A multi-agent task scheduling method and system based on task decomposition and collaborative strategy optimization

Abstract

This invention relates to the field of task prediction and scheduling technology, and discloses a multi-agent task scheduling method and system based on task decomposition and collaborative strategy optimization. The method includes: constructing a multi-dimensional state space; generating pre-decomposed decision information; performing spectral clustering and dynamic capability weight decomposition to obtain sub-task packages; forming an allocation scheme through path planning; and obtaining a corrected scheme by monitoring prediction deviations and replanning. Compared to existing technologies where fixed threshold triggering decomposition leads to lag, especially under conditions of sudden peaks in order tasks, dynamic coupling of task queue growth rate and robot load, the present application improves load balancing and response capabilities through exponential smoothing prediction and dynamic capability weight decomposition mechanisms.

Description

Technical Field

[0001] This invention relates to the field of task prediction and scheduling technology, and in particular to a multi-agent task scheduling method and system based on task decomposition and collaborative strategy optimization. Background Technology

[0002] Currently, in e-commerce warehousing and sorting systems, dozens of handling robots receive order picking tasks and perform handling through a central scheduling system. Existing scheduling methods generally adopt a first-come, first-served approach or a greedy allocation strategy based on immediate load, triggering batch splitting or reallocation when the task queue length exceeds a preset threshold. This approach treats task decomposition as a passive response to an existing backlog, and its decision-making is mostly based on a static snapshot of the current moment. Intervention is only initiated when the absolute number of tasks to be processed or the instantaneous waiting time exceeds a threshold, lacking forward-looking inferences about the rate of task arrival and the evolution of the robot swarm's capabilities.

[0003] For example, during major e-commerce promotions or sudden surges in orders, the task arrival rate can increase several times within minutes. However, the threshold-triggered decomposition mechanism can only initiate splitting after the queue length significantly exceeds the static value. At this point, a large number of tasks are already waiting, and some robots cannot be assigned tasks due to insufficient remaining power, path conflicts, or reaching the maximum concurrent capacity, causing instantaneous congestion. Meanwhile, other robots with spare capacity are not assigned tasks because they were not pre-split into subtasks of matching granularity, resulting in alternating oscillations of local idleness and local overload in the system. This decomposition method based on lagging observation cannot use trend signals such as the net increase in task acceleration to predict future congestion moments, nor can it automatically split large-scale task packages into a set of subtasks adapted to the dynamic capabilities of each robot before congestion actually occurs, making response lag and severe load imbalance the norm.

[0004] Therefore, there is an urgent need for a method that can predict future congestion moments based on the real-time task queue growth rate and the current load of each robot under sudden peak order tasks, and automatically decompose a large number of tasks into sub-tasks with matching granularity before congestion actually occurs. This method should complete pre-decomposition decisions and sub-task package generation when the task arrival rate increases sharply but the task pool is not yet completely congested. This allows the allocation mechanism to push appropriately granular tasks to robots with matching capability weights in advance, thereby maintaining high throughput and low latency even when the system faces impactful loads, improving scheduling foresight, and mitigating resource waste and heavy load imbalance caused by response delays. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the present invention aims to propose a multi-agent task scheduling method based on task decomposition and collaborative strategy optimization. This method addresses the technical problems of delayed response and unbalanced heavy load in existing technologies, particularly under conditions of sudden peaks in order tasks, dynamic coupling of task queue growth rate and robot load. These conditions prevent the method from anticipating future congestion and automatically decomposing tasks into sub-tasks with matching granularity.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a multi-agent task scheduling method based on task decomposition and cooperative strategy optimization.

[0007] The multi-agent task scheduling method based on task decomposition and cooperative strategy optimization includes: Step S10: Obtain the original data of the task queue and the original data of the robot cluster. Based on the original data of the task queue and the original data of the robot cluster, perform the state representation task using a multi-dimensional state feature aggregation mechanism and output a multi-dimensional state space. Step S20: Based on the multidimensional state space, the congestion early warning mechanism using Holt exponential smoothing prediction is used to perform a pre-decomposition decision task and output pre-decomposition decision information; Step S30: Based on the pre-decomposition decision information, an adaptive decomposition task is executed using a decomposition mechanism of spectral clustering and dynamic capability weights, and a set of subtask packages is output. Step S40: Based on the set of sub-task packages, use A The path planning and allocation mechanism performs task allocation and outputs the allocation plan. Step S50: Based on the allocation scheme, a pullback and replanning mechanism with prediction deviation monitoring is used to perform a closed-loop correction task and output the corrected allocation scheme.

[0008] Preferably, step S10, which involves acquiring the original data of the task queue and the original data of the robot cluster, and performing a state representation task based on the original data of the task queue and the original data of the robot cluster using a multi-dimensional state feature aggregation mechanism to output a multi-dimensional state space, specifically includes: Step S101: The scheduling center collects global data at a fixed frequency and defines the task queue state vector. ,in for The total number of tasks to be assigned at any given time. Preset data acquisition time step Net increase in tasks within the period for The total number of tasks to be assigned at any given time is defined as the task arrival rate. , To characterize the net incremental acceleration of a task reaching its burst intensity, ; The variance of the estimated execution time for all tasks in the task pool. Set the average priority of tasks; Step S102: Define the robot cluster state matrix For any robot in the robot swarm Its state vector includes: the x-coordinate of the current position. Current position ordinate Number of assigned but incomplete tasks Dynamic average processing rate Remaining battery percentage Movement state Power consumption coefficient fitted from historical data Expected battery life and maximum number of concurrent tasks ; Step S103: Convert the task queue state vector With the robot cluster state matrix The aggregation forms the multidimensional state space.

