Multi-agent cooperative task processing method and device, equipment and medium

Abstract

The application relates to the technical field of artificial intelligence, can be applied to business scenes such as pension services, financial technology and medical health, and discloses a multi-agent cooperative task processing method, device, equipment and medium, which comprises the following steps: acquiring a task instruction and analyzing a core target, and decomposing the core target into multiple subtasks; acquiring environment information, dividing a task area, combining an agent capability with an area weight to generate a cooperation framework; acquiring an agent real-time state, matching an optimal agent based on the cooperation framework to generate a task allocation table; issuing the task allocation table to control the agent to execute a task and upload execution information; monitoring an execution process, dynamically adjusting and updating the task allocation table when a path conflict or a device fault is detected; and acquiring environment completion state data after the subtasks are completed and comparing the environment completion state data with a preset standard model to check and accept. The application realizes efficient task decomposition and intelligent allocation by fusing task semantics, environment information and agent capability, and introduces real-time state sensing and an adaptive mechanism to improve cooperation stability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a multi-agent collaborative task processing method and device, equipment and a storage medium. BACKGROUND

[0002] With the continuous development of multi-agent systems in service robots, intelligent manufacturing, space collaboration and other fields, multi-agent collaborative mechanisms for complex tasks have gradually become the focus of research and application. However, the existing multi-agent task processing methods still have several key problems in task analysis, task area division, capability perception matching, task allocation and collaborative execution process, which seriously restrict the overall execution efficiency and collaborative stability of the system.

[0003] In the elderly care service scene, rehabilitation robots are often used to assist in completing tasks such as home organization and area cleaning. However, in the existing system, tasks are mostly manually set or predefined template triggered, lacking the ability to analyze the semantics of natural language task instructions, making it difficult to flexibly identify specific targets according to the actual oral commands issued by the elderly. In addition, task division usually does not consider the room structure and area characteristics, resulting in task scheduling being not targeted. There is a lack of efficient capability alignment mechanism among multiple robots, and the optimal matching cannot be made in task allocation combined with the current capability state of each robot, causing some agents to be overloaded, while others are in idle state, resulting in low overall efficiency.

[0004] In the medical and health business field, intelligent devices are gradually applied to tasks such as ward patrol, environmental maintenance and auxiliary logistics. The current technology generally has the problem of rough modeling of medical environment information, making it difficult to effectively divide the task area according to the spatial distribution. Due to the inability to accurately match device capabilities with area needs, high-capability devices are often assigned low-complexity tasks, and low-capability devices have difficulty completing high-precision operations. At the same time, real-time state changes of devices are not timely incorporated into the task scheduling mechanism, resulting in task allocation lagging and collaboration being disjointed, affecting the continuity and responsiveness of task execution.

[0005] In the field of financial technology business, intelligent service robots or multi-agent platforms are applied to intelligent site tour, customer reception and background data process business. However, in the face of tasks with time sensitivity and process dependence, the existing system is difficult to build a stable and efficient task collaboration framework. The system often cannot effectively utilize work environment data and regional resource characteristics for task decomposition, lacks a mechanism for dynamically matching task requirements and device capabilities, resulting in a lack of scientificity in task allocation, low efficiency in resource allocation, and thus affecting service quality and user experience. SUMMARY

[0006] The main purpose of the present application is to provide a multi-agent collaborative task processing method, device, equipment and storage medium, aiming at solving the technical problems that the existing multi-agent task processing cannot be jointly modeled and dynamically allocated based on task semantics, environment structure and agent capability state, resulting in low cooperation efficiency, poor resource utilization, and lack of global optimization of task collaboration mechanism.

[0007] To achieve the above purpose, the present application provides a multi-agent collaborative task processing method, comprising:

[0008] Obtaining a task instruction, performing semantic analysis on the task instruction to identify a core task target, and decomposing a task corresponding to the task instruction into a plurality of subtasks based on the core task target;

[0009] Obtaining working environment information, dividing a plurality of task areas according to the working environment information and determining the area weight of each task area, combining the capability parameters of a plurality of agents and the area weight, and generating a global collaboration framework;

[0010] Obtaining real-time state information of a plurality of agents, combining the global collaboration framework and the real-time state information, matching the optimal agent for each subtask, and generating a task allocation table;

[0011] Issuing the task allocation table to the plurality of agents, controlling the agents to execute the subtasks, and uploading the task execution information collected in the execution process to a shared information platform;

[0012] Monitoring the execution process of the agents through the shared information platform, triggering a dynamic adjustment strategy for conflict resolution or task re-allocation when detecting agent path conflict or device failure, generating an updated task allocation table and re-issuing;

[0013] After all the subtasks are executed, obtaining environment completion state data, comparing the environment completion state data with a preset standard model for acceptance.

[0014] Further, to achieve the above purpose, the present application provides a multi-agent collaborative task processing device, comprising:

[0015] A task analysis module for obtaining a task instruction, performing semantic analysis on the task instruction to identify a core task target, and decomposing a task corresponding to the task instruction into a plurality of subtasks based on the core task target;

[0016] A task planning module for obtaining working environment information, dividing a plurality of task areas according to the working environment information and determining the area weight of each task area, combining the capability parameters of a plurality of agents and the area weight, and generating a global collaboration framework;

[0017] The task allocation module is configured to acquire real-time state information of the plurality of agents, match an optimal agent for each subtask in combination with the global cooperation framework and the real-time state information, and generate a task allocation table;

[0018] The task execution module is configured to issue the task allocation table to the plurality of agents, control the agents to execute the subtasks, and upload task execution information collected in the execution process to the shared information platform;

[0019] The dynamic scheduling module is configured to monitor the execution process of the agents through the shared information platform, trigger a dynamic adjustment strategy to resolve conflicts or re-allocate tasks when detecting path conflicts or device failures of the agents, generate an updated task allocation table, and re-issue the updated task allocation table.

[0020] The task acceptance module is configured to acquire environment completion state data after all the subtasks are executed, and compare the environment completion state data with a preset standard model for acceptance.

[0021] Further, to achieve the above object, the present application further provides a computer device, which comprises a memory, a processor, and a multi-agent cooperative task processing program stored in the memory and executable on the processor, and the multi-agent cooperative task processing program implements the steps of the multi-agent cooperative task processing method when executed by the processor.

[0022] Further, to achieve the above object, the present application further provides a computer readable storage medium, which stores a multi-agent cooperative task processing program, and the multi-agent cooperative task processing program implements the steps of the multi-agent cooperative task processing method when executed by a processor.

[0023] Beneficial effects: The application relates to the technical field of artificial intelligence, can be applied to business scenarios such as pension services, financial technology and medical health, and discloses a multi-agent collaborative task processing method, device, equipment and medium, which comprises the following steps: acquiring a task instruction and performing semantic analysis to identify a core task target, decomposing the task into multiple subtasks; acquiring working environment information and dividing a task area, combining an agent capability parameter and a region weight to generate a global collaboration framework; acquiring real-time state information of the agent, performing optimal agent matching based on the global collaboration framework to generate a task allocation table; issuing the task allocation table to control the agent to execute the task and upload task execution information; monitoring the execution process based on a shared information platform, triggering a dynamic adjustment strategy when a path conflict or a device fault is detected, generating and issuing an updated task allocation table; and acquiring environment completion state data after the subtasks are executed and comparing the environment completion state data with a preset standard model for acceptance. The application realizes efficient conversion of task instructions to multi-agent collaborative execution by combining task semantic information, environment structure information and agent capability parameters for unified modeling, and introducing real-time state sensing and dynamic allocation mechanisms. Through real-time identification and adaptive adjustment of path conflicts and device faults, the stability and reliability of system collaboration are improved. BRIEF DESCRIPTION OF DRAWINGS

[0024] The application will be further described below in combination with the drawings and embodiments, and the drawings are as follows:

[0025] Figure 1 An application environment schematic diagram of the multi-agent collaborative task processing method in an embodiment of the application;

[0026] Figure 2 A flowchart of the multi-agent collaborative task processing method in an embodiment of the application;

[0027] Figure 3 A function module schematic diagram of a preferred embodiment of the multi-agent collaborative task processing device of the application;

[0028] Figure 4 A structure schematic diagram of a computer device in an embodiment of the application;

[0029] Figure 5 Another structure schematic diagram of a computer device in an embodiment of the application. DETAILED DESCRIPTION

[0030] It should be understood that the specific embodiments described herein are only used to explain the application, and are not used to limit the application.

[0031] The multi-agent collaborative task processing method provided by the embodiments of the application can be applied to business scenarios such as Figure 1In an application environment of the present application, the user end communicates with the service end through the network. The service end can obtain task instructions through the user end and perform semantic analysis to identify the core task target, decompose the task into multiple subtasks, obtain working environment information and divide the task area, generate a global collaboration framework in combination with the agent capability parameters and the area weight, obtain real-time state information of the agent, perform optimal agent matching based on the global collaboration framework to generate a task allocation table, issue the task allocation table to control the agent to execute the task and upload the task execution information, monitor the execution process based on the shared information platform, trigger a dynamic adjustment strategy when detecting path conflicts or device failures, generate and issue an updated task allocation table, and obtain the environment completion state data after the subtasks are executed and compare it with the preset standard model for acceptance. The present application realizes efficient conversion of task instructions to multi-agent collaborative execution by combining task semantic information, environment structure information and agent capability parameters for unified modeling, and introducing real-time state perception and dynamic allocation mechanism. Through real-time identification and adaptive adjustment of path conflicts and device failures, the stability and reliability of system collaboration are improved. The user end can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers and portable wearable devices. The service end can be implemented by an independent server or a server cluster composed of multiple servers. The present application will be described in detail below through specific embodiments.

