Prediction-based dual-stage distributed task allocation method

Abstract

The invention relates to the technical field of multi-agent cooperative control, and particularly provides a prediction-based dual-stage distributed task allocation method, in the first stage, an initial solution is generated through task removal and task inclusion, and in the task inclusion stage, a cut-off sensitive cost function is used for expanding a task gap; in the task removal stage, bid prediction is introduced to reduce invalid iteration; in the second stage, the overall scheme is further optimized through local search. The method has remarkable advantages in the aspects of rapid propagation of task information, reduction of the number of iterations and improvement of task distribution efficiency, and is particularly suitable for application scenes with extremely high requirements on task response speed, such as search and rescue. The task information can be efficiently transmitted to the whole cluster in less iteration, so that the collaborative decision process can be greatly accelerated, the task execution efficiency is improved, and the response speed and adaptability of the cluster in the search rescue task are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-agent collaborative control, and more specifically, to a prediction-based dual-stage distributed task allocation method. Background Art

[0002] With the rapid development of robot technology, mobile robots have made remarkable progress in terms of functionality and operational stability. However, a single robot still has limitations in perception, decision-making, and execution capabilities and is difficult to independently complete complex tasks. Through communication and collaboration, multi-robot systems have become an efficient solution for performing tasks in complex environments, and multi-robot task allocation is the core issue in the practical deployment of multi-robot systems, with the goal of optimizing the overall system performance by reasonably allocating tasks.

[0003] In search and rescue scenarios, task allocation faces many practical constraints, such as fuel limitations, deadline requirements, etc., and at the same time, maximizing the rescue of survivors is the primary goal. Such problems generally belong to the task allocation problem under time constraints (ST-SR-TA), which is characterized by strong time dependencies between tasks and a vast solution space, greatly increasing the complexity of the problem.

[0004] According to the different system architectures, multi-robot task allocation methods in search and rescue scenarios can be divided into two categories: centralized and distributed. Centralized methods rely on a central node to collect global information and optimize the objective. Although they can generate relatively optimal solutions, their computational complexity is high, and they are vulnerable to single-point failures and difficult to meet the requirements of large-scale task allocation. In contrast, distributed methods significantly improve the system robustness through decentralized computing and task execution and are more advantageous in practical applications. Therefore, designing efficient and feasible distributed task allocation algorithms to generate reasonable task allocation schemes is more practical and has attracted extensive research attention. Summary of the Invention

[0005] Aiming at the problems of unstable system communication and low task completion rate in multi-robot systems in search and rescue environments, the present invention provides a prediction-based dual-stage distributed task allocation method (Dual-stage distributed taskallocation(PDTA)) to improve the task planning ability of multi-robot systems.

[0006] To solve the above technical problems, the technical solutions adopted by the present invention are as follows: The prediction-based dual-stage distributed task allocation method specifically includes:

[0007] Construct a task allocation objective function based on task requirements;

[0008] Define the time cost function for the robot to execute tasks, and optimize the task allocation strategy based on the time cost function;

[0009] Aiming to maximize the number of allocated tasks, an improved performance impact algorithm is used to generate a task allocation plan;

[0010] A local search mechanism is used to perturb the task allocation plan, and the unallocated tasks are integrated into the task allocation plan to generate the final task allocation plan.

[0011] Furthermore, the task allocation objective function is defined as follows:

[0012]

[0013] where α i represents the task list assigned to robot i, and the number of tasks in the task list |α i | does not exceed its maximum task capacity L i , and robot i executes tasks in sequence according to the task list order, where t iκ (α i ) represents the time to reach the κ-th task , f i represents the maximum travel time of the robot, represents its deadline, h i,j ∈[0,1] represents whether the robot can execute task T j , h i,j =1 means it can execute the task, otherwise it is 0.

[0014] Furthermore, the time cost function is defined as follows:

[0015]

[0016] where C i (α i ) represents the total cost for robot i to execute all the tasks assigned to it, represents the travel time cost of the κ-th task in the task list α i of robot i with number i, represents the task deadline cost of the κ-th task in the task list α i of robot i with number i.

[0017] Furthermore, in the process of optimizing the task allocation strategy based on the time cost function, the tasks with smaller deadline costs are preferentially allocated, and the tasks are quantitatively evaluated by combining the travel time cost and the deadline cost to optimize the task allocation strategy.

