Multi-agent cooperative task execution method and system

Abstract

The application discloses a multi-agent cooperative task execution method and device and a storage medium, and relates to the technical field of multi-agent cooperative control. The method first constructs a cooperative task description carrying a global time reference, sets a unified time sequence judgment basis for a full cooperation group; drives each agent to update a local cooperation view based on real-time interaction execution state information, and relies on the global time reference to carry out consistency verification of cooperative progress, thereby constraining the execution rhythm of each agent; when task execution is abnormal, candidate adjustment strategies generated by each agent based on abnormal information and the current state of the agent are obtained, consistency negotiation convergence is performed on all candidate strategies, and a unified cooperative adjustment scheme is obtained to drive continuous task execution. The method can effectively suppress the accumulation of time sequence deviation, ensure the cooperative and orderly progress of multi-agent execution, quickly complete global strategy alignment in abnormal scenarios, and thus improve the execution efficiency and running stability of multi-agent cooperative tasks.

Description

Technical Field

[0001] This invention relates to the field of multi-agent cooperative control technology, and in particular to a method and system for multi-agent cooperative task execution. Background Technology

[0002] The execution of existing multi-agent collaborative tasks generally adopts a model of subtask splitting and allocation combined with periodic state synchronization. Each agent starts execution with its own local clock as the timing reference and synchronizes its progress according to a preset period. When anomalies such as path obstruction or execution timeout occur, each agent autonomously adjusts its execution plan based on its own local state.

[0003] Because each agent always makes decisions based on local information and lacks a global unified collaborative reference benchmark that runs through the entire task lifecycle, the behavioral decisions of each agent cannot form effective mutual constraints and collaborative calibration in the two key stages of execution progress determination and abnormal strategy adjustment. Local deviations and local adjustments gradually accumulate and amplify in the absence of a unified benchmark constraint, eventually evolving into global execution conflicts and collaborative failures. Summary of the Invention

[0004] This invention provides a method and system for multi-agent collaborative task execution to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a multi-agent cooperative task execution method, comprising: Construct a collaborative task description that carries a global time reference; During task execution, each intelligent agent updates its local collaborative view based on real-time interactive execution status information and uses a global time base to verify the consistency of collaborative progress. When an anomaly occurs, the agent obtains the candidate adjustment strategies generated by each agent based on the anomaly information and its current state, and conducts consensus negotiation and convergence on all candidate strategies to obtain a unified collaborative adjustment scheme for continued execution.

[0006] Preferably, the construction of the collaborative task description carrying a global time reference includes: Obtain the semantics of each action in the collaborative task and its temporal constraints; A global time base is generated based on the deviation between the local clocks of each agent in the collaborative group and the standard reference clock; Configure a global execution time window for each action semantic based on a global time base; The semantics of each action configured with a global execution time window and their temporal constraints are integrated to generate a collaborative task description.

[0007] Preferably, the execution status information includes: the degree of action completion of each agent in the process of executing the current action, as well as the position coordinates and the local time tag of the current moment bound to the degree of action completion.

[0008] Preferably, the consistency verification of collaborative progress using a global time base includes: The local time stamp in the execution status information reported by each agent is compared with the global time base, and the execution status information with the deviation value within the preset tolerance range is marked as a valid synchronization state. For execution status information marked as valid synchronization state, a preliminary completion degree verification is performed based on the temporal dependency relationship of each action semantic in the collaborative task description. The consistency verification state of the agent that passes the preliminary completion degree verification is recorded as qualified.

[0009] Preferably, the step of obtaining candidate adjustment strategies generated by each agent based on anomaly information and its current state includes: The abnormal information is identified by type to determine the abnormal triggering action and the scope of the abnormal impact. Based on the abnormal triggering action and the scope of the abnormal impact, and combined with the current action completion rate of the agent, the unexecuted action sequence in the agent's local collaborative view is replanned to obtain the alternative action subgraph. Resource consumption is evaluated for each action path in the candidate action subgraph to obtain the resource requirement vector for each action path; Based on the matching result between the resource demand vector and the agent's current remaining resource vector, candidate adjustment strategies are extracted from the alternative action subgraph.