[0032] Please refer to Figure 2 , Figure 2 The flowchart of an embodiment of the multi-agent collaborative task processing method provided by the present application is shown. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0033] As shown in Figure 2 The multi-agent collaborative task processing method provided by the present application includes the following steps:

[0034] S10, obtaining a task instruction, performing semantic analysis on the task instruction to identify a core task target, and decomposing a task corresponding to the task instruction into multiple subtasks based on the core task target;

[0035] In this embodiment, the operation of obtaining task instructions generally faces the input channel of human-computer interaction for task acquisition, and the input forms include but are not limited to voice, text, image and other multi-modal input data. The source of the task instruction can be a manual service request, a scheduling information of an autonomous scheduling platform, a task content returned by an external system interface, etc. For example, in a home service scenario, a user can issue a "please clean the bedroom" instruction through a voice assistant, or send a list type text task through a mobile terminal. In a financial or medical scenario, the task information can also come from an event request pushed by a business scheduling platform according to rules. After receiving the task instruction, the system needs to standardize the translation, and convert the original task data into a structured format for subsequent processing modules to call. Structured processing includes operations such as unified semantic identification, keyword extraction, specification of task entity boundaries, and elimination of ambiguous items. The translation process can use a word segmentation engine, a named entity recognition module, and a rule engine to assist in completing syntax analysis. Finally, the task content is represented as multiple fields in a unified structure, such as task verbs, target entities, execution objects, and environmental constraints.

[0036] After completing the instruction format conversion, enter the semantic analysis stage, the goal of which is to extract the behavior intention and operation target implied in the task. Semantic analysis can use pre-trained language models for context understanding, and combine domain-specific task ontology libraries for matching reasoning. After embedding the instruction text through vector space semantic representation, the language model identifies the core action type and target object through a classifier or matching rules, for example, identifying "clean up the living room" as the action "clean up" and the target area "living room". The semantic analysis process can also include context association judgment, such as whether the location of "desk" mentioned in the previous instruction affects the target of "cleaning" this time, or whether there are time, sequence, constraints and other complex semantic relationships. The semantic tags extracted through these analysis methods can include but are not limited to task target, operation limitation condition, priority constraint, task trigger condition, etc.

[0037] After identifying the core task target, the task target needs to be task decomposed to form multiple sub-tasks. The decomposition process is based on the task type structure defined in the knowledge base for hierarchical decomposition, and combines the environment structure and the existing task atlas to automatically match the sub-task type. For example, "cleaning the living room" can be decomposed into "cleaning the tea table", "putting the remote control in place", and "sweeping the floor". Task decomposition can also be performed according to the distribution of entities in the target area, such as decomposing "cleaning the kitchen" into "stove cleaning", "floor cleaning", and "kitchenware placement" according to the spatial layout information. During the decomposition process, the pre-set task operation tree, task dependency graph, and device accessibility model can be referred to for sub-task generation, and each sub-task will be closely bound to the subsequent capability matching and task scheduling process. Task decomposition is not limited to parallel sub-tasks, but can also handle sequential dependent tasks, condition triggered tasks, and mutually exclusive tasks, etc. Special task forms, by defining execution order, condition statements or execution constraints in the sub-task description, make the sub-task have complete schedulability and executability.

[0038] The task instruction acquisition can be realized through an input terminal equipped with a natural language interaction interface, such as an edge computing device embedded with a speech recognition chip, which can listen and receive human language instructions in real time and convert the speech signal into a text stream. The converted text is input to the natural language processing engine through the semantic recognition module, and the language model with encoding-decoding structure is used for semantic unit extraction, and the semantic slot filling is combined with the rule template to further extract structured targets. For example, after receiving "help me clean up the living room", the system automatically extracts "clean up" as the action, "living room" as the target entity, and "help me" as the service request type, and stores the structure information as a standard task format.

[0039] The semantic analysis of the task target can be completed by fusing the context-aware language model based on BERT and the semantic matching engine based on the knowledge graph. For complex synthetic instructions, first use dependency syntax analysis to preliminarily analyze the semantic relationship, and then use the domain semantic library to determine the actual meaning of each task unit and exclude ambiguity. For example, in "please clean the desk and the floor", the desk needs to clean the drawers and the table, and the floor needs to be cleaned, so the parser will split it into two different types of task targets.

[0040] In the task decomposition implementation process, a mapping atlas between task types and environmental objects can be constructed, and a sub-task set can be automatically generated according to the mapping relationship. The task decomposition algorithm can adopt a hierarchical task tree structure modeling, wherein the root node represents the overall task target, the child nodes are divided according to the operation dimension or target area, and the leaf nodes represent specific execution instructions. The system refines the task to the spatial entity granularity according to the identified target area characteristics, and generates a sub-task set that can be directly issued to a specific execution unit. In addition, each task after sub-task splitting is assigned a logical execution identifier, including task priority, parallel marker, preset execution window and other information, providing a basis for subsequent allocation and scheduling process.

[0041] Example: In the smart elderly care service scenario, when an elderly user issues a request of "please help me clean up the room" through a voice assistant, the system receives the task instruction and performs semantic analysis, identifying "clean up" as an operation action and "room" as a target area. The analysis system further confirms multiple task targets implied in the instruction "clean up the room" based on semantic relationships, such as "tidy the bed" "put things in place" "clean the floor" and the like. The task decomposition engine refines the task into multiple executable sub-tasks according to the room layout information, furniture and item distribution map and historical service data, such as "move floor debris to the storage cabinet" "clean the surface of the bedside table with a wet wipe" "adjust the bedding to be neatly placed" and the like. Each sub-task will carry specific spatial positioning information and operation action instructions, and the subsequent scheduling engine will assign them to service robots with corresponding capability parameters, such as robots with strong fine operation capability for item placement, robots with high cleaning efficiency for floor cleaning, and the like. Through this processing method, the system not only improves the rationality of task allocation and the accuracy of execution, but also significantly reduces the risk of repetitive work and missed areas, thereby providing elderly users with a safer, more comfortable and dignified living service experience.

[0042] In the medical and health business field, nursing instructions such as "please help the patient tidy up the bed and clean the ward" can identify "tidy up the bed" and "clean the ward" as two independent targets through semantic analysis, and further decompose them into sub-tasks such as "align the pillow" "smooth the bed sheet" "clean the floor" and the like. These sub-tasks can be accurately issued according to the structure of the ward space and the capability characteristics of intelligent nursing equipment, realizing efficient collaborative nursing work.

[0043] In the financial technology business field, a work order task such as "audit customer points this month and clean up abnormal records" in a customer service system can extract two task targets "audit points" and "clean up records" through task analysis. The system automatically splits the task into sub-tasks such as "query customer points data" "identify points anomalies" "perform points repair" according to the data table structure and processing logic, and then schedules corresponding data processing engines and log cleaning service modules to complete the task collaboratively.

[0044] The task recognition mechanism based on semantic analysis can accurately parse the task instruction into a structured task target, avoiding the one-sidedness and ambiguity of traditional rule matching in task understanding. Further decomposing the task parsing result into multiple sub-tasks with clear operation objects and execution constraints can help improve the accuracy and flexibility of task scheduling. The operation chain avoids overlapping responsibilities and unclear division of labor in the task execution process, improving the clarity, response speed and expandability of the overall system task processing.

[0045] S20, obtaining working environment information, dividing a plurality of task areas according to the working environment information and determining a region weight of each task area, combining the ability parameters of the plurality of agents and the region weight, and generating a global collaboration framework;

[0046] In the embodiment, the operation of obtaining working environment information usually relies on a device with spatial perception capability to comprehensively scan the service place. The environment information includes but is not limited to spatial layout structure, area size parameters, obstacle distribution, furniture position, use frequency identifier, environment cleanliness state and function purpose classification, etc. The above information can be collected by a laser radar, a depth camera, an inertial navigation device or an infrared sensor, and a structured spatial data model is generated. The model represents the geometric structure and function distribution of the target environment in a digital way, providing support for subsequent operations.

[0047] After establishing the environment space model, the working space needs to be logically divided. The division method can be based on the space purpose (such as living room, kitchen, bedroom, etc.), or can be combined with physical structure boundaries (such as doors, walls, furniture placement lines) for automatic area cutting. Each area will be an independent task processing unit, and its boundary, function label and processing difficulty will be recorded in the task area definition data set.

[0048] The determination of the region weight is based on the comprehensive calculation of multiple dimensions of parameters. The primary dimension is the area size, which determines the task intensity. The second is the function use frequency, which can be extracted from the historical behavior log, for example, the kitchen and bathroom have higher priority in cleaning tasks. In addition, it also includes the region accessibility coefficient, task complexity score, cleaning target density, etc. The above indexes can be calculated by a weighted function to obtain the normalized region weight, which determines the priority and resource allocation proportion of the region in collaborative scheduling.

[0049] The capability parameters of the plurality of agents are generally dependent on a unified capability parameter database, and the capability parameters include mechanical structure parameters (such as arm span, load, degree of freedom), environmental adaptability indexes (such as terrain passing rate), operation precision level, task completion success rate, energy consumption efficiency, operation time record and the like. These capability parameters can be obtained and updated based on manufacturing specifications, system initialization calibration, historical execution data and the like. All capability parameters will be encoded as structured vectors for processing by the agent task matching model.

[0050] The collaborative fusion of the environment division result and the agent capability parameter is the key to building a global collaboration framework. The system builds a multi-objective optimization matching matrix to establish a capability matching score relationship between all task areas and agents. The scoring mechanism can measure the similarity between the required capability vector of the region and the capability vector of the agent, and further consider the adjustment effect of the region weight on the task resource allocation. After completing all matching degree evaluation, a distribution structure covering all task areas can be generated through greedy algorithm, linear programming algorithm, graph matching algorithm and the like. This structure defines the task priority sequence, task area attribution and predicted execution time window of each agent, and finally forms a global collaboration framework for control execution.

[0051] The collection of environmental information can be achieved by a multi-modal sensor network fixedly installed on the ceiling or wall, or by a mobile robot carrying sensors for patrol inspection. When dividing the open space into regions, the convolutional neural network can be used for functional semantic segmentation of space images to realize automatic space function labeling. When dividing the complex structure environment, the BIM model or pre-stored CAD drawing can be used to assist space modeling. In the calculation of region weight, the coefficients of parameter weight can be set by experience, or can be optimized by data-driven model regression in historical task execution.

[0052] The management mode of capability parameters can adopt a centralized capability database maintenance model, in which the capability parameters of each agent will be dynamically updated after the task is completed. If it is necessary to adapt to the high dynamic change environment of the task, the task adaptability feedback index can also be temporarily called before the construction of the collaboration framework to realize real-time capability evaluation. The task matching score matrix can use cosine similarity as the basic score function, or can introduce a task adaptation loss function for reverse evaluation. The output structure of the collaboration framework supports embedding in the real-time scheduling module to realize continuous correction of the task allocation table.