[0018] Furthermore, the improved performance impact algorithm includes:

[0019] Based on the traditional performance impact algorithm, by introducing a task prediction mechanism in the task removal stage to predict the bids of competitors, reducing the interference of invalid bids on the allocation process;

[0020] Construct a deadline-aware cost function, and avoid subsequent task allocation blockages by preferentially allocating tasks with earlier deadlines;

[0021] Construct a local search mechanism to broaden the solution space based on the local search mechanism.

[0022] Furthermore, the process of generating a task allocation plan using the improved performance impact algorithm includes:

[0023] Each robot adds the tasks in the assigned task set to the task list until there are no more eligible tasks or the task list reaches its capacity;

[0024] When there is a conflict in the allocation of the same task among robots, execute the task removal process based on the task prediction mechanism to generate the task allocation plan.

[0025] Furthermore, the task list includes a global importance list, a marginal importance list, and a winner list;

[0026] When each robot adds the tasks in the assigned task set to the task list, based on the inclusion of performance impact IPI and removal performance impact RPI to measure the contribution of adding or deleting tasks in the robot task list to the total cost of the robot. When the IPI of the candidate task is less than the global importance, insert the task with the largest importance difference into the task list and update the winner list. When the IPI of the candidate task is not less than the global importance, directly update the importance list.

[0027] Furthermore, executing the task removal process based on the task prediction mechanism includes:

[0028] Determine the conflict task set and the robot i corresponding to the tasks in the conflict task set. If robot i is not the corresponding winner in the winner list, delete the corresponding task in the conflict task set from the task list of robot i;

[0029] Calculate the significance difference, and use the greedy algorithm to identify the task that maximizes the significance difference If Exists in the task lists of both robots i and j at the same time, then predict whether deleting Will result in a reduction in the total cost. If so, continue to delete tasks Otherwise, retain the tasks

[0030] Furthermore, the process of perturbing the task allocation scheme by using the local search mechanism includes:

[0031] Obtaining a task allocation result with reduced cost by swapping the entire task lists between robots or locally swapping individual tasks in their respective task lists;

[0032] Calculating a marginal importance list for all tasks in the unassigned task set and obtaining the highest marginal importance. If there is a task in the unassigned task set that is already included in the task list or does not meet the allocation constraints, assign the marginal importance of this task as M;

[0033] If the highest marginal importance is less than M, insert the tasks in the unassigned task set into the task list of the robot;

[0034] If the highest marginal importance is not less than M, it means that no new tasks can be added. At this time, remove the assigned task with the highest RPI value and calculate the maximum importance difference value between robots If Then further optimize the task allocation result, and then perform local search optimization again until

[0035] Furthermore, the further optimization of the task allocation result is realized based on the following formula:

[0036]

[0037] Wherein, means removing task from the task list of robot and inserting task into the task list of robot at the position, represents the global importance of task T1 in the task list and represents the boundary importance of task T2 in the task list .

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] The present invention constructs an optimization model considering deadlines, fuel limitations, and matching constraints for the multi-robot multi-task allocation problem in search and rescue scenarios, and proposes a prediction-based two-stage distributed task allocation method to maximize the number of rescues.

[0040] The PDTA algorithm proposed in this paper shows outstanding advantages in the task negotiation process. Compared with PI and PIassMax, which rely on multiple rounds of task negotiation and bid updates to optimize task allocation, this method reduces unnecessary task removal operations in the task removal stage by predicting the bids of competing robots, thereby effectively reducing the number of iterations and accelerating consensus.

[0041] The present invention has shown significant advantages in terms of rapid dissemination of task information, reduction of iteration times and improvement of task allocation efficiency, and is particularly suitable for application scenarios such as search and rescue that have extremely high requirements for task response speed. Since task information can be efficiently transmitted to the entire cluster in fewer iterations, this method can greatly speed up the collaborative decision-making process and improve task execution efficiency, thereby enhancing the response speed and adaptability of the cluster in search and rescue tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0043] Figure 1 Schematic diagram of the PDTA algorithm flow chart proposed in the present invention

[0044] Figure 2 It is a flow chart of a traditional PI algorithm in an embodiment of the present invention;

[0045] Figure 3 It is a schematic diagram of the overall process of the task inclusion stage in an embodiment of the present invention;

[0046] Figure 4 Schematic diagram of a conflict resolution process based on prediction in an embodiment of the present invention;