[0009] Preferably, step S20, which involves performing a pre-decomposition decision-making task based on the congestion warning mechanism using Holt exponential smoothing prediction in the multi-dimensional state space and outputting pre-decomposition decision information, specifically includes: Step S201: Set the task arrival rate and the acceleration of net increase in tasks Holt's two-parameter exponential smoothing method was used to extract the smoothed arrival rate. and smooth acceleration ,in

[0010]

[0011] in, and For smoothing coefficients; express Smooth arrival rate corresponding to each time point; Step S202: Calculate the system's dynamic digestion capacity index

[0012]

[0013] in The total number of robots, For indicator functions, Set a preset warning threshold for running out of power; Step S203: In the preset prediction window Get the future from within Expected task queue length at time ; Step S204: Define the expected congestion intensity index

[0014]

[0015] When both conditions are met and Furthermore, when the above two conditions are continuously met for a preset number of consecutive periods, the pre-decomposition decision information is generated; wherein, This is a preset congestion threshold.

[0016] Preferably, in step S202, the and These are used to exclude situations where the remaining battery power is below a preset battery depletion warning threshold. The number of robots and the number of assigned but unfinished tasks have reached the maximum concurrent task capacity. The robot.

[0017] Preferably, step S30, which involves performing an adaptive decomposition task based on the pre-decomposition decision information using a spectral clustering and dynamic capability weighting decomposition mechanism, and outputting a set of subtask packages, specifically includes: Step S301: Obtain the current task pool to be decomposed from the pre-decomposition decision information, regard each task in the current task pool to be decomposed as a node, construct a task association graph, and define the weight of the edges between nodes;

[0018] in For the task and Manhattan distance from the starting shelf coordinates For the task and The difference between the latest start times, Number the aisle of the shelf where the task is located. Using weighted coefficients, the Shi-Malik spectral clustering algorithm is used to divide the task association graph into... A cluster of tasks; Step S302: For any robot in the robot swarm Define load margin factor , Efficiency factor Battery life factor ,in, To preset the minimum estimated time window for completing an emergency task, Indicates that the robot swarm is The maximum dynamic average processing rate at any given time is calculated, and the overall capability value is determined. , ,in Adjustable weights; Step S303: Based on comprehensive capability value The Task clusters can be combined or broken down into A set of subtask packages is formed by subtask packages with a granularity equal to the quota.

[0019] Preferably, in step S40, A is used based on the set of sub-task packages. The path planning and allocation mechanism performs task allocation and outputs an allocation plan through the following steps: Step S401: Define the robot Execute subtask package The cost

[0020]

[0021] in To adopt A Heuristic algorithm planning for robots Current location to subtask package The optimal path travel time from the center of the starting region. To evaluate the subtask package based on the maintained global spatiotemporal occupancy grid map Assigned to robot The additional waiting or detour time introduced when its path conflicts with the assigned robot path. Indicates the current allocation status. Assigning robots based on assumptions Loading factor, These are the weighting coefficients; Step S402: Based on cost Tasks are assigned, and each robot i is checked individually to see if the total number of tasks assigned to it exceeds its maximum concurrent task capacity. If the limit is exceeded, the excess sub-task packages will be forcibly abandoned, and the final allocation scheme will be output.

[0022] Preferably, step S50, which involves executing a closed-loop correction task based on the allocation scheme using a prediction deviation monitoring and replanning mechanism, and outputting the corrected allocation scheme, specifically includes: Step S501: Calculate prediction deviation in real time ,in for Predicted task queue length at any given time. for The actual task queue length at any given time, when The deviation exceeds the preset threshold, or the remaining battery percentage of any robot. If the price drops by more than a preset percentage within a preset time, a correction will be triggered. Step S502: After triggering the correction, mark all subtasks currently being executed by all robots as frozen, and simultaneously revert all subtasks that have been assigned but not yet started in each robot's list, along with the remaining tasks in the task pool, to the waiting-to-be-assigned pool. ; Step S503: Use the pool to be allocated For the new task pool, and with the frozen subtasks as hard constraints, steps S20 to S40 are re-executed to finally generate the revised allocation scheme.

[0023] This invention also provides a multi-agent task scheduling system based on task decomposition and cooperative strategy optimization, comprising: The state aggregation module is used to acquire the original data of the task queue and the original data of the robot cluster, and to perform the state representation task based on the original data of the task queue and the original data of the robot cluster using a multi-dimensional state feature aggregation mechanism, and output a multi-dimensional state space. The congestion warning module is used to perform a pre-decomposition decision-making task based on the congestion warning mechanism using Holt exponential smoothing prediction based on the multi-dimensional state space, and output pre-decomposition decision information. The task decomposition module is used to perform adaptive decomposition tasks based on the pre-decomposition decision information using a decomposition mechanism of spectral clustering and dynamic capability weights, and output a set of sub-task packages. The task allocation module is used to assign tasks based on the set of sub-task packages using A. The path planning and allocation mechanism performs task allocation and outputs the allocation plan. The closed-loop correction module is used to perform closed-loop correction tasks based on the allocation scheme using a rollback and replanning mechanism that monitors prediction deviations, and outputs the corrected allocation scheme.