[0053] Example: In the field of elderly care services, the system first conducts a comprehensive scan of the elderly's living space through sensors, identifying typical living areas such as bedrooms, living rooms, kitchens, etc., and analyzing the usage frequency and functional intensity of each area based on the elderly's daily activity data. The system sets the bedroom as a high-weight area, especially focusing on the task of tidying up before and after the elderly's afternoon nap. The agent capability parameter shows that A1 robot is good at tidying tasks, while A2 robot is more efficient in floor cleaning. Based on this matching logic, the global collaboration framework assigns the bedroom tidying task to A1 and the living room cleaning task to A2, effectively improving task completion quality and reducing the elderly's discomfort.

[0054] In the field of medical health services, the system is deployed in the logistics management scenario of medical institutions to collaboratively process medical environment tasks with multiple cleaning / moving / delivering agents. Environmental modeling identifies functional areas such as operating rooms, wards, corridors, etc., and the system assigns different weights based on area infection risk level, operation frequency, and spatial complexity, such as high-frequency disinfection in operating rooms and targeted medicine delivery per bed in wards. The capability parameter shows that B1 agent has high-precision positioning and stable navigation capabilities, suitable for completing equipment arrangement tasks in operating rooms; B2 is suitable for cleaning tasks in corridors. Through the collaboration framework modeled based on the combination of weights and capabilities, the repeated entry and exit paths in high-infection-risk areas can be reduced, and the overall operation efficiency and infection control capability can be improved.

[0055] In the field of financial technology services, the system is applied to environmental maintenance tasks in large data centers, such as monitoring rooms, machine rooms, power module key areas, etc. The system divides cleaning areas and sets weights based on device distribution density, ventilation demand, and ground cable complexity. C1 robot has anti-electromagnetic interference capability and complex terrain adaptation capability, and C2 robot has high-efficiency dust removal capability and obstacle avoidance capability. Based on the matching relationship between capabilities and task requirements, the collaboration framework assigns the machine room cable arrangement task to C1 and the air conditioner outlet dust removal task to C2. This allocation maximizes the efficiency of the agents and improves the stability and safety of high-sensitivity area operations without affecting the operation of the data center.

[0056] This embodiment can effectively improve the distinction and allocation accuracy of task units and enhance the adaptability of resource scheduling by structuring the working environment and introducing task area division and weight evaluation mechanisms. Based on the capability parameter, a task matching mechanism can be established to avoid mismatches between tasks and performers, improving agent utilization efficiency and task execution quality. The resulting collaboration framework has strong interpretability and adjustability, providing a data foundation for dynamic scheduling.

[0057] S30, obtain real-time state information of the plurality of agents, match an optimal agent for each sub-task based on the global collaboration framework and the real-time state information, and generate a task allocation table;

[0058] In this embodiment, obtaining real-time state information of multiple agents is a basic link for implementing a dynamic scheduling mechanism, and its essence is to capture the instantaneous capability boundaries and action constraints of each execution agent in the collaborative system. The state of each agent is not only whether it is currently idle, but also a real-time mapping of a multi-dimensional attribute set, including but not limited to position information, motion state, attitude parameter, current task context, system health index, energy remaining level, task execution accuracy, module load rate, perception sensor feedback stability, path accessibility and network communication state, etc. The collection of these state information relies on the local perception system and task control unit of the agent, and is synchronized to the scheduling center or edge control cluster by the embedded communication module with low delay link to form a real-time state vector matrix based on timestamp identification. This matrix needs to maintain consistency and availability within the scheduling period, so it needs to combine state sampling window mechanism and data integrity fault tolerance strategy, such as using Kalman filter to predict and complete intermittent state, or enabling redundant channels to improve communication robustness in high interference environment.

[0059] Real-time state information is not processed in isolation, but needs to be integrated into the global collaboration framework constructed in the previous stage. The global collaboration framework establishes an initial matching relationship between agents and sub-tasks based on task area characteristics, capability parameters and area weights, and its output is a static priority allocation structure, which only has strategic reference value in real dynamic environment. Therefore, in the scheduling execution process, the matching structure needs to be adapted and adjusted based on real-time state information.

[0060] In the matching process, first, each sub-task is bound to its corresponding task area, and the definition of the task area can come from spatial segmentation algorithm, physical boundary annotation or semantic label mapping. Then, the initial matching degree score between each sub-task and all candidate agents is extracted from the global collaboration framework as a static baseline. Next, the influence coefficient of the current state of each candidate agent on the completion efficiency and success rate of the sub-task is dynamically calculated, including but not limited to the following dimensions:

[0061] Idle state: whether it is currently not bound by other tasks, or the expected task will be completed in a short time;

[0062] Position information: path cost from the current position to the target task area (heuristic path search such as A* can be used to calculate the shortest time path);

[0063] Energy state: whether the remaining power is greater than the minimum execution requirement of the task (safety return power needs to be reserved);

[0064] Task congestion coefficient: whether there is a bound task causing queuing, delay or area conflict;

[0065] Resource availability: whether the current agent has the tool module or interface capability required to complete the task;

[0066] Communication reachability: whether there is a communication blind area in the task area, affecting the real-time issuance of task instructions or the reporting of state;

[0067] Execution history stability: success rate, completion time, failure reason, and other behavior modeling statistical data under similar tasks.

[0068] These factors can construct a multi-dimensional state vector, input a weighted scoring model or a trained scheduling scoring network model (such as gradient boosting tree, attention neural network, etc.) to output the current matching score. This score is the priority estimate of the agent completing the subtask at the current time.

[0069] To prevent overfitting or skewed allocation, the system can introduce random perturbation terms (such as softmax temperature control) to achieve scheduling diversity, or add scheduling fairness restrictions (such as task history acceptance frequency restrictions) to improve overall resource utilization efficiency.

[0070] When generating the task allocation table, the system selects the execution subject for each subtask from the set of idle agents with the highest matching score based on subtask priority ranking. Resource conflict detection is required during the matching and allocation process to avoid multiple tasks binding the same agent or a single agent taking on overloaded tasks. The task allocation table should be a structured data carrier containing fields such as task identification, task area, execution time limit, scheduling source, allocated agent identification, expected state change, task completion threshold, etc., for subsequent task issuance and state tracking.

[0071] Under different task densities, the allocation strategy can have different granularity and distribution characteristics. For example, in the case of low task density and abundant agent resources, a greedy scheduling algorithm can be used to quickly allocate resources; in the case of high task density and resource scarcity, approximate global optimization algorithms such as genetic algorithms, reinforcement learning strategy networks, and simulated annealing algorithms can be used to improve the quality of scheduling solutions.

[0072] The task matching model can introduce a weight adaptive mechanism according to different task types. For example, in the item carrying task, the load capacity and path stability of the agent are dominant; in the fine operation task, the mechanical arm precision and stability score should be the core scoring factors. The task matching model should support plug-in strategy switching to adapt to different business logic.

[0073] The periodic control of real-time state acquisition can be dynamically adjusted according to the urgency of the task. Important area tasks or critical path agents need higher frequency state reporting, while the edge area can appropriately reduce the sampling frequency to save communication bandwidth. The state abnormality detection mechanism should also run in real time. When a state breakpoint, parameter mutation or logic conflict (such as a task being executed but the idle flag being true) is found, the state verification mechanism and task reconstruction logic should be triggered to improve system robustness.

[0074] The task allocation table can be designed as a double-layer structure: the first layer is a task dimension structure, including sub-task ID and allocation target; the second layer is an agent dimension structure, which identifies the current binding state of all tasks, facilitating subsequent conflict detection and abnormal recovery.

[0075] Example: In the elderly care scenario, multiple service robots need to complete the partition cleaning task in the room. In a certain task, robot A1 was originally assigned to clean the old man's bedroom, but the system detected that its current position was far from the bedroom, and the power was only 15%, which was not enough to complete the complete task process. At the same time, robot A4 has just completed the bathroom cleaning and is currently in an idle state and only 2 meters away from the bedroom. The scheduling system decides to temporarily switch the bedroom cleaning task to A4 by calculating its path cost, power redundancy and historical cleaning accuracy score, ensuring uninterrupted service and optimal resource utilization.

[0076] In the medical health field, a group of cleaning robots need to complete the alternating disinfection task of the ward corridor and the isolation area. Due to the sudden epidemic alert, it is required to speed up the disinfection of high-risk infection areas. The system obtains the current task state of all agents and the regional path traffic situation, finds that the robot B2 originally responsible for disinfection is trapped in path congestion, while B5 has completed the task in the adjacent area and has the ability of high protection module, immediately schedules B5 to perform the disinfection task, shortens the response delay, and ensures the timely completion of the task.

[0077] In the financial data center environment, the patrol robot needs to complete the line abnormal identification task of multiple cabinet areas. The scheduling system finds in real-time state that C1 robot needs to be urgently withdrawn for maintenance due to temperature sensing abnormality, and the task cannot continue, while C3 has just completed the adjacent task and has stable image recognition module and low interference tolerance characteristics, so it immediately switches the C1 task to C3, ensuring that the high-priority patrol task is not interrupted, and avoiding passive blind area of important data link.

[0078] The embodiment introduces real-time state information of multiple agents into task scheduling logic, dynamically perceives the load, path, resource and energy efficiency boundary of the task execution system, so that the task allocation can accurately reflect the current optimal undertaker of each subtask, and improves the task completion efficiency, scheduling response speed and resource utilization. When the system faces dynamic changes in the state of the agent (such as power reduction, task congestion, temporary communication interruption), it has adaptive redistribution capability, effectively reducing the task failure rate and overall collaborative cost caused by incorrect matching. The mechanism of real-time state participating in task matching makes the scheduling decision evolve from static preset to highly flexible response, adapting to high-dynamic, multi-disturbance and multi-target collaborative task execution scenarios.

[0079] S40, issuing the task allocation table to the plurality of agents, controlling the agents to execute the subtasks, and uploading the task execution information collected in the execution process to a shared information platform;

[0080] In the embodiment, the task allocation table is issued to the plurality of agents, which needs to rely on the synchronous and cooperative mechanism between the task scheduling module and the communication module. The task allocation table is the carrier of structured scheduling output, which includes multiple fields, such as subtask identification, task belonging area, allocation object identification, execution time limit, task priority, task type label, parameter template, etc. The table is transmitted in a wired or wireless manner through a network scheduling bus, usually using MQTT, ROS communication protocol or custom low-latency data protocol for task pushing. Before issuing the task allocation table, the scheduling system needs to perform table structure legality verification and content consistency verification to ensure that the task target, target area and agent capability match consistently, and avoid execution interruption or conflict caused by illegal instructions.