[0010] Preferably, the step of consensus negotiation and convergence of all candidate strategies to obtain a unified collaborative adjustment scheme includes: Action semantic parsing is performed on the candidate adjustment strategies reported by each agent to extract the action node identifiers and temporal partial order relationships between action nodes in each candidate adjustment strategy, thereby obtaining a set of strategy features. Based on the set of policy features, the overlap of all candidate adjustment policies at the action node level is calculated to obtain the set of overlapping action nodes and the set of conflicting action nodes. Generate a shared action resource allocation table based on the set of overlapping action nodes; Based on the set of conflicting action nodes and the shared action resource allocation table, the candidate adjustment strategies are iteratively adjusted until all strategies reach a consensus on action node identification and temporal partial order relationship, thus obtaining a unified collaborative adjustment scheme.

[0011] Preferably, the step of consensus negotiation and convergence of all candidate adjustment strategies further includes: For conflicting action nodes that fail to reach a consensus during each iteration adjustment process, conflict arbitration parameters are generated based on the task priority and global time base in the collaborative task description; Based on the conflict arbitration parameters, assign a globally unique execution order identifier to the conflict action node; Based on the globally unique execution order identifier, update the temporal partial order relationship and generate mandatory order constraints between conflicting action nodes; Forced order constraints are added to the local action sequence of each candidate adjustment policy to perform policy alignment and obtain aligned candidate adjustment policies. The aligned candidate adjustment strategy is used as the basis for the next iteration adjustment until all strategies reach a consensus.

[0012] Preferably, the step of consensus negotiation and convergence of all candidate strategies to obtain a unified collaborative adjustment scheme for continued execution further includes: Based on the unified collaborative adjustment scheme, the semantics of unexecuted actions and the global execution time window in the collaborative task description are updated to obtain the updated collaborative adjustment task description. The semantics of each action in the updated collaborative adjustment task description are projected to the corresponding agent action queue according to the globally unique execution order identifier, generating the adjusted action sequence for each agent. Based on the adjusted action sequence, each agent is driven to continue performing collaborative tasks, and the action execution feedback of each agent is monitored in real time.

[0013] A computer device includes a memory, a processor, and a computer program stored in the memory, characterized in that, when the computer program is executed by the processor, it implements the aforementioned multi-agent cooperative task execution method.

[0014] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned multi-agent cooperative task execution method.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By using a global time base as a unified reference to conduct consistency verification of collaborative progress, a timing judgment standard that runs through the entire life cycle of the task can be established for the entire collaborative group. This allows for continuous dynamic calibration and constraint of the execution progress of each agent, effectively suppressing the accumulation and spread of local clock deviations, ensuring that the execution actions of each agent are connected in an orderly manner under a unified timing system, and improving the smoothness of collaborative task execution and overall operational efficiency.

[0016] 2. By generating candidate adjustment strategies through resource matching and screening, and combining the iterative negotiation mechanism of action node overlap identification and conflict arbitration, the resource adaptability of single agent adjustment schemes can be improved from the source, and invalid negotiation samples can be reduced. At the same time, relying on the structured arbitration negotiation logic, the global alignment and convergence of multi-agent adjustment schemes can be quickly completed, eliminating resource conflicts and timing contradictions between adjustment actions, thereby reducing the duration of abnormal interruptions and strengthening the anti-interference ability and execution efficiency of collaborative tasks. Attached Figure Description

[0017] Figure 1 This is a flowchart of the multi-agent collaborative task execution method of the present invention; Figure 2 This is a flowchart of the candidate adjustment strategy acquisition method of the present invention; Figure 3 The flowchart below shows the method for generating a collaborative adjustment scheme according to the present invention. Figure 4 A structural diagram of a computer device provided by the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1, referring to Figures 1-3 As shown in the embodiment of the present invention, the multi-agent cooperative task execution method includes: S1. Construct a collaborative task description carrying a global time base, including: Obtain the semantics of each action in the collaborative task and its temporal constraints; A global time base is generated based on the deviation between the local clocks of each agent in the collaborative group and the standard reference clock; Configure a global execution time window for each action semantic based on a global time base; The semantics of each action configured with a global execution time window and their temporal constraints are integrated to generate a collaborative task description.