[0024] The present invention also provides a multi-agent task scheduling device based on task decomposition and cooperative strategy optimization. The multi-agent task scheduling device based on task decomposition and cooperative strategy optimization includes: a memory, a processor, and a multi-agent task scheduling program based on task decomposition and cooperative strategy optimization stored in the memory and executable on the processor. When the multi-agent task scheduling program based on task decomposition and cooperative strategy optimization is executed by the processor, it implements the above-mentioned method.

[0025] The present invention also provides a computer program product, the computer program product including a multi-agent task scheduler based on task decomposition and cooperative strategy optimization, the multi-agent task scheduler based on task decomposition and cooperative strategy optimization implementing the above method when executed by a processor.

[0026] The beneficial effects of this invention are as follows: This invention constructs a multi-dimensional state feature aggregation mechanism, which couples the arrival burst intensity of the task queue, the heterogeneity of task time, and the dynamic processing capability, battery life, and load margin of the robot cluster into a unified state space. This overcomes the problem of information separation between the task domain and the robot domain in traditional scheduling, and provides a more complete decision basis for congestion prediction.

[0027] By introducing congestion warning based on Holt exponent smoothing and adaptive decomposition of dynamic capability weights combined with spectral clustering, the system can complete the pre-decomposition of global capabilities before the task queue expands rapidly. This avoids the surge in path conflicts and the extension of robot waiting time caused by delayed decomposition, making task allocation more in line with the actual processing rhythm of the robots and improving the overall utilization rate and task completion time of the robot cluster. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the first embodiment of a multi-agent task scheduling method based on task decomposition and collaborative strategy optimization according to the present invention. Detailed Implementation

[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0030] 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.

[0031] Example 1: As Figure 1 The diagram shown is a flowchart of a first embodiment of a multi-agent task scheduling method based on task decomposition and cooperative strategy optimization according to the present invention.

[0032] In the first embodiment, the multi-agent task scheduling method based on task decomposition and cooperative strategy optimization includes: Step S10: Obtain the original data of the task queue and the original data of the robot cluster. Based on the original data of the task queue and the original data of the robot cluster, perform the state representation task using a multi-dimensional state feature aggregation mechanism and output a multi-dimensional state space. In step S10, the task queue raw data collected by the scheduling center at a fixed frequency includes at least the total number of tasks to be assigned, the estimated execution time of each task, the task priority, the starting shelf coordinates of the task, and the latest start time of the task. The robot cluster raw data covers the real-time position of each robot, the list of assigned but incomplete tasks, historical task execution time records, remaining battery power, motion status, and a power consumption coefficient obtained by fitting long-term operating data. This step utilizes a multi-dimensional state feature aggregation mechanism to organize these raw data into a unified multi-dimensional state space, specifically composed of the task queue state vector. Robot cluster state matrix Composition, in which This represents the total number of tasks currently awaiting assignment. For the most recent Net increase in tasks within a time step. The net incremental acceleration characterizes the intensity at which the mission reaches its burst. Reflecting the heterogeneity of task time, The average priority of tasks; for robots Its state vector includes location and the number of assigned but unfinished tasks. Dynamic average processing rate Remaining battery power Movement state Expected battery life and maximum number of concurrent tasks . and After being synchronized and aggregated in time, a multi-dimensional state space is formed. This space fully describes the pressure characteristics at the task arrival end and the real-time capability boundary of the robot end, providing a structured global supply and demand situation for congestion warning in subsequent steps.

[0033] By aggregating multidimensional state features, the previously scattered task queue statistics and individual robot states are integrated into a unified expression, enabling the system to simultaneously observe changes in task arrival acceleration and the dynamic evolution of the robot's processing capacity within a single framework. This introduces the net incremental acceleration of the task. Subsequently, the state space can distinguish whether task injection is in a steady state or an acceleration phase, while the dynamic average processing rate is calculated based on a sliding window. and expected battery life This transforms the robot's capabilities from static preset values ​​into dynamic quantities that update online based on the current execution status. The direct effect of this is that subsequent congestion warning modules no longer need to compare scattered thresholds, but can directly utilize the already defined thresholds that include growth trends and control domain constraints. and Reasoning is then performed. Because the characteristics of task burst intensity and robot available capacity saturation are already embedded in the state space, the predictive model can identify the trend of widening supply-demand gap earlier, creating a computable basis for generating reliable pre-decomposition decisions in advance, and reducing misjudgments caused by ignoring task acceleration characteristics or instantaneous changes in robot capabilities.

[0034] In multi-robot scheduling scenarios, traditional state representation methods often only record the current task queue length and the busy / idle status of each robot, or simply count the number of idle robots. This coarse-grained representation cannot reflect the changing trend of task arrival rate, nor can it capture the critical situation where a robot leaves the available set due to decreasing remaining power or reaching its maximum concurrent capacity. When there is a sudden surge in orders, the task queue length may change drastically within seconds, but simply looking at the current length cannot determine whether this is a short-term fluctuation or a continuous impact. The system is prone to prematurely triggering unnecessary decomposition before actual congestion, or responding only when it is already overloaded. This step, by incorporating the net incremental acceleration of tasks and the dynamic processing rate and endurance of robots into the state space, completes the time-varying details, making the state information itself contain early signals of future congestion risks. Without adding additional sensors or communication mechanisms, the scheduling system obtains sufficient state richness to support proactive decision-making, overcoming the problem of excessive decision-making time lag under the original state representation.