[0081] After receiving the task allocation table, the agent will parse the task content through the local scheduling interface, extract the target task area, execution target and task behavior chain template, and generate the action path and task behavior sequence according to the preset control strategy. Under the scheduling of the control module, the agent executes the subtask process one by one according to the task behavior sequence. The task control process may involve path planning, module wakeup, action execution, state collection, abnormality perception and feedback response, etc. The behavior template can load different subprograms according to the task type, such as cleaning task scheduling chassis driving module and dust collection module, sorting task scheduling mechanical arm module and image recognition module.

[0082] During task execution, the agent needs to continuously collect task execution information, including current position changes, target object processing status, perception sensor data, execution phase completion flags, error status codes, module load conditions, etc. The original collection format of the execution information is unstructured data, which needs to be converted into a structured format by a local data processing unit, labeling information type, timestamp, task context, and execution phase label. Some tasks also need to be attached with real-time environmental data, such as temperature and humidity, electromagnetic interference strength, space occupancy rate, etc.

[0083] The structured task execution information needs to be uploaded to the shared information platform after data processing. The shared information platform is the public information exchange center of the multi-agent collaborative system, supporting unified storage, real-time update, and access authorization of task information, state information, and space information. The information upload adopts a multi-channel parallel mechanism to ensure high-frequency and low-latency communication requirements. The system optimizes bandwidth occupation through block uploading strategy or streaming compression protocol, and supports data reporting feedback and abnormal retransmission mechanism to prevent information loss or miscommunication. After receiving the task execution information, the shared information platform automatically maps it to the current task space state and updates the corresponding task progress, area completion degree, and abnormal identification field for subsequent monitoring and scheduling modules to call.

[0084] The format of the task allocation table can use different field templates according to different types of tasks. For example, for structured space cleaning tasks, the task table contains cleaning path point sets, action mode parameters, area risk level, and expected completion precision level; for item sorting tasks, it contains item identification labels, target placement location, operation grasping posture template, and maximum retry count, etc.

[0085] The generation of task behavior sequence can rely on rule template library or strategy network model. For example, in the elderly care scenario, the sorting task template contains a five-stage process of "identify target → stretch arm to grab → carry to placement location → take photo for verification → mark as completed". The task execution control flow is driven by a state machine mechanism, and each stage needs to pass the state feedback success flag before entering the next stage.

[0086] The uploading strategy of task execution information supports two modes of periodic uploading and event-triggered uploading. Periodic uploading is suitable for continuous tasks (such as patrol and cleaning), and event uploading is suitable for result-oriented tasks (such as switch control and grasping operation). For agents executing tasks in high-dynamic and complex environments, the uploading mechanism can also combine frame rate compression and abnormal detection mechanisms to trigger information reporting only at key frames or abnormal frames to reduce platform data load.

[0087] In the shared information platform, the execution information can be stored in hash groups according to task identification, supporting task backtracking and execution trajectory visualization. At the same time, the platform supports synchronous interface with the superior management platform, pushes the task execution status to the supervision end or user interface, and realizes transparent management of tasks.

[0088] Example: In the elderly service scene, home robot A2 is assigned to perform the bedroom tidying task. The system issues a task allocation table to it, including the target area "master bedroom", the task type "item return", the behavior template "stretching arm -> recognition -> carrying -> placing", and at the same time stipulates the task completion time limit and environmental interference level. A2 parses and starts executing the task, and after recognizing the scattered clothes, it performs the grabbing action and carries them to the wardrobe area, while recording the state changes and sensor feedback of each action stage. After the tidying is completed, the structured task execution information is uploaded to the shared information platform, and other robots can determine whether the bedroom tidying is completed, avoiding repeated scheduling.

[0089] In the medical and health field, after the disinfection robot receives the task allocation table, it performs the disinfection task according to the preset area path. The system continuously collects its travel trajectory, spray release state and air quality change value, and uploads them in real time to the shared information platform according to the area position label, for the background monitoring system to remotely monitor the task execution quality and environmental improvement.

[0090] In the data center in the financial field, the patrol robot receives the task allocation table and goes to the specified cabinet area to perform the line inspection task. During each inspection process, it records key execution data such as cable connection state image, power indicator state, environmental temperature and humidity information, and uploads them in real time to the shared information platform. The background system judges whether there is abnormal fluctuation or manual intervention is needed based on the task execution information, realizing remote and efficient management of the inspection process.

[0091] This embodiment can realize high-concurrency and precise multi-task scheduling through issuing task allocation tables and executing tasks by agents according to behavior sequences. The task execution information is uploaded in real time to the shared information platform, so that the scheduling system can fully grasp the execution status of each task, the current running condition of the agent and the environmental change data. This mechanism improves the perception closed-loop capability and execution stability of the multi-agent system, helps to discover task abnormalities in time and trigger correction measures, and improves the overall collaboration efficiency and execution reliability. The data is uniformly collected to the shared information platform, which provides complete data support for subsequent task optimization and capability learning, and enhances the adaptability and evolution ability of the system.

[0092] S50, monitoring the execution process of the agent through the shared information platform, and when detecting path conflict or device failure of the agent, triggering a dynamic adjustment strategy to resolve the conflict or reassign the task, generating an updated task allocation table and reissuing it;

[0093] In the present embodiment, the monitoring of the execution process of the plurality of agents relies on the continuous data access capability of the shared information platform and the unified state modeling mechanism. The shared information platform continuously receives and integrates the task execution information uploaded by each agent, and constructs a global visual running map combining the dynamic state and spatial position data of the environment. The map records the spatial position, motion trajectory, current task progress, module running state and resource consumption state of all agents, and other key running parameters.

[0094] Path conflict detection is based on the spatiotemporal intersection analysis of the spatial trajectories of multiple agents. In the shared information platform, the system performs overlap analysis by comparing the future predicted trajectories of multiple agents with the currently occupied paths. If any two trajectories intersect within the future window and the intersection point is less than a preset safety threshold, it is marked as a path conflict event. The predicted trajectory can be obtained by a trajectory reasoning model based on historical motion state or a future travel path provided by the local planning module of each agent. Once a path conflict event is identified, the system immediately freezes the motion state of the low-priority agent, prevents it from entering the conflict area, and starts the conflict resolution process.

[0095] Device failure detection is based on abnormal codes in task execution information, sensor self-check feedback, module running state and energy level indicators. If there is an abnormal action response, execution stagnation, energy level below the set threshold, etc. in the state information uploaded by a certain agent, it is determined as a device failure event. The shared information platform identifies the fault impact range and task interruption point in combination with the historical task execution model and the ability tag library.

[0096] Dynamic adjustment strategy is divided into two types of paths: conflict resolution path and task reallocation path. For path conflict events, the system first evaluates the current task urgency of each agent involved in the conflict. The urgency calculation combines task priority, remaining time and spatial travel cost, etc. Agents with low urgency are controlled to pause movement and trigger path re-planning process to generate an avoidance path to avoid the conflict area. The avoidance path is generated by the path planning engine in the global map state, optimizing the path cost and delay indicators, and the result is written into the movement instruction for issuance.

[0097] For device failure events, the system identifies the current interrupted task and extracts its remaining workload parameters, and simultaneously retrieves idle agents with high matching degree for the task in the ability database. The matching criteria include historical task completion ability, geographical location, energy level and module health status. The selected replacement agent receives the task takeover instruction and registers its new task state in the shared information platform.

[0098] The process of updating the task allocation table is based on completed conflict avoidance or task succession decisions. It replaces the corresponding fields in the original task allocation table with new paths, new agent IDs, new task deadlines, and adjustment tags, and then re-verifies the table's consistency. The generated new task allocation table is redistributed to the relevant agents through the scheduling module. Simultaneously, the shared information platform updates the task status mapping table and the regional task distribution map to ensure that all system components have a consistent understanding of the current task status.

[0099] Path conflict detection can employ a sliding prediction window method to simulate trajectories within a time segment, identifying conflict pairs with intersection probabilities exceeding a threshold. The trajectory prediction model can be a linear prediction model based on Bayesian filters, or it can incorporate graph neural networks to model potential coupling behaviors between agents, enabling dynamic prediction of complex trajectories.

[0100] The urgency of a task can be calculated using the following formula: U=α·P+β·(1-T / Tmax)+γ·(1-D / Dmax), where P is the task priority, T is the remaining time of the task, D is the current path distance, and α, β, and γ are weighting coefficients that can be automatically adjusted in actual deployment using machine learning algorithms.

[0101] The selection of a replacement agent for a faulty task can incorporate a capability similarity scoring function, comprehensively considering the matching degree between multiple capability dimensions (such as organization ability, cleaning efficiency, and navigation accuracy) and task requirement parameters to achieve precise scheduling. Simultaneously, before task replacement, the system must verify the legality of the replacement agent's execution path to avoid conflicts between new and existing paths.

[0102] The task allocation table is updated incrementally, replacing only the affected fields while maintaining the original allocation logic for the remaining tasks, minimizing system fluctuations. The scheduling module can use an asynchronous update strategy to handle multiple task conflicts in parallel, improving system response speed.

[0103] Example Description: In an elderly care service scenario, two service robots, A1 and A2, perform bedroom and hallway tidying tasks respectively. When A2 needs to cross the hallway to enter the master bedroom while A1 is exiting the bedroom, their predicted trajectories intersect. The system determines there is a path conflict and pauses the lower-priority A2, planning an avoidance path. During this process, the task allocation table is updated to add an avoidance path field for A2. If A1 experiences a battery alarm during execution, the system immediately calls upon the idle A3 as a replacement, retrieves A1's remaining task data, and issues a task takeover command to ensure the successful completion of the master bedroom tidying task.

[0104] In the field of medical health, when the cleaning robot works in the ward area, its path prediction conflicts with the supply robot path. The system avoids intersection collision through passage order judgment and path re-planning to ensure the smooth passage of medical equipment. If the cleaning module of the cleaning robot fails, the system automatically switches the task to the standby robot and issues an updated task allocation table to complete the task connection.