[0020] As a preferred embodiment of the present invention, when constructing a collaborative task description carrying a global time reference, the entire process of the collaborative task is first decomposed layer by layer to extract the semantics of each action and its temporal constraints. The semantics of an action refers to the smallest operational unit in the task that has independent execution logic, and the temporal constraints refer to the pre-dependencies and parallel execution rules between different semantics of actions.

[0021] Specifically, the real-time readings of the local clocks of each agent in the collaborative group are collected first. The difference between each local clock reading and the synchronization reading of the standard reference clock is calculated one by one to obtain the clock deviation amount corresponding to each agent. Then, the average value of all clock deviations is statistically analyzed, and a global time reference is generated based on the statistically obtained deviation correction value.

[0022] It should be noted that the sampling interval of the local clock ranges from 10 milliseconds to 100 milliseconds. It can be determined through multiple rounds of actual testing based on the time accuracy requirements of the collaborative task. Usually, 50 milliseconds is selected as the standard value. This value can ensure the accuracy of clock deviation measurement while controlling the frequency of communication interaction.

[0023] Then, using the global time base as a unified time coordinate axis, and combining the estimated execution time of each action semantic with the temporal constraint relationship, a corresponding global execution time window is configured for each action semantic.

[0024] It should be emphasized that the deviation tolerance range of the global execution time window is 1 to 3 times the sampling interval. It can be determined by simulation verification according to the fault tolerance level of the task. Usually, 2 times the sampling interval is selected as the conventional value. This value can maintain the constraint rigidity of the task timing and reserve a reasonable buffer space for the execution fluctuation of the agent.

[0025] Finally, the semantics of each action configured with a global execution time window and their temporal constraints are structurally integrated to generate a complete collaborative task description.

[0026] For example, in a multi-agent collaborative warehousing and sorting scenario, the semantics of four types of actions—product identification, product grabbing, path transfer, and product shelving—and their corresponding sequential dependencies are first decomposed. Then, a global time reference is generated based on the deviation between the local clocks of the four sorting agents and the standard reference clock. Subsequently, corresponding global execution time windows are configured for each of the four types of actions. Finally, these are integrated to form a collaborative task description for the sorting task. This can unify the time reference system of all agents at the task definition level and reduce the risk of deviation in subsequent collaborative progress verification from the source.

[0027] S2. Drive each agent to update the local collaborative view based on the execution status information of real-time interaction during task execution, and use the global time base to verify the consistency of collaborative progress.

[0028] The execution status information includes: the degree of action completion of each agent in the process of executing the current action, as well as the position coordinates and the local time label of the current moment bound to the degree of action completion.

[0029] This includes using a global time base to verify the consistency of collaborative progress, including: The local time stamp in the execution status information reported by each agent is compared with the global time base, and the execution status information with the deviation value within the preset tolerance range is marked as a valid synchronization state. For execution status information marked as valid synchronization state, a preliminary completion degree verification is performed based on the temporal dependency relationship of each action semantic in the collaborative task description. The consistency verification state of the agent that passes the preliminary completion degree verification is recorded as qualified.

[0030] In a preferred embodiment of the present invention, when driving each intelligent agent to start the collaborative task execution process according to the collaborative task description, the execution progress data of its current action is collected at a fixed period to generate execution status information. The execution status information includes the action completion degree of each intelligent agent in the process of executing the current action, as well as the position coordinates and the local time tag of the current moment bound to the action completion degree. The action completion degree refers to the proportion of the current action execution process to the complete action process. The position coordinates and the local time tag are bound one-to-one with the action completion degree of the corresponding sampling time.

[0031] Then, the generated execution status information is broadcast and interacted with to the other agents in the collaboration group. Based on the execution status information received from the other agents, the agent synchronously updates its own local collaboration view. The local collaboration view refers to the summary view of the task execution progress of the entire collaboration group stored locally by each agent.