[0035] For example, taking the morning peak of a certain e-commerce warehouse sorting system as an example, the dispatch center uses... Data is collected in seconds; the total number of tasks waiting to be assigned in the current task queue is [number missing]. One, the previous second was If there are [number], then the net increase is [number]. for The net incremental acceleration was found to be [number] units per second after multiple consecutive calculations. A positive value indicates that task injection is accelerating. Some tasks in the task pool involve long-distance shelves, causing an increase in the variance of the estimated execution time. Larger. The robot swarm contains... Remaining battery power of the robot It has dropped to Below is its expected battery life. Less than the time window required to complete the urgent task, there is another The number of unfinished tasks has been assigned to the robots in Taiwan. Approaching their respective maximum concurrent task capacity By aggregating the above information to and The multidimensional state space clearly presents the accelerating accumulation trend of tasks and the gap in available capabilities of the robot, providing an accurate data profile for congestion warning in the next step, and can start the pre-decomposition logic before the task pool is completely congested.

[0036] Step S20: Based on the multidimensional state space, the congestion early warning mechanism using Holt exponential smoothing prediction is used to perform a pre-decomposition decision task and output pre-decomposition decision information; Step S20 performs a pre-decomposition decision task based on a multi-dimensional state space. The output pre-decomposition decision information is a signal triggered when it is determined that the system will experience congestion in the future and tasks are still accelerating upon arrival. This signal drives subsequent adaptive task decomposition. This step first considers the task arrival rate... and acceleration Apply respectively Two-parameter exponential smoothing method to extract smoothed arrival rate and smooth acceleration Smoothing coefficient and This is used to balance response speed and noise suppression. Next, the system's dynamic processing capacity index is calculated. , where the indicator function Excluding those with remaining power below the warning threshold Or the number of assigned tasks has reached The robot ensures Reflects the actual available global processing speed. (In the prediction window) Inside, assuming The expected task queue length remains unchanged. Therefore, the expected congestion intensity index is defined. .when and And the above conditions are maintained continuously. During each scheduling cycle, pre-decomposed decision information is generated.

[0037] By By combining exponential smoothing prediction with dynamic capacity calculation, this step incorporates the second-order trend of task arrival at the destination and the available capacity of the robot swarm into a unified congestion judgment criterion. The smoothing process suppresses false warnings caused by instantaneous spikes at individual data collection moments, thus... and It can stably reflect the overall situation of task input, while its dynamic digestion capacity... Robots that are already saturated or have low battery are automatically excluded, ensuring that only individuals capable of continuously contributing processing power are included when calculating future queue lengths. Calculations are based on a position-accelerated formula. Subsequently, the system can predict future congestion intensity in advance, rather than waiting for the queue length to exceed a static threshold. This is achieved by introducing a continuous hold period. Short-term disturbances are further filtered out, so that the pre-decomposition decision information is only triggered when it is under continuous impact pressure, thus providing a reliable start signal for subsequent task decomposition.

[0038] In existing technologies, triggering task congestion often employs a single queue length threshold method, where splitting or reallocation is initiated when the number of tasks to be assigned exceeds a certain fixed value. While this approach can limit the rate of task arrival in scenarios with a stable arrival rate, it becomes problematic in scenarios with sudden peaks, such as e-commerce promotions, where the arrival rate can surge rapidly within a short period. The threshold triggering mechanism only begins to respond after the queue length has significantly exceeded the static value. At this point, a large number of tasks are already waiting, and some robots cannot be assigned due to insufficient power or full concurrent capacity, resulting in instantaneous congestion. Meanwhile, other robots with spare capacity are not assigned tasks because they were not pre-split into appropriately granular subtasks, leading to severe load imbalance. This step utilizes acceleration... and It enables advance decision-making based on future predictions, shifting the trigger point from "currently congested" to "about to be congested," and reflects the actual available processing resources through dynamic digestion capacity calculation, so that the system can issue an early warning when the order wave has just accelerated and the task pool has not yet been completely congested.

[0039] For example, during the opening phase of an e-commerce promotion, the task arrival rate was... From within seconds The rate gradually climbed to The acceleration changes from zero to a positive value and continues to increase at a rate of [number] units per second. The scheduling system [does this]. Smoothing yields smoothed arrival rates and smooth acceleration Simultaneously calculate dynamic digestion capacity Approximately Quantities / second. Within the prediction window. Expected task queue length in seconds Will increase to One, and the total maximum number of clusters for One, expected congestion intensity index Approximately Exceeding the preset threshold And the acceleration is continuous The scheduling cycle remains positive. The system determines that an overload is imminent, and immediately generates pre-decomposition decision information and transmits it to the task decomposition module. At this point, the actual queue length is still insufficient. This allows ample decision-making time for subtask splitting and robot allocation adjustments.

[0040] Step S30: Based on the pre-decomposition decision information, an adaptive decomposition task is executed using a decomposition mechanism of spectral clustering and dynamic capability weights, and a set of subtask packages is output. Step S30 is triggered upon receiving the pre-decomposition decision information. Its input is the current task pool to be decomposed, and its output is the set of subtask packages obtained through adaptive decomposition. This step first treats each task in the task pool as a graph node, constructing a task association graph where the weights of the edges between nodes are... ,in For the task and Manhattan distance from the starting shelf coordinates It is the difference between the latest start times of the two. and Number the aisle of the shelf where the task is located. , , These are adjustable weight coefficients. The Shi-Malik spectral clustering algorithm is then used to divide the task graph into... Task clusters are defined to ensure that the starting points of tasks within a cluster are spatially clustered and temporally compatible. Simultaneously, for the robot... To calculate its dynamic capability weights, first define the load margin factor. Efficiency factor and battery life factor Overall ability value After normalization, the capability weights are obtained. Determine the total pre-decomposition task volume. And calculate the task quota that each robot should be allocated. Then Task clusters can be combined or broken down into Each granularity equals the quota. The subtask packages allow the starting points of tasks within each package to cluster together.