[0105] In the financial field, when the data center patrol robot exits the task due to sensor abnormalities during cabinet inspection, the system immediately allocates a standby robot to continue the task according to the task history and capability matching, and regenerates the task table and notifies the management platform to avoid missing critical task nodes and ensure the continuity of the operation and maintenance process.

[0106] This embodiment realizes real-time monitoring of the intelligent agent execution process through the shared information platform, which can identify path conflicts and device failures in real time during task execution, ensuring the stability and continuity of the multi-agent system operation process. Path conflict events can be effectively avoided through real-time trajectory prediction and dynamic path adjustment mechanism to improve the safety and space resource utilization efficiency of task execution. Device failure events can be timely replaced through task redistribution process to ensure uninterrupted tasks. The dynamic updating and synchronous issuing mechanism of the task allocation table ensures the consistency and integrity of the overall task coordination of the entire system when local abnormalities occur.

[0107] S60, after all sub-tasks are executed, the environment completion state data is obtained, and the environment completion state data is compared with a preset standard model for acceptance.

[0108] In this embodiment, after all sub-tasks are completed, the actual change amount of the environment state caused by task execution needs to be evaluated, and the task completion quality is judged based on the evaluation result. The environment completion state data represents a structured expression form of the current state of the target area after task execution, reflecting quantifiable information such as space layout, object position, cleanliness, and obstacle arrangement. This data is obtained by triggering the final state collection process after the intelligent agent completes the execution, usually combining visual, laser radar, depth camera or multi-modal sensor scanning of the task execution area, and filtering, fusing and reconstructing the scanning data to generate three-dimensional point cloud, grid map or RGB-D image expression forms.

[0109] In order to have comparability, the environment completion state data needs to be standardized through data structure and transformed into a unified format three-dimensional environment model through feature extraction algorithm. Feature extraction includes shape reconstruction based on geometry, object label recognition based on semantics, and space partition modeling based on topology. The three-dimensional environment model structure should have high resolution, spatial continuity and semantic integrity, so that it can support fine-grained comparison with the reference model.

[0110] The preset standard model indicates the expected expression of the target state after task completion, serving as the discrimination standard for acceptance judgment. The model can be obtained through expert rule setting, historical optimal task result sampling, or artificial modeling, and its data structure is consistent with the three-dimensional environment model, ensuring point-to-point, block-to-block, or semantic label-to-semantic label matching in the comparison process.

[0111] The comparison process is completed based on mechanisms such as spatial reconstruction registration, semantic coincidence matching, and error calculation. Spatial registration can use the Iterative Closest Point algorithm (ICP), global feature matching algorithm, or deep learning-driven feature pairing network to align the current three-dimensional environment model to the standard model coordinate system. Semantic matching identifies the displacement, missing, and incorrect placement of key objects, and then evaluates the task completion deviation. Error calculation can output global displacement metrics, key point error ranges, semantic matching rates, and other indicators to support acceptance decisions.

[0112] The final comparison result is used to generate an acceptance judgment label. If the error is within the set threshold, the subtask is considered qualified; if there are abnormal areas exceeding the allowed deviation, the area is marked as a difference area, and specific spatial coordinates, difference types, and impact levels are output. The comparison and judgment results are written into the acceptance report, which is structured to represent the task completion status, providing data for subsequent result archiving, rework scheduling, and system optimization.

[0113] The environment completion state data can be collected by a patrol robot equipped with an RGB-D camera. Image data is processed by a three-dimensional reconstruction algorithm to generate dense point clouds, and a spatial gridding process is used to construct a consistent three-dimensional model. Object recognition can use deep learning object detection algorithms such as YOLO or Mask R-CNN to extract controllable objects in the task area, such as furniture, clutter, medical devices, or device terminals.

[0114] The preset standard model can be sourced from different data sources. In the context of multiple task executions, historical task completion samples can be aggregated to generate diverse reference templates. In high-precision tasks, engineers can manually configure three-dimensional spatial structures, object placement states, and environmental cleanliness scoring models to form expert rule-defined standard models. In the comparison process, a hierarchical error evaluation mechanism is introduced. Coarse layer comparison is used for structure outline inspection, fine layer comparison locates key object displacement, and semantic layer comparison assesses object category errors or omissions. For example, the target is "desktop cleaning", and the system compares the occupancy rate change of the desktop area, whether the remaining item types and quantities are consistent with the expected model. The system crops and labels areas with comparison errors above the threshold to form difference areas for output. The acceptance result is output in a structured report format, supporting graphical visualization and structured database storage. In addition to error indicators, the report can also include confidence scores, task allocation record indexes, and recommended rework paths.

[0115] After the acceptance comparison of the task execution results, the system further analyzes the environment completion state data and task execution information obtained in the comparison to realize the continuous optimization of the ability cognition and scheduling mechanism. The environment completion state data represents the static space result after the task is completed, reflects the cleanliness, structure recovery and function achievement of the target area, and is derived from the environment scanning records of the inspection agent or terminal device after completing the task, commonly in the form of three-dimensional point cloud, depth image or semantic model, and is uniformly stored in the acceptance database. The task execution information is derived from the whole-process dynamic data flow generated during the execution process, including the start and end time, path trajectory, state change, execution exception, sensor input and system response of each agent, which is collected and archived in real time by the shared information platform.

[0116] Based on the environment completion state data, structured evaluation indicators can be constructed for the spatial state, such as spatial coverage integrity, residual area, object positioning accuracy, and interference area frequency, which are used to quantify the result achievement quality. Based on the task execution information, efficiency evaluation indicators can be constructed for the process behavior, such as unit time task quantity, path length optimization ratio, energy consumption and completion ratio, artificial intervention rate, and interruption frequency, which are used to depict the task execution efficiency and stability. These two types of evaluation indicators together constitute the task performance vector, reflecting the whole process performance of the task from the process to the result.

[0117] The above indicators are input into the machine learning model to establish the mapping relationship between the evaluation indicators and the ability parameters and the scheduling weights. The ability parameters are used to describe the execution ability performance of each agent under different task attributes, including but not limited to grasping accuracy, cleaning range efficiency, object recognition confidence, and path planning robustness. The task allocation weight parameter is used to determine the influence degree of each ability factor in the scheduling function, thereby affecting the future task allocation result.

[0118] The goals of model training and analysis include two dimensions: one is to dynamically adjust the agent ability parameters based on existing evaluation data to form a historical performance-driven ability perception mechanism; the other is to optimize the scheduling weight setting through evaluation data to make the scoring function closer to the actual performance difference. The model type can be selected according to the system size and task complexity, such as neural network, graph structure model or hybrid learning architecture, which supports nonlinear association modeling between heterogeneous abilities.

[0119] Finally, the system completes parameter update according to the model output result, writes the optimized ability parameters into the agent ability database as the basis for subsequent ability calling, and writes the optimized task allocation weight parameters into the task allocation strategy library for updating the task scheduling strategy. This mechanism builds a closed loop of ability evolution guided by acceptance results, realizing the self-adaptive evolution and precision enhancement of the multi-agent collaboration system in the long-term running.

[0120] The environmental completion state data can be based on the three-dimensional space image after semantic segmentation, extract key area features through space reconstruction software, extract object quantity, object position offset value, residual stain area, etc. Dimensional indicators, and convert them into uniform dimensional numerical vectors through standardization processing. Task execution information is aggregated by reading task life cycle logs and execution path records, including task duration, trajectory offset, path overlap, and abnormality frequency. Statistical methods and Bayesian modeling are used to build execution evaluation distribution.

[0121] The machine learning model can use a multi-task learning framework to build a dual-output path, one for learning ability parameter estimation and the other for learning weight parameter adjustment. Loss function is used to combine prediction bias and task completion score error for optimization. Model training can be updated online in real time, or asynchronously executed after the end of the task period. A sliding window is used to collect historical data samples to build a training set, and the model performance is evaluated based on validation indicators such as MSE, R 2

[0122] When the data volume is limited, transfer learning can be used to use the trained model in other similar scene tasks as initialization parameters to reduce model convergence time. In the task category diversification scenario, task type encoding vectors are introduced as additional inputs to the model to improve task heterogeneity adaptation capability.

[0123] Ability parameter updates are fused with old and new values in an exponential moving average manner to avoid drastic fluctuations in parameters caused by single abnormal results. Weight parameter updates are reconstructed in the form of score normalization to ensure the consistency of the weight sum of the scoring system.

[0124] Example: In the elderly care service scenario, after the health care robot completes the old room cleaning task, the inspection robot captures the distribution of items and the space cleanliness through three-dimensional scanning. The system compares the generated three-dimensional environment model with the standard bedroom cleanliness model, identifies that there are still items left under the bed that have not been cleaned, and marks the difference area in the acceptance report, providing accurate coordinates for subsequent task rework.

[0125] In the medical health scenario, the postoperative ward needs to be restored to a clean state. After the cleaning robot completes the task, the system collects the space state of key nodes such as the bed, medical cart, and disinfection area, and compares it with the standard sterile environment model to identify that a cart placement offset has caused a cleaning dead angle that has not been handled. The system determines that the cleaning task is not up to standard and notifies the re-scan.

[0126] In the unattended data center in the financial field, after the operation and maintenance robot completes the equipment aisle cleaning and ground inspection task, the system obtains the task area state modeling result, compares it with the baseline safety model, detects that there is still a messy cable area on the server side that has not been handled, triggers the unqualified determination, and generates a rework suggestion.​

[0127] The embodiment obtains environment completion state data and compares it with a preset standard model, so that the system can judge the task execution effect in a data-driven manner, and overcome the defect that the final state of space cannot be reflected only by relying on the task execution process log. Through visual modeling and structured comparison mechanism, the uncompleted or error completed area can be accurately identified and quantified as a difference index, and acceptance automation is further realized. The spatial positioning and task mapping ability of the difference area improves the abnormal closed loop processing ability of the system, so that the subsequent rework, evaluation and capacity update have a data basis.