[0032] It should be noted that the reporting period for execution status information ranges from 20 milliseconds to 200 milliseconds. It can be determined through multiple sets of comparative tests based on the real-time requirements of collaborative tasks. Usually, 100 milliseconds is selected as the standard value. This value can reduce the bandwidth load of communication within the group while taking into account the timeliness of status synchronization.

[0033] After completing the real-time update of the local collaborative view, the consistency of the collaborative progress is checked using the global time base. Specifically, the local time tag in the execution status information reported by each agent is extracted first, and the difference between the local time tag and the standard time corresponding to the global time base is calculated to obtain the time deviation value corresponding to each execution status information. Then, the time deviation value is compared with the preset tolerance range.

[0034] If the deviation value is within the preset tolerance range, the corresponding execution status information will be marked as a valid synchronization state. If the deviation value exceeds the preset tolerance range, the corresponding execution status information will be marked as a state to be calibrated and the clock fine-tuning operation of the corresponding agent will be triggered.

[0035] It should be emphasized that the preset tolerance range is between 5% and 15% of the width of the time window corresponding to the global time base. It can be determined through simulation verification according to the timing strictness of the task. Usually, 10% of the time window width is selected as the normal value. This value can filter out invalid verification fluctuations caused by small clock jitters and can also identify abnormal states with excessive timing deviations in a timely manner.

[0036] After completing the labeling and filtering of valid synchronization states, the execution state information marked as valid synchronization states is used to verify the prior completion degree based on the temporal dependency relationship of each action semantic in the collaborative task description. First, all prior action semantics corresponding to the current action are sorted out according to the temporal dependency relationship, and then the action completion degree of the agents responsible for each prior action semantic in the collaborative group is checked.

[0037] If the completion rate of all preceding action semantics reaches the preset completion threshold, the current agent's preceding dependency conditions are satisfied, and its consistency check status is recorded as qualified. If any preceding action semantics have not reached the preset completion threshold, the current agent's progress is determined to be ahead, and its consistency check status is recorded as waiting.

[0038] It should be noted that the preset completion threshold ranges from 95% to 100%, and can be determined through actual testing and calibration based on the fault tolerance requirements of the action execution. Usually, 100% is selected as the normal value. This value can strictly guarantee the execution order of the time-dependent relationship and avoid collaboration conflicts caused by the incomplete completion of the preceding action.

[0039] For example, in a multi-agent collaborative warehousing and sorting scenario, four sorting agents broadcast their own product grabbing completion rate, current shelf location coordinates, and local time stamp every 100 milliseconds. After each agent updates its local storage of the entire group's progress view, it first compares the local time stamp of each agent with the corresponding time of the global time base.

[0040] Then, states with deviations within 10% of the corresponding time window are marked as valid synchronization states. Next, it is verified whether the agent responsible for product identification has completed all the pre-identification actions. Only when the identification action completion rate reaches 100% is the consistency verification status of the agent responsible for product grabbing recorded as qualified. This ensures that the execution progress of each agent remains coordinated and consistent under a unified time benchmark, avoiding collaboration misalignment problems from both time and timing dimensions.

[0041] S3. When an anomaly occurs, obtain the candidate adjustment strategies generated by each agent based on the anomaly information and its own current state, and conduct consensus negotiation and convergence on all candidate strategies to obtain a unified collaborative adjustment scheme for continued execution.

[0042] This includes obtaining candidate adjustment strategies generated by each agent based on anomaly information and its current state, including: The abnormal information is identified by type to determine the abnormal triggering action and the scope of the abnormal impact. Based on the abnormal triggering action and the scope of the abnormal impact, and combined with the current action completion rate of the agent, the unexecuted action sequence in the agent's local collaborative view is replanned to obtain the alternative action subgraph. Resource consumption is evaluated for each action path in the candidate action subgraph to obtain the resource requirement vector for each action path; Based on the matching result between the resource demand vector and the agent's current remaining resource vector, candidate adjustment strategies are extracted from the alternative action subgraph.