[0041] A decomposition method combining spectral clustering and dynamic capability weights is employed to ensure that task splitting is not merely based on equal distribution of tasks, but also incorporates the dual constraints of spatial correlation between tasks and the robot's real-time carrying capacity. Spectral clustering achieves graph segmentation through eigenvalue decomposition of the graph Laplacian matrix, aggregating tasks with similar starting shelf locations and compatible time windows together. This reduces the empty travel distance of robots executing sub-task packages and the time loss caused by frequent task switching. The introduction of dynamic capability weights maps the robot's load margin, efficiency performance, and endurance online to the assigned task quota. This ensures that robots with strong capabilities receive a task load no less than their quota, while robots nearing saturation or with low battery levels receive a correspondingly reduced quota. The resulting sub-task packages are spatially compact and match the scale of the robot group's capability distribution, allowing the subsequent allocation module to complete matching in a shorter time and facilitating workload balancing among robots.

[0042] Traditional task decomposition often uses methods such as prioritizing tasks and then dividing them into equal blocks or distributing them equally based on the number of robots, rarely considering the spatial distribution of task starting points and the dynamic differences in robot capabilities. When task starting points are scattered across different shelving areas, the equally distributed task packages may require a single robot to travel back and forth between the east and west ends of the warehouse, significantly increasing idle time and prolonging the overall task completion time. Furthermore, when there are significant differences in processing speed, remaining battery power, and remaining capacity among robots, equal distribution can easily lead to some robots being overloaded while others are idle, resulting in uneven load distribution. This step uses spectral clustering to ensure spatial compactness within sub-task packages and uses capability weights to modulate the size of each package. Without increasing centralized computational complexity or requiring additional hardware, it alleviates the inefficiencies caused by spatial clutter and capability mismatch, thus generating sub-task packages suitable for rapid allocation even under sudden peak task pressure.

[0043] For example, the warehouse has just received a batch of orders, which involve The shelves in lane 1 and lane 3 of the district total One task, The shelves in lanes 8 and 10 of the district total Each task has a relatively flexible time window. The scheduling system determines the total amount of tasks to be pre-decomposed. for Spectral clustering generates two larger task clusters and three smaller clusters based on spatial distance and alleyway differences. Meanwhile, the three high-battery robots exhibit high efficiency and sufficient load margin, resulting in high overall capability values. Their capability weights are as follows: , and The other two units have a capability weight of only The system splits the two large clusters into sub-task packages that match the high-capacity quotas, and directly assigns the three smaller clusters as sub-task packages to low-capacity robots. This decomposition ensures that the starting points of the tasks within each robot's sub-task package are concentrated in a few adjacent aisles, significantly reducing the round-trip mileage during pickup. Furthermore, the load is matched to the remaining battery power, preventing interruptions due to insufficient power or overload during execution.

[0044] Step S40: Based on the set of sub-task packages, use A The path planning and allocation mechanism performs task allocation and outputs the allocation plan. Step S40 takes the set of subtask packages as input and outputs a specific allocation scheme for each robot. This step first defines the robot. Execute subtask package Cost function ,in To pass Heuristic algorithms plan the robot Current location to subtask package The optimal path travel time to the center of the starting region; The subtask packages are evaluated based on the maintained global spatiotemporal occupancy grid. Assigned to robots The additional waiting or detour time that may be introduced when its path conflicts with the assigned robot path Indicates the current allocation status; Assigning robots based on assumptions Loading factor, , , These are weighting coefficients. All robots calculate the marginal cost of inserting the unassigned subtask packet into their own path sequence. And use this as the standard value The bids are submitted to the central processing unit. The central processing unit collects all bids to form an initial allocation. If multiple robots submit bids for the same sub-task package that are similar in value and the highest, or if the allocation might lead to a foreseeable deadlock, a consensus processing unit is activated. Each robot determines whether to concede based on a collaborative concession function. The concession benefit is defined as the difference between the cost reduction from abandoning the package and the cost increase from winning the second-best package. A concession signal is only issued when the concession benefit is positive. After the processing unit finishes, it verifies for each robot whether the total number of tasks allocated exceeds its maximum concurrent task capacity. If the limit is exceeded, the excess sub-task package with the highest marginal cost will be forcibly abandoned and re-bid. The final allocation scheme will be generated after iteration until full compliance is achieved.

[0045] pass The system combines path planning with collaborative processing. The allocation process considers the shortest travel time while incorporating predictions of future path conflicts and a mechanism for inter-robot processing. Marginal cost calculation requires each robot to evaluate not only the cost of executing the sub-task package itself but also the incremental impact on the original path after inserting it into the existing task sequence. This effectively avoids global path congestion caused by locally optimal allocation. The consistency processing unit step in collaborative processing allows robots to dynamically adjust their allocation by quantifying concession benefits. A robot only issues a concession signal when it finds that abandoning the current competing package and switching to a suboptimal package can reduce its overall execution cost, making the allocation result closer to the globally optimal solution. The final overload check ensures that no robot will exceed its maximum concurrent task capacity, preventing execution delays or task failures due to overload. This ensures that the generated allocation scheme has high executability at both the path and capacity levels.