[0128] The application relates to the technical field of artificial intelligence, and can be applied to business scenarios such as pension services, financial technology and medical health, and discloses a multi-agent collaborative task processing method, device, equipment and medium, which comprises the following steps: acquiring a task instruction and performing semantic analysis to identify a core task target, decomposing the task into a plurality of subtasks; acquiring working environment information and dividing a task area, combining an agent capability parameter and a region weight to generate a global collaboration framework; acquiring real-time state information of the agent, performing optimal agent matching based on the global collaboration framework to generate a task allocation table; issuing the task allocation table to control the agent to execute the task and upload task execution information; monitoring the execution process based on a shared information platform, triggering a dynamic adjustment strategy when a path conflict or a device fault is detected, generating and issuing an updated task allocation table; and acquiring environment completion state data after the subtasks are executed and comparing the environment completion state data with a preset standard model for acceptance. The application realizes efficient conversion of task instructions to multi-agent collaborative execution by combining task semantic information, environment structure information and agent capability parameters for unified modeling, and introducing real-time state sensing and dynamic allocation mechanism; and through real-time identification and adaptive adjustment of path conflicts and device faults, the stability and reliability of system collaboration are improved.

[0129] In one embodiment, the above step S10 comprises:

[0130] S101, receiving a task instruction of voice input;

[0131] S102, converting the task instruction into structured text information through a voice recognition module;

[0132] S103, inputting the structured text information into a pre-trained language model for semantic analysis, and extracting core semantic elements of the task instruction through the pre-trained language model;

[0133] S104, identifying a core task target based on the core semantic elements;

[0134] S105, decomposing the core task target into a plurality of subtask categories based on task type characteristics;

[0135] S106, defining an execution priority parameter based on the region relevance of the subtask category;

[0136] S107, storing the core task target, subtask category, and execution priority parameter into a task planning database.

[0137] In this embodiment, the task instruction acquisition process relies on a natural language interaction mechanism, and its input form can be voice, text, or multi-modal composite expression. Voice input is an important entrance for issuing unstructured tasks in adaptive service scenarios. Voice signals are processed by a speech recognition module that cooperates with acoustic modeling and language modeling. Deep neural network structures such as CTC or RNN-T are often used to convert voice streams into structured text information, with functions such as sentence breaking, error correction, and context preservation to ensure the accuracy of subsequent semantic analysis.

[0138] The recognized structured text information is input into a pre-trained language model that has context awareness and task entity extraction capabilities. A multi-layer encoder based on the Transformer structure can be used, such as BERT, RoBERTa, T5, etc. The multi-head attention mechanism is used to extract semantic elements related to action intent, object type, spatial location, and priority from the text. These semantic elements are not directly mapped but serve as intermediate semantic representations to construct semantic graph structures or task vector expressions for downstream task planning processes.

[0139] The recognition process of the core task target needs to be based on the dependency relationship between semantic elements, using a graph attention mechanism (GAT) or an intent reasoning module based on logical rules to extract high semantic density target definitions from action verbs, target objects, and task backgrounds, such as "organize the bedroom", "clean the living room", "put medicine in place", etc. A task target representation is established for specific execution behaviors. This target representation not only includes the task name but also includes object categories, target spatial locations, and expected states.

[0140] Once the task target is established, it will be decomposed according to the existing task templates or rule systems in the task library. Task type features can include task operation complexity, object number, and spatial coupling degree. For example, "organize the room" can be broken down into "store clothes", "clean the floor", and "put books in place". Each subtask category needs to obtain the corresponding operation chain through an action mapping table, such as "identify clothes -> grab -> put into the closet" and other action sequences.

[0141] The definition of the execution priority parameter is not based on static rules, but is calculated in combination with the task type characteristics and the spatial layout relationship. The priority judgment can consider the precedence relationship between nodes in the spatial dependency graph (Spatial Dependency Graph), such as "sweeping the floor" should be performed after "moving obstacles", or can be determined according to the object importance weight, such as "returning medicine" is prior to "arranging bookshelves". In addition, the current time, environmental constraints or user preferences can also be used as weighting factors to participate in priority calculation, forming a dynamic priority vector.

[0142] Finally, the core task target, subtask category and corresponding execution priority parameter need to be written into the task planning database, which has the task semantic version control ability, execution instance storage ability and task relationship network modeling ability, providing a semantic basis for subsequent task area division, task scheduling and execution feedback. The database structure is recommended to use a graph database (such as Neo4j) or a document-type database supported by a nested structure (such as MongoDB) to facilitate the expression of logical dependencies, shared parameters and state inheritance relationships between subtasks.

[0143] The embodiment can realize automatic understanding and decomposition of complex tasks in non-regular and non-programming interaction through structured processing and semantic modeling of natural language task instructions, improving the system's ability to analyze human intentions. By introducing pre-trained language models, semantic element extraction and task structure mapping processes, the implicit information in the instructions is restored and action-level decomposition is achieved, avoiding ambiguity and information loss in instruction understanding. Further combining task type characteristics and spatial logic relationship to dynamically generate execution priority parameters makes the execution order of subtasks no longer dependent on pre-set static templates, and can be adjusted adaptively according to the current task content and environmental conditions. After storing the task decomposition results in the task planning database, not only does it support the quick call of subsequent task scheduling algorithms, but also forms a structured accumulation of task execution knowledge.

[0144] In one embodiment, the above step S20 comprises:

[0145] S201, obtaining three-dimensional space data of a working environment through a space scanning sensor;

[0146] S202, identifying environmental structure characteristics based on the three-dimensional space data;

[0147] S203, dividing a plurality of task areas according to the environmental structure characteristics;

[0148] S204, determining an area parameter and a functional demand parameter of each task area;

[0149] S205, determining a region weight based on the area parameter and the functional demand parameter;

[0150] S206, querying the capability parameters of the plurality of agents from the agent capability database;

[0151] S207, calculating the matching degree score of each agent with each sub-task category based on the capability parameters;

[0152] S208, generating a priority allocation order according to the matching degree score;

[0153] S209, generating a global collaboration framework based on the priority allocation order and the regional weight.

[0154] In this embodiment, the process of obtaining work environment information is based on spatial perception, and the goal is to establish a quantifiable and structured spatial expression model for task area division and resource scheduling. This process starts with modeling the space of the work environment through spatial scanning sensors. The devices used can include RGB-D cameras, laser radars, structured light sensors, or multi-modal fusion devices. These devices convert the environment into three-dimensional spatial data in the form of point clouds, voxel maps, depth maps, or semantic grids. The three-dimensional spatial data needs to be processed by denoising, registration, and stitching to form a globally analyzable digital model of the environment, which has geometric and topological information accessible to algorithms.

[0155] On the basis of three-dimensional data, spatial structure recognition techniques are used to extract environmental structure features, including spatial closure detection, wall and door / window position recognition, planar region segmentation, furniture recognition, and spatial boundary fitting. The implementation methods can use RANSAC plane extraction, hyper-voxel clustering, deep convolutional semantic segmentation networks (such as MinkowskiNet), etc. The output results include spatial boundary coordinates, obstacle distribution maps, and candidate sets of functional areas, providing semantic support for subsequent task area division.

[0156] After the environmental structure features are identified, the entire work environment needs to be divided into multiple task areas. The division of task areas is not simply based on geometric distance or size, but on the fusion of multi-dimensional features such as spatial function, object distribution density, and cleaning complexity. For example, a living room can be divided into three task areas: "coffee table area", "sofa periphery", and "front side of TV cabinet", according to spatial accessibility, obstacle quantity, and task type adaptation level. The division process can construct a spatial similarity graph to aggregate spatial sub-blocks into task areas with consistent functions.

[0157] The area parameter and the functional requirement parameter of each task area need to be extracted. The area parameter can be accurately calculated by a three-dimensional modeling algorithm, and the functional requirement parameter can be calculated according to the objects contained in the area, the spatial density, the task frequency and other indicators. The functional requirement parameter represents the task complexity and the task priority in the task processing, for example, the "dining table" has high functional requirements due to the dense objects, and the "corridor" has low functional requirements due to the single cleaning. The functional requirement parameter can be quantified into dimensions such as expected load of task processing and task complexity factor, forming an area attribute vector.

[0158] Based on the combination of the area parameter and the functional requirement parameter, the area weight is determined. The area weight not only reflects the task amount, but also comprehensively considers the space difficulty and the resource scheduling priority. In the weight calculation, a regularization method can be introduced to prevent the weight of an individual large but low complexity area from being too high, affecting the balance of task allocation. The weight value can be normalized and used as a multiplication factor in subsequent scheduling calculations.

[0159] The ability parameter is obtained from the agent capability database. The ability parameter includes but is not limited to cleaning speed, grabbing accuracy, recognition accuracy, obstacle avoidance ability, energy efficiency, etc., and is formed through a static parameter table or dynamic performance backtracking. After retrieving these parameters from the database, the matching degree between each agent and each type of subtask needs to be calculated. The matching degree measures the adaptability of the agent to complete a specific subtask, and the calculation method can be based on multi-dimensional parameter similarity, historical execution performance or learning model prediction output. The matching degree value is usually represented as a score matrix, where the rows correspond to agents, the columns correspond to subtask categories, and the cells represent the corresponding matching scores.

[0160] After calculating each agent, the agents are sorted according to the matching score to generate a priority allocation order. The sorting mechanism can use a weighted average method, a hierarchical scoring mechanism or a heuristic sorting function to support matching high-adaptability agents to specific subtask categories, and to construct a task scheduling sequence.

[0161] On this basis, the priority allocation order and the area weight are fused to generate a global collaboration framework for the entire task scene. The global collaboration framework should have a three-layer nested structure of task partitioning, ability allocation and priority level, support multi-agent concurrent scheduling and task reconstruction, and reserve dynamic adjustment interfaces. The collaboration framework can be represented as a multi-layer graph structure, with task areas as nodes, weights as edge attributes, and agent matching as graph embedding paths. This structure serves as the basis for subsequent task allocation and dynamic collaboration decisions, providing a structured, dynamic and data-driven task collaboration model for the system.

[0162] The embodiment can map the spatial environment into discrete task areas with execution attributes by dividing the task areas based on environmental spatial data and structural features, introducing area parameters and functional demand parameters as quantitative basis, and significantly improving the accuracy and operability of task modeling. The dynamic adaptation between agents and tasks is realized through the matching degree scoring mechanism between the ability parameters and the task demand of the area, thereby avoiding resource mismatch and improving task execution efficiency. The fusion of priority allocation order and area weight can balance the task urgency and optimal resource utilization, and finally form a global optimal collaboration structure.