[0047] Figure 5 A flowchart of local search and refined solution in an embodiment of the present invention;

[0048] Figure 6 1 is a communication network topology structure of the PDTA algorithm in an embodiment of the present invention, wherein (a) is a mesh communication topology; (b) is a linear communication topology; (c) is a ring communication topology; (d) is a star communication topology;

[0049] Figure 7 This is a comparison chart of the average number of iterations of different algorithms under high robot task ratio;

[0050] Figure 8 A comparison chart of the average number of tasks assigned to different algorithms under high robot-task ratio;

[0051] Figure 9 Box plot comparison of the number of iterations under four topologies when the robot task ratio is 12:84;

[0052] Figure 10 Box plot comparison of the number of task assignments under four topologies when the robot task ratio is 12:84;

[0053] Figure 11 Box plot comparison of the number of iterations under four topologies when the robot task ratio is 12:96;

[0054] Figure 12 Box plot comparison of the number of task assignments under four topologies when the robot task ratio is 12:96. Detailed implementation manner

[0055] Hereinafter, exemplary embodiments of the present disclosure will be described in more detail with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully conveyed to those skilled in the art. It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other. Hereinafter, the present invention will be described in detail with reference to the drawings and in combination with the embodiments.

[0056] This embodiment proposes a prediction-based two-stage distributed task allocation method, and the algorithm flow is as Figure 1 shown, and the specific content includes:

[0057] Assume that in a post-disaster rescue environment, consider a team of rescue robots, denoted as V = {υ1, υ2,..., υ n}, and a group of m survivors, denoted as T = {T1, T2,..., T m}. Our task objective is to provide emergency support to the survivors, some of which are medicine support tasks and some are food support tasks. To successfully rescue the survivors, appropriate rescue robots must be dispatched to the survivor locations. Each robot can only execute the assigned tasks in the order specified in its task list, and the entire rescue team does not need to return to the starting position. All tasks are classified according to requirements, and each task is only assigned to the robots that can execute them. The matching relationship is defined by the compatibility matrix H, where each element h i,j ∈ [0, 1] indicates whether the robot υ i can execute the task T j , h i,j = 1 means it can execute the task, otherwise it is 0. Each task T j has a deadline η j and a fixed execution time τj , the robot must start executing the task before the deadline. Additionally, the activity time of robot υ i is restricted by fuel, and the maximum travel time is denoted as f i . Due to limited transportation capacity, each robot can be assigned at most L i tasks.

[0058] In the present invention, the goal of task allocation is to maximize the number of survivors successfully rescued while satisfying the constraints related to robot capabilities, task deadlines, fuel limitations, and task matching conditions. Based on the above description, this problem can be written as the following optimization problem:

[0059]

[0060]

[0061] where α i represents the task list assigned to robot i, and the number of tasks in the task list |α i | does not exceed its maximum task capacity L i . Robot i executes tasks in sequence according to the task list, where t iκ (α i ) represents the time to reach the κ-th task , represents its deadline. Additionally, there should be no conflict between the task lists α i and α j for different robots i and j.

[0062] When a multi-robot team executes tasks, the total activity time of each robot can be divided into two parts: task execution time and travel time. The task execution time refers to the cumulative operation time required to complete all tasks in its task list, while the travel time covers the time consumed by the robot moving from the initial position to each task location and between tasks. Since the task execution time is usually relatively fixed, shortening the travel time between adjacent tasks can enable the robot to complete more tasks within the limited activity time. Based on this principle, the present invention combines the travel time with the task deadline cost to quantitatively evaluate the task-related cost, thereby achieving the optimization of task scheduling.

[0063] For the κ-th task in the task list α i of robot numbered i, the travel time cost is defined as follows:

[0064]

[0065] The task deadline cost is defined as:

[0066]

[0067] Therefore, the total cost for robot i to execute all its assigned tasks can be defined as:

[0068]

[0069] The present invention adopts a Performance Impact (PI) algorithm to solve the above problems. This algorithm is based on the concept of market transactions and is specifically designed for the task allocation problem in search and rescue environments. It belongs to a distributed task allocation algorithm. Although it originated from the CBBA algorithm, different from the CBBA algorithm that only focuses on minimizing local costs, the PI algorithm introduces the concept of "importance" to more directly optimize the overall objective - that is, the sum of the arrival times of all robots. The implementation process of the traditional PI algorithm is as Figure 2 shown. During the auction process, the algorithm continuously constructs and updates the task bundles of each robot, and reaches a consistent winning bid result through the conflict resolution stage, finally generating a conflict-free task allocation scheme. However, the traditional PI algorithm has problems such as the inability to maximize the number of task allocations, being prone to falling into local optima, and poor convergence. Therefore, the present invention has improved the traditional PI algorithm in the following aspects:

[0070] (1) Introduce a task prediction mechanism

[0071] Different from the existing methods, the present invention introduces a task prediction mechanism in the task removal stage to predict the bids of competitors. This process significantly reduces the impact of invalid bids on the allocation process, thereby reducing the number of iterations required for the algorithm to converge. In addition, it also enhances the robustness of the algorithm in low-bandwidth communication environments.

[0072] (2) Design a deadline-aware cost function

[0073] A deadline-aware cost function is developed to optimize task allocation by effectively integrating task deadlines. By giving priority to tasks with earlier deadlines, this function avoids inserting tasks with later deadlines prematurely, which may prevent the allocation of other critical tasks. This strategic prioritization can maximize the number of allocated tasks, optimize resource utilization, and improve the overall system efficiency.

[0074] (3) Combine a local search mechanism

[0075] The present invention adopts a local search mechanism to broaden the solution space and ensure that key rescue tasks are not overlooked.

[0076] The specific steps of the present invention are mainly divided into two stages. The first stage is the generation of an initial solution based on prediction, and the second stage is the local search mechanism and refinement strategy.

[0077] I. Generation of an initial solution based on prediction

[0078] In this initial solution generation stage, the core process involves iterative optimization through two main steps: task inclusion and conflict resolution.

[0079] 1.1 Task inclusion

[0080] In the task inclusion stage, each robot sequentially adds the tasks in the unassigned task set to its own task list until there are no more eligible tasks or the task list reaches its capacity. The present invention uses the inclusion performance impact (IPI) and removal performance impact (RPI) to measure the contribution of adding or deleting tasks in the robot task list to the total cost of the robot. Specifically, when inserting task t j into the current task list α of robot i i , the calculation method of IPI is as follows:

[0081]

[0082] where represents the task list α after inserting task t j at position I i . When deleting task t i from the task list α of robot i j , the RPI is defined as follows:

[0083]

[0084] The overall process of the task inclusion stage is described as follows: Each robot i independently selects a candidate task set from the task pool to be executed This candidate set includes tasks that meet the following conditions: First, the task requirement type must match the robot type, represented by h ij = 1. Second, the tasks in the candidate set should not have been assigned to any other robot, that is Subsequently, the robot iteratively adds tasks to its task list within the limit permitted by its capacity (|α i | ≤ L i ). During the task inclusion process, robot i maintains three lists: the global importance list the marginal importance list and the winner list δ i。The global importance list records the final bids of all robots for each task after the previous round of task assignment negotiation phase. The marginal importance list records the local bids of robots for all tasks in this round, and the winner list records the results of task assignment. The overall process is as Figure 3 shown.

[0085] 1.2 Conflict Resolution

[0086] In the conflict resolution stage, it consists of two parts: consensus among robots and prediction-based task removal.

[0087] The consensus process among robots is described as follows: Since in the process of task inclusion, each robot independently constructs its own task list, this may lead to conflicts in the assignment of the same task among robots. To resolve these conflicts, the robot team must coordinate to reach a consensus. Specifically, in the consensus stage, robots communicate through the connected network and broadcast three types of lists among adjacent robots: the winner list δ i , which is used to record the task assignment results; the global importance list ξ iκ , which is used to record the bidding status of robots for tasks; and the timestamp list, which is used to capture the latest communication time to facilitate information synchronization.

[0088] The consensus process follows the communication protocol in the CBBA algorithm. Each robot compares its global importance list with the lists of neighboring robots. For each task, tasks with a higher RPI can significantly reduce the total cost when removed, so they are preferentially assigned to robots with a lower RPI. During this process, the lists are continuously updated and propagated throughout the network. After the consensus stage, all robots obtain a consistent global importance list and winner list, indicating that a consensus has been reached. Then, each robot enters the task removal process to eliminate tasks that are not designated as its own winners.