[0043] This involves consensus negotiation and convergence across all candidate strategies to obtain a unified collaborative adjustment scheme, including: Action semantic parsing is performed on the candidate adjustment strategies reported by each agent to extract the action node identifiers and temporal partial order relationships between action nodes in each candidate adjustment strategy, thereby obtaining a set of strategy features. Based on the set of policy features, the overlap of all candidate adjustment policies at the action node level is calculated to obtain the set of overlapping action nodes and the set of conflicting action nodes. Generate a shared action resource allocation table based on the set of overlapping action nodes; Based on the set of conflicting action nodes and the shared action resource allocation table, the candidate adjustment strategies are iteratively adjusted until all strategies reach a consensus on action node identification and temporal partial order relationship, thus obtaining a unified collaborative adjustment scheme.

[0044] This includes achieving consensus and convergence for all candidate adjustment strategies, which also includes: For conflicting action nodes that fail to reach a consensus during each iteration adjustment process, conflict arbitration parameters are generated based on the task priority and global time base in the collaborative task description; Based on the conflict arbitration parameters, assign a globally unique execution order identifier to the conflict action node; Based on the globally unique execution order identifier, update the temporal partial order relationship and generate mandatory order constraints between conflicting action nodes; Forced order constraints are added to the local action sequence of each candidate adjustment policy to perform policy alignment and obtain aligned candidate adjustment policies. The aligned candidate adjustment strategy is used as the basis for the next iteration adjustment until all strategies reach a consensus.

[0045] This includes consensus negotiation and convergence of all candidate strategies to obtain a unified collaborative adjustment scheme for continued execution, and also includes: Based on the unified collaborative adjustment scheme, the semantics of unexecuted actions and the global execution time window in the collaborative task description are updated to obtain the updated collaborative adjustment task description. The semantics of each action in the updated collaborative adjustment task description are projected to the corresponding agent action queue according to the globally unique execution order identifier, generating the adjusted action sequence for each agent. Based on the adjusted action sequence, each agent is driven to continue performing collaborative tasks, and the action execution feedback of each agent is monitored in real time.

[0046] In a preferred embodiment of the present invention, when abnormal situations such as execution timeout or action execution failure occur during the execution of a collaborative task, candidate adjustment strategies generated by each agent based on the abnormal information and its current state are first obtained.

[0047] Specifically, the collected abnormal information is first identified by type. By matching the preset abnormal features, the execution links and impact boundaries corresponding to the abnormal are sorted out to obtain the abnormal triggering action and the scope of abnormal impact. The abnormal triggering action refers to the specific action semantics that directly cause the abnormal state, and the scope of abnormal impact refers to the subsequent action semantics and the scope of associated intelligent agents affected by the abnormality.

[0048] Then, based on the abnormal triggering action and the scope of the abnormal impact, combined with the current action completion rate of the agent, the unexecuted action sequence in the agent's local collaborative view is replanned, the action execution order and execution path are adjusted, and a candidate action subgraph is obtained. The candidate action subgraph refers to the set of all feasible unexecuted action paths that satisfy the timing constraints. Then, the resource consumption of each action path in the candidate action subgraph is evaluated, and the resource consumption such as computing power, mobile energy consumption, and job space required for each path is calculated to obtain the resource demand vector of each action path.

[0049] Next, the resource requirement vector of each action path is matched with the agent's current remaining resource vector. If the matching degree meets the preset matching threshold, the corresponding action path is extracted from the candidate action subgraph as a candidate adjustment strategy.

[0050] It should be noted that the preset matching threshold ranges from 70% to 95%, and can be determined through multiple sets of simulation tests based on the resource redundancy of the agent. Usually, 85% is selected as the conventional value. This value can both screen out adjustment paths with good resource adaptability and retain a sufficient number of alternative solutions to support subsequent negotiations.

[0051] After collecting candidate adjustment strategies for each agent, a consensus negotiation is then conducted on all candidate strategies to achieve convergence and obtain a unified collaborative adjustment scheme.

[0052] Specifically, the action semantics of the candidate adjustment strategies reported by each agent are first analyzed, and the execution steps and sequential logic of each strategy are sorted out one by one. The action node identifiers and temporal partial order relationships between action nodes in each candidate adjustment strategy are extracted to obtain the strategy feature set.