[0163] In one embodiment, the above step S30 comprises:

[0164] S301, associate each sub-task to the corresponding execution area;

[0165] S302, obtain the real-time state information of each agent including idle state, energy level and current position;

[0166] S303, determine the distance parameter of the current position of each agent and the sub-task execution area;

[0167] S304, extract the matching degree score of the agent and the sub-task from the global collaboration framework;

[0168] S305, determine the comprehensive score by combining the matching degree score, idle state, energy level and distance parameter;

[0169] S306, select the idle agent with the highest comprehensive score for each sub-task, and generate a task allocation table containing sub-tasks, execution areas and allocated agents;

[0170] S307, store the task allocation table to the task allocation database.

[0171] In the embodiment, associating each sub-task to the corresponding execution area is the first step to realize accurate task distribution, and this process establishes the mapping relationship between task semantic information and spatial topology. Each sub-task has a category label and a target behavior description in the semantic planning stage, such as "arranging the desktop", "cleaning the floor", "moving the clutter", etc. Combined with the task area division result generated in the environmental structure analysis, the area binding is completed according to the matching degree between task semantics and area functions, for example, the "arranging the desktop" task is one-to-one mapped with the "desktop area in the study". This process can be automatically completed through similarity calculation between task labels and area semantic labels, or mapping rules can be pre-set in the rule base to ensure the accuracy of task spatial positioning.

[0172] Acquiring real-time state information of each agent is the prerequisite for dynamic scheduling. The state information includes at least three dimensions: idle state, energy level, and current position. The idle state indicates whether the agent is in a schedulable state, which is usually obtained from the control instruction response system or the execution thread management module, and identifies whether the current number of task execution threads is zero. The energy level reflects the endurance capability of the agent, which is obtained from the energy monitoring module and can provide information such as the remaining battery percentage and battery health status. The current position represents the real-time coordinates of the agent in the environment space, which is obtained from the positioning module (such as visual SLAM, laser SLAM, or UWB positioning system) and is usually represented as a three-dimensional position vector in the world coordinate system or a two-dimensional plane coordinate.

[0173] Determining the distance parameter between the current position of each agent and the sub-task execution area is an important means of quantifying the scheduling cost. The task execution area can be modeled as a spatial boundary or a center point, while the current position of the agent is a dynamic point. Based on this, the distance between the two can be calculated using Euclidean distance, A* path length, navigation cost, etc. The specific method is determined according to the environmental complexity and computational resource constraints. In high-degree-of-freedom environments, path cost distance should be used to reflect the real navigation complexity, and graph search or topological graph cost function should be introduced in obstacle-dense or multi-layer structure scenarios.

[0174] Extracting the matching degree score of the agent and the sub-task from the global collaboration framework is a quantitative inheritance of the historical adaptability between the agent's ability and the task requirements. The matching degree score is modeled by the ability parameter, task category attribute, and historical success rate in the previous stage, and is in the form of a two-way mapping matrix, with the agent as the row and the task category as the column. The score represents the expected adaptability of completing a certain type of task. When extracting, the scheduling score module in the collaboration framework structure can be directly accessed to quickly return the evaluation value of the current agent for the target task.

[0175] The matching degree score, idle state, energy level, and distance parameter together constitute the basis for scheduling decisions. The calculation of the comprehensive score requires the construction of a scheduling scoring function, which can use weighted linear combination or nonlinear evaluation models. In simple strategies, the following function can be defined:

[0176] S = a x M + b x E + g x (1 - D) + d x I

[0177] where S is the comprehensive score, M is the matching degree score, E is the normalized energy level, D is the normalized distance cost, I is the idle state identifier (binary), a, b, g, d are scheduling weight parameters. The weight values can be obtained through empirical configuration, historical task performance regression, or machine learning model dynamic learning. In more complex scenarios, a neural network structure can also be used, with the above parameters as input and the optimal scheduling value as output.

[0178] After the comprehensive score calculation is completed, for each subtask, the system will select the one with the highest comprehensive score from the set of agents currently in an idle state as the performer, ensuring that optimal agent resources are allocated first. To ensure balanced resource distribution, a task load record or cooling time mechanism can be introduced to suppress the number of consecutive dispatches of the same agent, preventing resource overload.

[0179] The final task allocation result is composed of a triple, namely (subtask ID, execution area ID, allocated agent ID). The task allocation table summarizes the allocation information of all subtasks, and is usually in the form of a task item list or a mapping table structure, supporting subsequent instruction issuance and dynamic adjustment. The task allocation table is stored in the task allocation database, ensuring traceability and auditability of the scheduling results, and providing basic data support for fault recovery and task reallocation. The database supports an asynchronous update mechanism to maintain data consistency with the shared information platform.

[0180] The spatial binding mechanism of the task execution area and the subtask in this embodiment makes the task scheduling process have clear spatial constraints, avoiding positioning ambiguity and path redundancy in task execution. By constructing an agent state portrait based on multi-dimensional state information, accurate identification of schedulable resources is achieved, effectively avoiding the risk of task failure caused by insufficient energy or excessively long paths. The comprehensive score mechanism integrates multiple factors such as matching degree, energy consumption, distance, and availability during scheduling, with good strategy flexibility and scheduling robustness. The structured storage of task allocation results ensures the integrity and traceability of the task management process, forming a closed-loop scheduling system.

[0181] In one embodiment, the above step S40 comprises:

[0182] S401, parsing the subtask content, execution area, and allocated agent in the task allocation table;

[0183] S402, sending a task start instruction containing the execution area to the allocated agent, controlling the allocated agent to move to the execution area and start the task program;

[0184] S403, collecting environmental dynamic data through the sensors carried by the allocated agent;

[0185] S404, converting the environmental dynamic data into structured execution information;

[0186] S405, marking the state label of the structured execution information;

[0187] S406, uploading the structured execution information with state labels to the shared information platform in real time;

[0188] S407, updating the working area state in the shared information platform.

[0189] In this embodiment, parsing the task allocation table is not only reading data, but also the first step of deconstructing the scheduling intent. The task allocation table encapsulates information such as sub-task identifiers, target execution areas, target state expectations, assigned agents, task priorities, time limit constraints, etc. The parsing module needs to support complex structured data processing, including multi-table joining, enumeration type identification, field integrity checking, syntax and semantic consistency checking. At the same time, it needs to correct historical scheduling records and boundary conditions, such as whether the current agent has executed such tasks. The process uses Abstract Syntax Tree Parsing (AST) and data constraint checking mechanism to ensure the reliability of parsing.

[0190] Sending task start instructions to designated agents involves high coordination between task delivery channels and control interfaces. The start instruction encapsulates area information (such as geometric boundary points, job height, task path suggestions, etc.), behavior templates (action sequence, execution module call), fault tolerance mechanism (number of retries, rollback conditions) and priority label. The system uses REST, DDS or ROS2 service call protocol to safely communicate the instructions to the agent control unit. After receiving the start instruction, the control module triggers local path planning, safe entry logic, job module activation logic and execution monitoring sub-module, ensuring the chain coordination from delivery to execution start.

[0191] During execution, the agent carries a sensor array to continuously collect environmental dynamic data. Dynamic data includes visual information (images, depth maps), laser scanning information, tactile feedback, sound events, force sensor readings, etc. Data collection follows a high-frequency sampling strategy and an asynchronous event triggering mechanism to ensure that key environmental changes are accurately captured. The raw format of sensor data is a continuous time series with high-dimensional characteristics, which needs to be processed through filtering (such as Kalman filtering, time window moving average) and noise reduction (such as median filtering, Fourier filtering).

[0192] Structuring environmental dynamic data into execution information involves mapping from perception conditions to execution semantics. Structured information includes object positions, state changes (such as whether the door is closed), execution completion sub-steps (such as objects being grasped), abnormal events (such as grasping failure, obstacle obstruction), etc. The conversion method combines image semantic segmentation (such as Mask R-CNN), object recognition models, state machine labeling mechanisms to construct event log tables and state change timestamps.

[0193] The state label of the structured execution information is marked to reflect the task progress and abnormal types. The state label is divided into several categories: preparing, executing, stage completed, abnormal to be handled, retrying, completed, etc. The agent needs to write the state label into the information packet header through the execution feedback loop, and add state code and detailed execution data, so that the scheduling and monitoring modules can interpret it in real time.

[0194] Real-time uploading of structured execution information with status tags requires a high-bandwidth low-latency information channel. Parallel batch uploading and incremental event uploading mechanisms are adopted: regular data is uploaded in batches by time segments, while abnormal or status change events trigger immediate single uploading. The communication framework can use WebSocket or MQTT QoS2 guarantee mechanism, combined with local transmission buffering module and uploading confirmation mechanism to ensure uploading reliability.

[0195] After the shared information platform receives the data, it needs to map it into the "work area state model", update the task completion degree, abnormal nodes, resource occupation state, area safety state, etc. in the space, and provide an interface for the scheduling, monitoring and learning modules to access in real time. The update model ensures state consistency and supports multi-view mapping, such as two-dimensional plan view, three-dimensional environment view and time axis view.

[0196] This embodiment realizes high-fidelity communication of scheduling decisions to the agent operation layer by explicitly parsing the task allocation table content and standardizing the task start mechanism, significantly improving the task execution accuracy. In the execution process, multi-source dynamic data acquisition and structured processing flow are introduced, making the task state not only controllable but also verifiable, ensuring the system to maintain global consistency and continuity under the asynchronous cooperation of multiple agents. The standardized uploading and unified labeling of task execution information make each stage of task fulfillment highly traceable and automatically assessable, providing reliable support for subsequent parameter optimization, fault response and scheduling redistribution of the scheduling model.