[0089] In the present invention, different from the PI algorithm that relies on the outdated global importance list in the previous round of consensus for task removal, the method proposed in the present invention adopts a task prediction mechanism. The traditional PI method ignores the fact that other robots are simultaneously performing task removal operations, which may lead to the incorrect removal of tasks that should be retained. To solve this problem, the PDTA method in the present invention uses the exchange of task lists during communication to predict the subsequent bids of other robots, providing a more up-to-date and reliable reference for the task removal stage.

[0090] As Figure 4 shown, the task removal process based on the prediction mechanism is described as follows: First, determine the conflict task set d i , which is composed of tasks in the task list α i , and the winner list δ iThe corresponding winner among them is not robot i. These tasks will be removed from robot i's task list. Similar to the task inclusion process, the decision to continue removing tasks is based on calculating the significant difference. A greedy approach is adopted to identify the tasks that maximize this significant difference. Remove this task from the task list α i and the conflict task set d i Finally, update the cost values of the remaining tasks in α i and the global importance list ξ i .

[0091] Next is the prediction phase. If the task is also included in the task list α of another robot j j , then it is necessary to predict the impact of removing this task from robot j to determine whether the task really needs to be removed. Using the same principle outlined in the previous task removal phase, a new predicted importance list is established for the task If the condition is satisfied, it indicates that removing the task from robot i will not result in a reduction in the total cost. Therefore, robot i retains this task and waits for the results of the subsequent consensus round. On the contrary, if then the task removal process continues, iteratively removing tasks until the conflict set d i is exhausted or this indicates that further removing tasks will not reduce the total cost.

[0092] Subsequently, the task inclusion phase is executed for each robot, and the above procedure is repeated until no changes occur in the consensus, task removal, and task inclusion phases, marking the end of the first stage of the algorithm.

[0093] II. Local Search Mechanism and Refinement Strategy

[0094] Although the algorithm generated an initial solution that satisfies the deadline constraint in the previous step, some tasks may still not be assigned. In an emergency rescue scenario, the highest priority is to maximize the rescue of survivors. Therefore, this embodiment proposes a rescheduling method aimed at optimizing the initial task assignment so that the robot team can undertake additional tasks while ensuring that all constraints are met.

[0095] 2.1 Local Search Mechanism

[0096] As Figure 5As shown, the present invention explores whether it is possible to achieve a lower total cost by integrating unassigned tasks based on the initial solution through a local search mechanism, so that more tasks can be assigned. The present invention uses two methods to conduct the exploration. First, consider two robots with the same performance, i and j, and represent their respective task lists as α i =[α i1 ,α i2 ,...,α iκ and α j =[α j1 ,α j2 ,...,α jλ , where α iκ and α jλ represent the κ-th task and the λ-th task of robots i and j respectively. The first method is to exchange the entire task lists of the two robots, exchange α i with α j to explore the task allocation between robots. The second method requires a local exchange within the task list by exchanging the task α i in α ik with the task α j in α jλ . This method prevents the algorithm from converging to a local optimum and can further explore the solution space.

[0097] 2.2 Refinement Strategy

[0098] After the local search mechanism perturbs the initial solution, a task allocation result D with reduced cost is obtained. Subsequently, similar to the task inclusion process outlined in the initial solution, an attempt is made to integrate the tasks in the unassigned task set U into the current solution.

[0099] The refinement strategy is described as follows: Calculate the marginal importance list ξ * (α υ ,T) for all tasks in the unassigned task set U, where the highest marginal importance is denoted as This list represents how the insertion of each candidate task affects the cost of the current robot. is used as the criterion for deciding whether to include the task.

[0100] When attempting to insert the task T into the task list of robot , if inserting the task T into the task list α of robot υ , if the insertion of this task If T is infeasible (the task is already included in the task list or does not meet the assignment constraints), then assign an infinite value M to the marginal importance of task T to indicate that the task is not selectable at the current stage, thus effectively excluding it from further optimization calculations.

[0101] If it means the possibility of adding other unassigned tasks to the assignment result. Tasks need to be added, and the tasks to be added are determined using the following formula:

[0102]

[0103] where the triple (υ * , T * , ρ * ) indicates that the combination with the minimum marginal importance is in the task list α * of robot υ υ , and task T * is inserted at position ρ * .

[0104] Subsequently, the task list and the unassigned task set U are updated according to the following formula: where represents inserting task T * into the task list of robot at position ρ * , and U = U{T *} means removing task T * from the unassigned task set U.