[0053] The action node identifier refers to the unique identity marker of each action semantic, the temporal partial order relationship refers to the sequential execution constraint relationship between action nodes, and then, based on the policy feature set, the overlap of all candidate adjustment policies at the action node level is calculated, and the action nodes contained in all policies and the action nodes with conflicting execution logic are counted to obtain the overlapping action node set and the conflicting action node set.

[0054] Then, based on the set of overlapping action nodes, the shared job resources and time periods corresponding to each node are sorted out to generate a shared action resource allocation table. Then, based on the set of conflicting action nodes and the shared action resource allocation table, each candidate adjustment strategy is iteratively adjusted, and the execution logic and timing relationship of conflicting nodes are corrected one by one until all strategies reach a consensus on action node identification and timing partial order relationship, and a unified collaborative adjustment scheme is obtained.

[0055] It should be emphasized that the threshold for calculating overlap ranges from 80% to 100%, and can be determined by actual measurement and calibration based on the degree of collaboration in the task. Usually, 90% is selected as the standard value. This value can accurately distinguish between consensus actions and divergent actions, reduce ineffective negotiation links, and improve convergence efficiency.

[0056] During the iterative adjustment process, for conflicting action nodes that have not reached a consensus in a single iteration, an arbitration mechanism is used to promote negotiation convergence. First, the task priority and global time benchmark in the collaborative task description are combined to generate conflict arbitration parameters by comprehensively considering the importance level of the task and the time urgency. Then, based on the sorting results of the conflict arbitration parameters, a globally unique execution order identifier is assigned to the conflicting action nodes to clarify the sequential execution authority of each conflicting node.

[0057] Then, based on the globally unique execution order identifier, the temporal partial order relationship between the original action nodes is updated, a mandatory order constraint between conflicting action nodes is generated, and the mandatory order constraint is added to the local action sequence of each candidate adjustment strategy. The execution logic of conflicting nodes of all strategies is unified, and the aligned candidate adjustment strategies are obtained after strategy alignment.

[0058] Finally, the aligned candidate adjustment strategy is used as the basis for the next iteration adjustment, and the conflict identification and arbitration process is repeated until all strategies reach a consensus. It should be noted that the maximum number of iterations for the adjustment ranges from 3 to 10, which can be determined through simulation verification based on the number of conflicting nodes. Usually, 5 is selected as a common value. This value can ensure sufficient negotiation in complex conflict scenarios, and also avoid infinite iteration that slows down the task execution pace, thus balancing the sufficiency of negotiation and the execution efficiency.

[0059] After the negotiation converges and a unified collaborative adjustment plan is obtained, the task configuration is updated synchronously and each agent is driven to resume execution.

[0060] Specifically, based on a unified collaborative adjustment scheme, the semantics of unexecuted actions and the global execution time window in the collaborative task description are synchronously corrected, the execution order of actions and the corresponding time intervals are adjusted, and the updated collaborative adjustment task description is obtained. Then, the semantics of each action in the updated collaborative adjustment task description are projected to the corresponding agent action queue according to the globally unique execution order identifier, replacing the original sequence of unexecuted actions, and generating the adjusted action sequence for each agent.

[0061] Then, based on the adjusted action sequence, each agent is driven to restart the task execution process, and the action execution feedback of each agent is monitored in real time according to the original status reporting cycle to continuously verify the consistency of collaborative progress. For example, in a multi-agent collaborative warehouse sorting scenario, when a sorting agent encounters a channel blockage anomaly, the anomaly triggering action is first identified as a path transfer action. The scope of the anomaly covers the subsequent goods shelving action and the two adjacent transfer agents. Each agent, based on its remaining power and operating space, replans to obtain alternative transfer paths, and after resource matching, generates its own candidate adjustment strategy.

[0062] Then, through multiple rounds of negotiation, overlapping shelf resources and conflicting passage paths are sorted out. After priority arbitration, a global passage order is assigned to the conflicting paths. Finally, a unified detour adjustment plan is generated and the action queue of each agent is updated. This drives all agents to continue to execute the sorting task according to the adjusted path. A globally consistent adjustment plan can be quickly reached after an anomaly occurs, avoiding secondary conflicts caused by each agent's adjustment and ensuring the continuous and stable execution of the task.