[0046] Traditional multi-robot task allocation often employs single-round processing or real-time greedy algorithms, rarely considering path conflicts and load limits. In warehouse environments with dozens of robots operating intensively, single-round processing easily leads to the convergence of planned paths at narrow passages or intersections, causing frequent avoidance and waiting, resulting in actual execution times far exceeding estimates. Greedy algorithms, on the other hand, tend to push a large number of tasks to a few more efficient robots, causing them to quickly reach or even exceed their concurrent processing capacity limits, thus reducing the overall task flow speed. This step addresses this by explicitly adding a conflict expectation penalty term to the cost function. and Load Factor Item Furthermore, an iterative processing unit and an over-limit backoff mechanism are introduced during the allocation process, making the allocation results smoother in terms of both path space-time occupancy and robot carrying capacity. This reduces the disorderly stagnation of robots caused by congestion and improves the timeliness and stability of the solution in actual execution.

[0047] For example, suppose there are currently three sub-task packages. , , Awaiting assignment, robot and All of them participated in the auction. pass The algorithm calculates the path from the current position to... The optimal path in the starting area needs to be traveled Seconds, but the spatiotemporal occupancy raster shows that the path will be in seconds. to During the second period with the assigned robot The planned trajectory intersects, and it is expected to introduce... The marginal cost of waiting for a conflict in seconds, plus the load factor, is: Second; Go to Driving required Seconds, conflict waiting Seconds, marginal cost is the same Seconds. The central processing unit initiates consistency processing. If the calculation is abandoned Instead, they won the second-best package. The total cost of itself is reduced. If the yield is positive within seconds, a yield signal is issued, and ultimately... get , get This avoided potential deadlocks caused by similar prices. Subsequent verification revealed... and The total number of tasks did not exceed their respective The allocation scheme was determined, and there were few path conflicts in actual execution, with the overall task completion time approaching the planned value.

[0048] Step S50: Based on the allocation scheme, a pullback and replanning mechanism with prediction deviation monitoring is used to perform a closed-loop correction task and output the corrected allocation scheme.

[0049] Step S50 takes the allocation scheme and the real-time feedback system status as inputs and outputs the corrected allocation scheme. This step calculates the prediction deviation in real time. ,in It is predicted in step S20. Task queue length at any given time This is the actual length of the task queue collected by the scheduling center. When... Exceeding the preset deviation threshold, or the remaining battery power of any robot. If the robot's performance drops by more than a preset percentage within a short period, or if robot malfunction reports are received for multiple consecutive scheduling cycles, a correction process is triggered. Upon triggering, all subtasks currently being executed by all robots are marked as frozen, but this does not affect their continued execution. Meanwhile, all subtasks assigned but not yet started in each robot's task list, along with the remaining tasks in the task pool, are all moved back to the waiting-to-be-assigned pool. Then with For the new task pool, using the current state of each robot and the frozen tasks as hard constraints, steps S20 to S40 are re-executed to generate a revised allocation scheme. This closed-loop mechanism enables the system to autonomously recover from prediction deviations and execution anomalies.

[0050] By monitoring prediction deviations and implementing rollback replanning, the system can automatically identify anomalies and initiate local rescheduling when the initial allocation plan deviates from reality due to environmental changes or prediction errors. Marking already started subtasks as frozen ensures the continuity of on-site handling operations, preventing interruptions to ongoing physical operations due to replanning and thus preventing secondary disturbances. Rolling back subtasks that have not yet started releases the planned resources they occupy, allowing them to enter a new round of decomposition and allocation along with other remaining tasks. This phased freezing and rollback strategy limits the correction scope to tasks that have not yet been actually executed, quickly correcting scheduling deviations caused by robot battery depletion, occasional malfunctions, or task prediction errors without interfering with ongoing handling operations, thus maintaining the time determinism of the overall task flow.

[0051] In traditional scheduling strategies, once a task allocation plan is issued, modifications are often only possible during the next complete scheduling cycle or through manual intervention. If a robot experiences an unexpected power outage or task prediction becomes significantly inaccurate during this period, assigned but not yet executed tasks cannot be promptly revoked and rerouted, causing some tasks to remain stalled for extended periods, while other idle robots are unable to intervene. This step addresses this by continuously monitoring prediction deviations online. By triggering closed-loop replanning in response to sudden changes in robot state and reverting unstarted tasks, the scheduling system gains the ability to autonomously repair unexpected situations in dynamic environments. This compensates for the vulnerability of open-loop assignment to uncertainty and reduces the risk of a large number of misassignments due to the accumulation of prediction model errors over time.

[0052] For example, after a batch of tasks is assigned, the robot... During the execution of the first sub-task, a motor malfunction caused accelerated power consumption, and the remaining power was... within seconds by Descending to Exceeding the preset The percentage decrease threshold; meanwhile, the actual task queue experienced a prediction deviation due to the upstream order replenishment growth rate exceeding the predicted value. The preset threshold has been exceeded. The closed-loop correction module is immediately triggered, and... The ongoing picking task is marked as frozen and continues to execute. At the same time, the two sub-task packages that have not yet been started in the robot's task list, as well as the sub-task packages that have not yet been started in the other robot lists, are withdrawn and together with the newly arrived tasks in the current task pool, they form the waiting-to-be-assigned pool. After re-executing congestion warning, adaptive decomposition, and collaborative processing, for Only a small number of tasks nearing the finish line are retained; the remaining, more demanding tasks are transferred to high-battery robots, and all sub-task packages are reassigned. The revised approach avoids... The task was delayed due to power depletion, but the overall task completion time did not show a significant extension.