[0105] If the highest marginal importance it means that new tasks cannot be added through the local search mechanism. In this case, remove the assigned task with the highest RPI value and add unassigned tasks to further refine the solution space. This process aims to reduce the overall cost, and then reapply the local search optimization.

[0106] The maximum importance difference value between robots is as follows:

[0107]

[0108] If it can be inferred that there is further room for optimizing the task assignment result. Then, for the existing assignment results of the robot group V, perform the operations of task removal and re - assignment. First, based on the maximum difference between the marginal importance and the global importance, determine the task - robot pairs that need to be adjusted, and remove the task from the task list of robot ​ removed from, and the task inserted into the robot task list in The position can maximize the gain between the marginal importance and the global importance. Performing these two operations can create empty spaces for the insertion of subsequent tasks. Use the following formula to calculate the robot-task pair.

[0109]

[0110] where represents removing the task from the robot 's task list and inserting the task into the robot task list at the position, represents the global importance of task T1 in the task list and represents the marginal importance of task T2 in the task list .

[0111] This process aims to reduce the overall cost, and then reapply local search optimization until there is no room for improvement in the overall allocation scheme when θ′ ≤ 0, and the algorithm ends.

[0112] Furthermore, this embodiment constructs a simulation model of a post-disaster rescue scenario, aiming to simulate the complexity of the actual rescue environment. Under the condition of limited communication, quickly deploying a heterogeneous unmanned rescue robot team to provide emergency support for survivors becomes a key task. In the system settings: the survivor needs are divided into two categories: medical supplies and food, corresponding to the formation of two professional rescue teams for medicine transportation and food supply; the spatial distribution of survivors is generated randomly in a three-dimensional space of 10,000m × 10,000m × 1,000m (X × Y × Z) using the Monte Carlo method, where the Z-axis coordinate simulates the distribution of different altitude layers; all tasks are set to have equal priority, among which the medicine transportation task has a service time window of 300 seconds, and the time window of the food supply task is extended to 350 seconds; the earliest start time of each task is uniformly set to 0 seconds, and the actual deadline is randomly generated in the interval from 0 to 4,000 seconds; the rescue robots are randomly initialized at the starting positions in a two-dimensional surface space of 10,000m × 10,000m, and are respectively configured with fixed cruising speeds of 30m / s and 50m / s; the fuel endurance time of each robot is limited in the interval from 3,500 to 4,000 seconds.

[0113] ​In this embodiment, the proposed algorithm is compared and analyzed with three well-known distributed task allocation algorithms, CBBA, PI, and PImaxAss. The simulation is carried out under four topologies. To eliminate the influence of randomness, each algorithm runs 50 Monte Carlo simulations by randomly generating task scenarios and uses the same random seed to ensure fairness. The relevant parameter settings in this embodiment are shown in Table 1.

[0114] Table 1

[0115]

[0116] In this embodiment, the task allocation scenarios under different robot task ratios are first tested. The main performance indicators include the average number of task allocations and the average number of algorithm iterations. The number of robots is 8, 12, 14, and 16 respectively, and the ratios are 5, 6, 7, and 8 respectively. Table 2 (experimental results of the average number of task allocations) and Table 3 (experimental results of the average number of iterations) list the simulation results of each algorithm under these different robot task ratios. The maximum and minimum number of iterations for task allocation in each scenario are highlighted in bold.

[0117] As can be seen from Table 2 and Table 3, the test results under the four robot task ratios show that the PDTA algorithm proposed in the present invention is superior to the other three algorithms, achieving a higher average number of task allocations and requiring fewer iterations. It should be noted that when the number of tasks is small, although some tasks are not completed due to an earlier deadline (some robots cannot arrive before the deadline), the proposed algorithm, the PI algorithm, and the PImaxAss algorithm can all complete the remaining tasks with a relatively high completion rate, while the completion rate of the CBBA algorithm is relatively low. This shows that the present invention can effectively reduce the communication volume between robots and give full play to the capabilities of the multi-robot system to maximize the number of rescues.

