee colony algorithm in the first After the adaptive abandonment strategy is updated in the nth iteration Individual bees in a colony To improve the artificial bee colony algorithm in the first After the update of the local-global hybrid search strategy in the second iteration Individual bees in a colony To adaptively abandon the random factor of the strategy, and Control large-amplitude and small-amplitude disturbances separately. For the first The number of times an individual bee in a colony has not been improved. The preset threshold for abandonment, The first in the current bee colony The objective function value of each individual This represents the objective function value of the best individual in the entire population during the current iteration.

8. The two-layer optimized path planning method for multi-robot task collaboration according to claim 1, characterized in that, The adaptive weighting mechanism in S4 is as follows: by introducing a weighting factor. Constructing individual weighted fitness Weighted fitness is defined as ;in, , It is the first in the entire population The objective function value of each individual This represents the average of the objective function for all individuals. These are weighting coefficients. To prevent tiny constants with a denominator of zero, This represents the objective function value of the best individual in the entire population during the current iteration.

9. The two-layer optimized path planning method for multi-robot task collaboration according to claim 1, characterized in that, In the improved artificial bee colony algorithm, the breakpoint factor of an individual is calculated based on the number of tasks performed by each robot: Let the breakpoint factor of the individual bee colony be calculated based on the number of tasks performed by each robot. The number of tasks for each robot is ,but: Among them, the former Number of tasks for each robot exist Randomly generated within the range, the number of tasks for the last robot is determined by... Sure, , .

10. The two-layer optimized path planning method for multi-robot task collaboration according to claim 1, characterized in that, The decoding method in S5 is as follows: ;in, , , The 0 in the code represents a warehouse node; Represents robots Decoding path, This represents the task node sorting factor for robot 1; Represents robots Task node sorting factor; This represents the task node sorting factor for robot M.

Images