[0063] Example 2, as Figure 4 As shown, a computer device includes one or more processors 100 and a memory 200 communicatively connected to one or more processors 100, wherein the processors 100 are responsible for executing program instructions to complete task scheduling and computational processing operations, and the memory 200 is responsible for storing various configuration and status data of computer programs and collaborative tasks.

[0064] Specifically, the processor 100 and the memory 200 establish a bidirectional communication connection through a data transmission bus. First, the memory 200 transmits the computer program instructions to be run and the basic configuration data of the cooperative task to the processor 100. Then, the processor 100 sends the cooperative task description, consistency verification results and cooperative adjustment scheme generated by the operation back to the memory 200 for persistent storage.

[0065] It should be noted that the number of processors 100 ranges from 1 to 8, and can be determined through actual deployment testing based on the computational complexity of the collaborative task and the total scale of the connected intelligent agents. Usually, 2 processors are selected as the standard configuration. This configuration can meet the computational requirements of multi-threaded parallel processing while controlling the overall hardware cost of the device, and effectively improve the execution efficiency of task scheduling and policy negotiation.

[0066] It is important to emphasize that when the processor 100 calls and runs the computer program stored in the memory 200, it fully executes the entire process of building a collaborative task description with a global time reference, driving each agent to update its local collaborative view and complete the consistency verification of collaborative progress, and collecting candidate strategies and converging consensus negotiations in abnormal scenarios. All execution logic strictly follows the timing rules and constraints of the aforementioned multi-agent collaborative task execution method, ensuring that the output collaborative scheduling instructions have a globally unified time reference and execution logic.

[0067] For example, in a multi-agent collaborative warehousing and sorting scenario, the device uses dual processors 100 to run program instructions in memory 200 in parallel, generating sorting task descriptions with global time references for four sorting agents, receiving execution status information of each agent in real time and completing consistency verification, and quickly completing negotiation and convergence of candidate strategies and issuing a unified adjustment plan when channel blockage occurs. This can effectively improve the execution efficiency and anomaly recovery speed of collaborative warehousing and sorting tasks.

[0068] Example 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements all the steps of the aforementioned multi-agent cooperative task execution method.

[0069] Specifically, a computer-readable storage medium carries and stores complete program instruction code in a non-volatile manner. When an external processor reads the computer program in the storage medium and starts running it, the entire execution process of the multi-agent collaborative task, from task construction and progress verification to anomaly adjustment, can be completely reproduced, achieving a collaborative execution effect that is completely consistent with the aforementioned method.

[0070] It should be noted that the range of carrier types for computer-readable storage media includes three categories: magnetic storage media, optical storage media, and semiconductor storage media. The selection can be determined through verification based on the deployment environment, storage capacity requirements, and read / write performance requirements of the application scenario. Semiconductor flash memory media is usually selected as the conventional choice. This type of media has the advantages of small size, strong shock resistance, and fast read / write response speed, and can adapt to the usage needs of various embedded devices and fixed deployment scenarios.

[0071] It is important to emphasize that the computer program stored in the computer-readable storage medium contains complete method execution logic, covering the entire process of task construction, progress verification and exception negotiation. It can run independently to implement all method steps without the need for additional program components, effectively ensuring the convenience of method application and cross-device compatibility.

[0072] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for multi-agent cooperative task execution, characterized in that, The method includes: Construct a collaborative task description that carries a global time reference; During task execution, each intelligent agent updates its local collaborative view based on real-time interactive execution status information and uses a global time base to verify the consistency of collaborative progress. When an anomaly occurs, the agent obtains the candidate adjustment strategies generated by each agent based on the anomaly information and its current state, and conducts consensus negotiation and convergence on all candidate strategies to obtain a unified collaborative adjustment scheme for continued execution.

2. The multi-agent cooperative task execution method as described in claim 1, characterized in that, The description of constructing a collaborative task carrying a global time reference includes: Obtain the semantics of each action in the collaborative task and its temporal constraints; A global time base is generated based on the deviation between the local clocks of each agent in the collaborative group and the standard reference clock; Configure a global execution time window for each action semantic based on a global time base; The semantics of each action configured with a global execution time window and their temporal constraints are integrated to generate a collaborative task description.