[0053] Example 2: Furthermore, the present invention provides a multi-agent task scheduling system based on task decomposition and cooperative strategy optimization, employing a multi-agent task scheduling method based on task decomposition and cooperative strategy optimization from the above embodiments, which can solve the technical problem of multi-agent task scheduling based on task decomposition and cooperative strategy optimization. The beneficial effects of the multi-agent task scheduling system based on task decomposition and cooperative strategy optimization provided by the present invention are the same as those of the multi-agent task scheduling method based on task decomposition and cooperative strategy optimization provided in the above embodiments, and other technical features of the multi-agent task scheduling system based on task decomposition and cooperative strategy optimization are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0054] Example 3: This invention provides a multi-agent task scheduling device based on task decomposition and cooperative strategy optimization. The device includes at least one processor and a memory communicatively connected to the processor. The memory stores instructions executable by the processor, which are then executed to enable the processor to perform the multi-agent task scheduling method based on task decomposition and cooperative strategy optimization described in Example 1. This multi-agent task scheduling device can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. This multi-agent task scheduling device is merely an example and should not limit the functionality or scope of this invention. A multi-agent task scheduling device based on task decomposition and cooperative strategy optimization may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) that can execute various appropriate actions and processes according to a program stored in read-only memory or a program loaded from a storage device into random access memory. The random access memory also stores various programs and data required for the operation of the multi-agent task scheduling device based on task decomposition and cooperative strategy optimization. The processing unit, read-only memory, and random access memory are interconnected via a bus. An I / O interface is also connected to the bus. Typically, the following systems can be connected to the I / O interface: input devices including touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices including liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices including magnetic tapes, hard disks, etc.; and communication devices. The communication device allows the multi-agent task scheduling device based on task decomposition and cooperative strategy optimization to communicate wirelessly or wiredly with other devices to exchange data. While a multi-agent task scheduling device based on task decomposition and cooperative strategy optimization with various systems has been described, it should be understood that implementation or possession of all described systems is not required. It can be implemented alternatively or with more or fewer systems.

[0055] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the multi-agent task scheduling method based on task decomposition and cooperative strategy optimization described above. The computer program product provided by this invention can solve the technical problem of multi-agent task scheduling based on task decomposition and cooperative strategy optimization. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the multi-agent task scheduling method based on task decomposition and cooperative strategy optimization provided in the above embodiments, and will not be repeated here.

[0056] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a read-only memory. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0057] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0058] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A multi-agent task scheduling method based on task decomposition and collaborative strategy optimization, characterized in that, The methods include: Step S10: Obtain the original data of the task queue and the original data of the robot cluster. Based on the original data of the task queue and the original data of the robot cluster, perform the state representation task using a multi-dimensional state feature aggregation mechanism and output a multi-dimensional state space. Step S20: Based on the multidimensional state space, the congestion early warning mechanism using Holt exponential smoothing prediction is used to perform a pre-decomposition decision task and output pre-decomposition decision information; Step S30: Based on the pre-decomposition decision information, an adaptive decomposition task is executed using a decomposition mechanism of spectral clustering and dynamic capability weights, and a set of subtask packages is output. Step S40: adopting A The path planning allocation mechanism performs task allocation and outputs an allocation scheme. Step S50: Based on the allocation scheme, a pullback and replanning mechanism with prediction deviation monitoring is used to perform a closed-loop correction task and output the corrected allocation scheme.

2. The multi-agent task scheduling method based on task decomposition and collaborative policy optimization according to claim 1, wherein, Step S10, which involves acquiring the original data of the task queue and the original data of the robot cluster, and performing a state representation task based on the original data of the task queue and the original data of the robot cluster using a multi-dimensional state feature aggregation mechanism to output a multi-dimensional state space, specifically includes: Step S101: The scheduling center collects global data at a fixed frequency, defines the task queue state vector wherein is the total number of tasks to be allocated at time is the net increment of tasks within the preset collection time step is the total number of tasks to be allocated at time , is the acceleration of the net increment of tasks representing the intensity of the task arrival burst ; is the variance of the estimated execution time of all tasks in the task pool is the average priority of the tasks​ Step S102: Define the robot swarm state matrix For any robot in the robot swarm Its state vector contains: current position x-coordinate , current position y-coordinate , number of assigned unfinished tasks , dynamic average processing rate , percentage of remaining power , motion state , historical data fitted power consumption coefficient , expected endurance working time and maximum concurrent task carrying number ; Step S103: Convert the task queue state vector With the robot cluster state matrix The aggregation forms the multidimensional state space.

3. The multi-agent task scheduling method based on task decomposition and cooperative strategy optimization as described in claim 2, characterized in that, Step S20, which involves performing a pre-decomposition decision-making task based on the congestion warning mechanism using Holt exponential smoothing prediction in the multi-dimensional state space and outputting pre-decomposition decision information, specifically includes: Step S201: Set the task arrival rate and the acceleration of net increase in tasks Holt's two-parameter exponential smoothing method was used to extract the smoothed arrival rate. and smooth acceleration ,in in, and For smoothing coefficients; express Smooth arrival rate corresponding to each time point; Step S202: Calculate the system's dynamic digestion capacity index in The total number of robots, For indicator functions, Set a preset warning threshold for running out of power; Step S203: In the preset prediction window Get the future from within Expected task queue length at time ; Step S204: Define the expected congestion intensity index When both conditions are met and Furthermore, when the above two conditions are continuously met for a preset number of consecutive periods, the pre-decomposition decision information is generated; wherein, This is a preset congestion threshold.