[0118] Table 2

[0119]

[0120] Table 3

[0121]

[0122]

[0123] As can be seen from Table 2, as the number of tasks increases, the task allocation ability of some algorithms will decrease significantly due to the constraint of the robot fuel limit. The proposed PDTA algorithm shows a significant improvement in task allocation performance, especially in scenarios with a relatively high number of tasks. Further, we conducted further tests with 16 and 18 robots and gradually increased the number of tasks from 32 to 144. Figure 7 andFigure 8 It clearly demonstrates the advantages of the PDTA algorithm in task allocation performance under high task robot ratio.

[0124] In addition, in practical applications, if the distance between robots exceeds the communication range, they cannot directly exchange information. To this end, this embodiment uses four communication network topologies, namely mesh, row, ring, and star, as the communication framework in each iteration to simulate the actual network communication of robot clusters within a limited perception range and evaluate the performance of the algorithm under different topological conditions. These four topological structures are as follows: Figure 6 As shown, it is randomly regenerated in each iteration, two-way communication is adopted between nodes, and the sequence number is for reference only.

[0125] This example systematically compares the performance of four algorithms, PDTA, CBBA, PI, and PIassMax, under different topologies. Figure 9 - 10 and Figure 11 - 12 The number of iterations and task allocation when the robot to task ratio is 12:84 and 12:96 are shown respectively. The results show that in the mesh topology, task information can be quickly propagated to all robots, thus achieving the highest negotiation efficiency and significantly reducing the number of iterations. In contrast, the information propagation range of star and ring topologies is limited, resulting in an increase in the number of iterations, while the linear topology has the highest number of iterations due to the large network diameter and limited information transmission speed.

[0126] The PDTA algorithm proposed in this paper shows outstanding advantages in the task negotiation process. Compared with PI and PIassMax, which rely on multiple rounds of task negotiation and bid updates to optimize task allocation, this method reduces unnecessary task cancellation operations in the task removal phase by predicting the bids of competing robots, thereby effectively reducing the number of iterations and accelerating consensus. In contrast, although CBBA has fewer iterations, it has the lowest number of task allocations, reflecting its limitations; PI and PIassMax perform better in allocating tasks, but at the cost of a higher number of iterations, affecting overall efficiency.

[0127] Based on the traditional Performance Impact (PI) algorithm, the present invention introduces a task prediction mechanism in the task removal stage to predict the bids of competitors, reducing the interference of invalid bids on the allocation process, thereby reducing the number of iterations required for the algorithm to converge and enhancing its robustness in low-bandwidth communication environments. At the same time, a deadline-aware cost function is designed to optimize resource utilization, improve the overall system efficiency, and maximize the number of allocable tasks by preferentially allocating tasks with earlier deadlines to avoid subsequent task blockages. In addition, this method combines a local search mechanism to search within the solution space in one step to optimize the task allocation strategy, ensuring that critical tasks are not overlooked, enabling the robot to maximize the number of survivors successfully rescued in a search and rescue environment, thereby enhancing the rationality and stability of the algorithm.

[0128] Different from the defect of the existing method that may wrongly remove tasks that should be retained, the present invention introduces a task prediction mechanism in the task removal stage. This mechanism utilizes the exchange of task lists during the communication process to predict the possible subsequent bids of other robots, thereby providing a more timely and reliable reference for task removal decisions. By reducing the occurrence of invalid task removals, this method effectively accelerates the convergence speed of the algorithm and improves the efficiency and stability of task allocation.

[0129] The present invention introduces a local search mechanism to expand the solution space and ensure that critical rescue tasks are not overlooked. Although the initial solution meets the deadline constraints, some tasks may still be unallocated, especially in emergency rescues where maximizing survivor rescue is crucial. Therefore, this method proposes a rescheduling strategy to optimize task allocation through local search, enabling the robot to undertake more tasks. Specifically, based on the cost function, the total cost is reduced, and the available time intervals of the task list are increased to facilitate the insertion of additional tasks. The task exchange strategy includes two methods: one is to exchange the robot task lists, and the other is to locally adjust within the task list to optimize the task order. Without changing the total number of tasks, this method improves the task allocation efficiency, avoids local optima, and explores the solution space more comprehensively.