3. The multi-agent cooperative task execution method as described in claim 2, characterized in that, The execution status information includes: the degree of action completion of each agent in the process of executing the current action, as well as the position coordinates and the local time tag of the current moment bound to the degree of action completion.

4. The multi-agent cooperative task execution method as described in claim 3, characterized in that, The consistency verification of collaborative progress using a global time base includes: The local time stamp in the execution status information reported by each agent is compared with the global time base, and the execution status information with the deviation value within the preset tolerance range is marked as a valid synchronization state. For execution status information marked as valid synchronization state, a preliminary completion degree verification is performed based on the temporal dependency relationship of each action semantic in the collaborative task description. The consistency verification state of the agent that passes the preliminary completion degree verification is recorded as qualified.

5. The multi-agent cooperative task execution method as described in claim 4, characterized in that, The acquisition of candidate adjustment strategies generated by each agent based on anomaly information and its current state includes: The abnormal information is identified by type to determine the abnormal triggering action and the scope of the abnormal impact. Based on the abnormal triggering action and the scope of the abnormal impact, and combined with the current action completion rate of the agent, the unexecuted action sequence in the agent's local collaborative view is replanned to obtain the alternative action subgraph. Resource consumption is evaluated for each action path in the candidate action subgraph to obtain the resource requirement vector for each action path; Based on the matching result between the resource demand vector and the agent's current remaining resource vector, candidate adjustment strategies are extracted from the alternative action subgraph.

6. The multi-agent cooperative task execution method as described in claim 1, characterized in that, The process of consensus negotiation and convergence of all candidate strategies to obtain a unified collaborative adjustment scheme includes: Action semantic parsing is performed on the candidate adjustment strategies reported by each agent to extract the action node identifiers and temporal partial order relationships between action nodes in each candidate adjustment strategy, thereby obtaining a set of strategy features. Based on the set of policy features, the overlap of all candidate adjustment policies at the action node level is calculated to obtain the set of overlapping action nodes and the set of conflicting action nodes. Generate a shared action resource allocation table based on the set of overlapping action nodes; Based on the set of conflicting action nodes and the shared action resource allocation table, the candidate adjustment strategies are iteratively adjusted until all strategies reach a consensus on action node identification and temporal partial order relationship, thus obtaining a unified collaborative adjustment scheme.

7. The multi-agent cooperative task execution method as described in claim 6, characterized in that, The consensus negotiation and convergence of all candidate adjustment strategies also includes: For conflicting action nodes that fail to reach a consensus during each iteration adjustment process, conflict arbitration parameters are generated based on the task priority and global time base in the collaborative task description; Based on the conflict arbitration parameters, assign a globally unique execution order identifier to the conflict action node; Based on the globally unique execution order identifier, update the temporal partial order relationship and generate mandatory order constraints between conflicting action nodes; Forced order constraints are added to the local action sequence of each candidate adjustment policy to perform policy alignment and obtain aligned candidate adjustment policies. The aligned candidate adjustment strategy is used as the basis for the next iteration adjustment until all strategies reach a consensus.

8. The multi-agent cooperative task execution method as described in claim 7, characterized in that, The step of consensus negotiation and convergence of all candidate strategies to obtain a unified collaborative adjustment scheme for continued execution also includes: Based on the unified collaborative adjustment scheme, the semantics of unexecuted actions and the global execution time window in the collaborative task description are updated to obtain the updated collaborative adjustment task description. The semantics of each action in the updated collaborative adjustment task description are projected to the corresponding agent action queue according to the globally unique execution order identifier, generating the adjusted action sequence for each agent. Based on the adjusted action sequence, each agent is driven to continue performing collaborative tasks, and the action execution feedback of each agent is monitored in real time.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the computer program is executed by a processor, it implements the multi-agent cooperative task execution method as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When a computer program is executed by a processor, it implements the multi-agent cooperative task execution method as described in any one of claims 1 to 8.

Images