4. The multi-agent task scheduling method based on task decomposition and cooperative strategy optimization as described in claim 3, characterized in that, In step S202, the and These are used to exclude situations where the remaining battery power is below a preset battery depletion warning threshold. The number of robots and the number of assigned but unfinished tasks have reached the maximum concurrent task capacity. The robot.

5. A multi-agent task scheduling method based on task decomposition and cooperative strategy optimization as described in claim 2, characterized in that, Step S30, which involves performing an adaptive decomposition task based on the pre-decomposition decision information using a spectral clustering and dynamic capability weighting decomposition mechanism, and outputting a set of subtask packages, specifically includes: Step S301: Obtain the current task pool to be decomposed from the pre-decomposition decision information, regard each task in the current task pool to be decomposed as a node, construct a task association graph, and define the weight of the edges between nodes; in For the task and Manhattan distance from the starting shelf coordinates For the task and The difference between the latest start times, Number the aisle of the shelf where the task is located. Using the weight coefficients, the Shi-Malik spectral clustering algorithm is used to divide the task association graph into... A cluster of tasks; Step S302: For any robot in the robot swarm Define load margin factor , Efficiency factor Battery life factor ,in, To preset the minimum estimated time window for completing an emergency task, Indicates that the robot swarm is The maximum dynamic average processing rate at any given time is calculated, and the overall capability value is determined. , ,in Adjustable weights; Step S303: Based on comprehensive capability value The Task clusters can be combined or broken down into A set of subtask packages is formed by subtask packages with a granularity equal to the quota.

6. The multi-agent task scheduling method based on task decomposition and cooperative strategy optimization as described in claim 1, characterized in that, In step S40, A is used based on the set of sub-task packages. The path planning and allocation mechanism performs task allocation and outputs an allocation plan through the following steps: Step S401: Define the robot Execute subtask package The cost in To adopt A Heuristic algorithm planning for robots Current location to subtask package The optimal path travel time from the center of the starting region. To evaluate the subtask package based on the maintained global spatiotemporal occupancy grid map Assigned to robot The additional waiting or detour time introduced when its path conflicts with the assigned robot path. Indicates the current allocation status. Assigning robots based on assumptions Load factor, These are the weighting coefficients; Step S402: Based on cost Tasks are assigned, and each robot i is checked individually to see if the total number of tasks assigned to it exceeds its maximum concurrent task capacity. If the limit is exceeded, the excess sub-task packages will be forcibly abandoned, and the final allocation scheme will be output.

7. The multi-agent task scheduling method based on task decomposition and cooperative strategy optimization as described in claim 1, characterized in that, Step S50, which involves executing a closed-loop correction task based on the predicted deviation monitoring and replanning mechanism according to the allocation scheme, and outputting the corrected allocation scheme, specifically includes: Step S501: Calculate prediction deviation in real time ,in for Predicted task queue length at any given time. for The actual task queue length at any given time, when The deviation exceeds the preset threshold, or the remaining battery percentage of any robot. If the price drops by more than a preset percentage within a preset time, a correction will be triggered. Step S502: After triggering the correction, mark all subtasks currently being executed by all robots as frozen, and simultaneously revert all subtasks that have been assigned but not yet started in each robot's list, along with the remaining tasks in the task pool, to the waiting-to-be-assigned pool. ; Step S503: Use the pool to be allocated For the new task pool, and with the frozen subtasks as hard constraints, steps S20 to S40 are re-executed to finally generate the revised allocation scheme.

8. A multi-agent task scheduling system based on task decomposition and cooperative strategy optimization, applied to the multi-agent task scheduling method based on task decomposition and cooperative strategy optimization as described in any one of claims 1 to 7, characterized in that, The system includes: The state aggregation module is used to acquire the original data of the task queue and the original data of the robot cluster, and to perform the state representation task based on the original data of the task queue and the original data of the robot cluster using a multi-dimensional state feature aggregation mechanism, and output a multi-dimensional state space. The congestion warning module is used to perform a pre-decomposition decision-making task based on the congestion warning mechanism using Holt exponential smoothing prediction based on the multi-dimensional state space, and output pre-decomposition decision information. The task decomposition module is used to perform adaptive decomposition tasks based on the pre-decomposition decision information using a decomposition mechanism of spectral clustering and dynamic capability weights, and output a set of sub-task packages. The task allocation module is used to assign tasks based on the set of sub-task packages using A. The path planning and allocation mechanism performs task allocation and outputs the allocation plan. The closed-loop correction module is used to perform closed-loop correction tasks based on the allocation scheme using a rollback and replanning mechanism that monitors prediction deviations, and outputs the corrected allocation scheme.

9. A multi-agent task scheduling device based on task decomposition and collaborative strategy optimization, characterized in that, The multi-agent task scheduling device based on task decomposition and cooperative strategy optimization includes: a memory, a processor, and a multi-agent task scheduling program based on task decomposition and cooperative strategy optimization stored in the memory and executable on the processor. When the multi-agent task scheduling program based on task decomposition and cooperative strategy optimization is executed by the processor, it implements a multi-agent task scheduling method based on task decomposition and cooperative strategy optimization as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a multi-agent task scheduler based on task decomposition and cooperative strategy optimization. When the multi-agent task scheduler based on task decomposition and cooperative strategy optimization is executed by the processor, it implements a multi-agent task scheduling method based on task decomposition and cooperative strategy optimization as described in any one of claims 1 to 7.

Images