[0130] In summary, the present invention demonstrates significant advantages in terms of rapid task information dissemination, reducing the number of iterations, and improving task allocation efficiency, and is particularly suitable for application scenarios with extremely high requirements for task response speed such as search and rescue. Since task information can be efficiently transmitted to the entire cluster in fewer iterations, this embodiment can significantly accelerate the collaborative decision-making process, improve task execution efficiency, thereby enhancing the response speed and adaptability of the cluster in search and rescue tasks.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent replacements can still be made to the specific implementation manners of the present invention, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A prediction-based two-stage distributed task allocation method, characterized in that, Including: Construct a task allocation objective function based on task requirements; Define the time cost function for the robot to execute tasks, and optimize the task allocation strategy based on the time cost function; Aiming at maximizing the number of allocated tasks, use an improved performance impact algorithm to generate a task allocation plan; Adopt a local search mechanism to perturb the task allocation plan, integrate the unallocated tasks into the task allocation plan, and generate the final task allocation plan.

2. The prediction-based two-stage distributed task allocation method according to claim 1, wherein The task allocation objective function is defined as follows: Among them, α i represents the task list assigned to robot i, and the number of tasks in the task list |α i | does not exceed its maximum task capacity L i . Robot i executes tasks in sequence according to the order of the task list, where t iκ (α i ) represents the time to reach the κ-th task . f i represents the maximum travel time of the robot, represents its deadline, and h i,j ∈[0,1] represents whether the robot can execute task T j . h i,j =1 means that the task can be executed, otherwise it is 0.

3. The prediction-based two-stage distributed task allocation method according to claim 1, wherein The time cost function is defined as follows: Among them, C i (α i ) represents the total cost for robot i to execute all the tasks assigned to it, represents the task list α of robot with number i i in which, the travel time cost of the κ-th task, represents the task list α of robot with number i i in which, the task deadline cost of the κ-th task.

4. The prediction-based two-stage distributed task allocation method according to claim 3, wherein In the process of optimizing the task allocation strategy based on the time cost function, tasks with smaller deadline costs are preferentially allocated, and the travel time cost and deadline cost are combined to quantitatively evaluate the tasks, and the task allocation strategy is optimized.

5. The prediction-based two-stage distributed task allocation method according to claim 1, wherein The process of using an improved performance impact algorithm to generate a task allocation plan includes: Each robot adds the tasks in the allocated task set to the task list until there are no more eligible tasks or the task list reaches its capacity; When conflicts occur in the allocation of the same task among robots, perform a task removal process based on the task prediction mechanism to generate the task allocation plan.

6. The prediction-based two-stage distributed task allocation method according to claim 5, wherein The task list includes a global importance list, a marginal importance list, and a winner list; In the process of each robot adding the tasks in the allocated task set to the task list, based on the inclusion of performance impact IPI and removal performance impact RPI to measure the contribution of adding or deleting tasks in the robot task list to the total cost of the robot. When the IPI of the candidate task is less than the global importance, insert the task with the largest importance difference into the task list and update the winner list. When the IPI of the candidate task is not less than the global importance, directly update the importance list.

7. The prediction-based two-stage distributed task allocation method according to claim 5, wherein Performing the task removal process based on the task prediction mechanism includes: Determine the conflict task set and the robot i corresponding to the tasks in the conflict task set. If robot i is not the corresponding winner in the winner list, delete the corresponding task in the conflict task set from the task list of robot i; Calculate the significant difference and use the greedy algorithm to identify the task that maximizes the significant difference If exists in the task lists of both robots i and j, then predict whether deleting will result in a reduction in the total cost. If so, continue to delete the task Otherwise, keep the task 8. The prediction-based two-stage distributed task allocation method according to claim 1, wherein The process of using a local search mechanism to perturb the task allocation plan includes: Obtain a task allocation result with reduced cost by swapping the entire task lists between robots or locally swapping individual tasks in their task lists respectively; Calculate the marginal importance list for all tasks in the unallocated task set and obtain the highest marginal importance. If there are tasks in the unallocated task set that are already included in the task list or do not meet the allocation constraints, assign the marginal importance of this task to M; If the highest marginal importance is less than M, insert the tasks in the unallocated task set into the task list of the robot; If the highest marginal importance is not less than M, it means that no new task can be added. In this case, the assigned task with the highest RPI value is removed, and the maximum importance difference value among the robots is calculated. If then the task assignment result is further optimized, and then local search optimization is performed again until 9. The prediction-based two-stage distributed task allocation method according to claim 8, wherein Further optimize the task allocation result based on the following formula: Among them, indicates removing the task from the task list of the robot and inserting the task into the position of the task list of the robot . Indicates the global importance of task T1 in the task list . Indicates the boundary importance of task T2 in the task list .

Images