[0197] In one embodiment, the above step S50 comprises:

[0198] S501, monitoring the movement trajectory of the agent in the shared information platform;

[0199] S502, when detecting that the movement trajectories of two agents intersect at a distance less than a safety threshold, determining a path conflict event;

[0200] S503, determining the task urgency of the agent that has a path conflict based on the work area state and task progress in the shared information platform;

[0201] S504, generating a passing order decision according to the task urgency;

[0202] S505, controlling the low-urgency agent to pause movement based on the passing order decision;

[0203] S506, planning an avoidance path for the agent that pauses movement, and updating the space occupation state of the shared information platform;

[0204] S507, when processing the path conflict event, sending a moving instruction containing the avoidance path to the temporarily paused agent;

[0205] S508, when detecting that the energy level of the agent is lower than the failure threshold, determining that a device failure event occurs;

[0206] S509, calling a capability-matched idle agent for the failed agent in the device failure event;

[0207] S510, obtaining a work progress parameter of the failure task from the shared information platform;

[0208] S511, determining a remaining workload parameter based on the work progress parameter;

[0209] S512, when processing the device failure event, sending a new task start instruction containing the remaining workload parameter to the idle agent;

[0210] S513, generating an updated task allocation table based on the moving instruction and the new task start instruction;

[0211] S514, issuing the updated task allocation table to the corresponding agent.

[0212] In this embodiment, the continuous monitoring mechanism of the shared information platform is developed for the whole cycle task execution process of the multi-agent system, and the core is real-time collection, centralized synchronization, dynamic analysis, and driving immediate response based on task behavior data. The agent moving track can be constructed by the odometer, visual SLAM module, IMU sensor, UWB module and the like of each agent end, and uploaded to the shared information platform through the message middleware, to form a continuously trackable multi-agent position data set in space-time dimension. The data set supports refreshing the state at a millisecond level, ensuring that the platform has immediate control ability over the global motion situation.

[0213] On this basis, by calculating the intersection set of the moving tracks of any two agents and the nearest distance value thereof, and comparing with the set safety threshold, automatic detection of potential path conflict is realized. The setting of the safety threshold needs to be determined comprehensively in combination with the robot size, navigation accuracy, task density and other parameters in specific scenes, which essentially defines the behavior safety boundary. When the track intersection area is less than the threshold, it is immediately determined that a path conflict event occurs.

[0214] The processing of path conflict events introduces the concept of task urgency, which is calculated based on the current task progress and region state of the working area. Task progress reflects the degree of completion of the current area task, which can be determined by task markers, completion identifier counting, or task feedback logs; the region state can be constructed based on the spatial cleanliness and spatial coverage information fed back by the environment perception module. Through the urgency calculation function, each conflicting agent is given a priority score.

[0215] Based on the urgency score, a passing order is formed to control the suspension of movement of low-priority agents to ensure the smooth passage of high-priority paths. The suspension operation is performed by sending a motion command to inhibit the navigation controller, while maintaining local state perception and avoiding path overlap.

[0216] For suspended agents, further avoidance path planning is performed, calling global or local path planning modules (such as A*, DWA, RRT, or optimization graph methods) to generate temporary movement paths that do not conflict with other agent trajectories based on the current spatial occupancy map, and to update the spatial occupancy state of the regions involved in the path on the shared information platform in real time.

[0217] Thereafter, the platform sends specific movement instructions for the avoidance path to the suspended agent, which can include path point sequences, end coordinates, speed parameters, etc., to ensure that its behavior runs independently under obstacle avoidance logic.

[0218] When the monitoring module detects that the energy level of an agent (provided by the battery management system or power consumption prediction module) is below a preset threshold, a device failure event is triggered. Such events require dynamic retrieval of idle agents with high compatibility with the failed task, by querying the ability database and selecting entities currently in an idle state, and sorting and selecting based on the ability matching degree with the task type.

[0219] Subsequent operations need to obtain the working progress parameters of the failed task from the shared information platform, including the task completion ratio, the remaining path of the task, the amount of data collected, etc., and calculate the remaining workload based on it, such as cleaning area, number of transported objects, or number of scanning frames.

[0220] Based on the remaining workload, a new task start instruction is generated for the selected idle agent, including task target, execution area, remaining sub-task identifier, etc. The new instruction directly starts the task execution process to ensure that the task continues to advance without interruption.

[0221] Finally, the platform integrates all the issued mobile instructions and new task instructions, reconstructs the task allocation table to cover the current task state and execution subject changes. The new task allocation table needs to pass through consistency check and state mapping confirmation before being issued to ensure that no execution conflicts or resource reuse problems are introduced. The table is synchronized to all related agents through message queue or instruction bus to achieve overall adjustment and linkage recovery of distributed tasks.

[0222] The embodiment can realize rapid identification of path conflict and device failure events by continuously monitoring the moving trajectory and energy state of the agent in the shared information platform; dynamically adjust the behavior of low-priority agents by generating a passing order based on the task urgency and planning an avoidance path, thereby effectively relieving local path congestion problems; realize agent replacement without interrupting the task by extracting the work progress of the failed task and generating a replacement instruction for the idle agent; ensure the task cooperation continuity and execution consistency of the entire multi-agent system by integrating the mobile instruction and the replacement task instruction to generate an updated task allocation table and reissue it.

[0223] In one embodiment, the above step S60 comprises:

[0224] S601, controlling a patrol device to scan a task execution area and collecting environmental space data of the task execution area through the patrol device;

[0225] S602, generating a three-dimensional environment model based on the environmental space data;

[0226] S603, obtaining preset standard model data;

[0227] S604, comparing features of the three-dimensional environment model and the preset standard model, and marking a difference area in the three-dimensional environment model that does not meet the standard;

[0228] S605, generating an acceptance report containing coordinates of the difference area, and storing the acceptance report in an acceptance database.

[0229] In the embodiment, after the task is completed, to realize systematic verification of the execution result, the spatial state expression information of the task execution area after the task is completed needs to be obtained, and the state is compared and analyzed with the predefined expected standard, thereby realizing structured acceptance of the task quality. First, a patrol device is controlled to scan the task execution area. The device is usually a mobile agent, which is equipped with spatial perception components such as laser radar, RGB-D camera, TOF depth camera, and multi-band visual sensor, and can independently plan a scanning path to perform full-coverage data collection on the specified area without disturbing the on-site arrangement.

[0230] The original environmental spatial data collected during the inspection process reflects the real spatial state after the task is completed. The data types can be dense point cloud, depth image, structured light reconstruction fragment, and local mesh graph. Time stamps, pose information, and scene labels can be attached during the collection process to support subsequent temporal and spatial reconstruction. The original data needs to be preprocessed, including registration, filtering, denoising, coordinate unification, and view fusion, to generate a spatial expression result with topological continuity and semantic consistency, i.e., a three-dimensional environmental model.

[0231] The constructed three-dimensional environmental model should have the characteristics of complete spatial coverage, clear geometric boundary, and accurate semantic partitioning, to be used for comparison with the preset standard model. The preset standard model represents the spatial state under ideal task completion conditions and is generated based on task categories, regional functions, historical high-quality completion samples, or user-set expectations. Its data form can be a three-dimensional point cloud template, a semantic occupancy map, a CAD modeling file, or a multi-scale spatial diagram, and it is compared with the current environmental model through feature comparison operations.

[0232] During the comparison process, a multi-level spatial feature matching strategy is adopted. First, the geometric level spatial position alignment is performed, and through the Iterative Closest Point (ICP) algorithm, global registration network, and feature point matching, the coarse and fine registration of the two models is realized. Subsequently, the semantic level is analyzed to analyze the compliance of object categories, spatial boundaries, and item positions, and to identify abnormal placement, residual foreign matter, and incomplete cleaning that violates the standard. All positions or regions that do not meet the standard conditions are labeled as difference regions, and their corresponding spatial coordinates, physical dimensions, and deviation types are extracted.

[0233] The above difference regions are used as the acceptance results to be corrected, and a structured acceptance report is generated. The report records the spatial positioning information, problem category label, and deviation degree quantization value of each difference region in detail, forming an interpretable result data for the execution system, user, or supervisor. The acceptance report is generated in the form of a data table, a labeled graph, or a model annotation, and is written into an acceptance database. The database supports task-level indexing, regional-level comparison, and temporal evolution analysis, ensuring the traceability and sustainable optimization capability of the acceptance results.

[0234] Example: In a smart nursing home with multi-agent service capabilities, the system receives a voice task instruction from the administrator: "Please arrange for the robot to clean the room of the elderly A101 today morning, make the bed, pick up the floor debris, and wipe the table." The task is first converted into structured text information through the voice recognition module, and the pre-trained language model is used for semantic analysis to identify key actions and objects such as "cleaning", "making the bed", "picking up floor debris", and "wiping the table". The system further divides the task into subtasks: ① bed making, ② floor cleaning, and ③ table wiping, and determines the required operation capability type and device support capability based on the task type.

[0235] Subsequently, the system dispatches the spatial scanning sensor to collect the three-dimensional spatial data of A101's room, identifies the room layout, furniture distribution, and passage channel structure characteristics, and accordingly divides the task area, such as the bed area, floor area, and table area, and labels their area, purpose label, complexity level, and other parameters to further determine the task weight of each area. For example, the bed area is directly related to the elderly's sleep, with a weight of 0.45, the floor area is 0.35, and the table area is 0.2. At the same time, the system retrieves the current in-house all idle service robot capability parameter data from the agent capability database, including grabbing accuracy, surface wiping ability, navigation ability, obstacle avoidance stability, and other indicators.

[0236] The system calculates the matching degree of each subtask category and the agent capability parameters, generates a matching score, and generates a global collaboration framework based on the weight of each task area to form a priority allocation order of the agent to the task area. At this time, the system does not immediately assign the task, but synchronously retrieves the real-time state information of the agent, such as location information, power level, idle state, execution history, etc., quantifies the distance parameters between the current location and the target task area, and combines the task matching score, energy margin, and executability to comprehensively evaluate the current adaptability of each agent to each subtask.

[0237] For example: cleaning robot A3 is currently closest to the floor area, and its floor cleaning ability score is higher, with sufficient energy and in idle state, so the system assigns it to subtask ②; while robot A1 has high-precision operation ability, although its navigation speed is slightly lower, it is assigned to subtask ① because it has the strongest grabbing and detail arrangement ability. Finally, the system forms a task allocation table, which includes the names of each subtask, the corresponding execution area, the specific robot number assigned, and the expected execution window time, and stores the task allocation table in the database, and then issues it to each relevant agent.

[0238] After receiving the task, each service robot starts moving to the designated task area and starts the corresponding task program. In this process, each agent continuously collects dynamic data during execution through the sensors carried (such as infrared camera, pressure touch module, surface smoothness identifier), including environmental changes, item status, completion status, etc., and structures these data into execution information with state labels, and uploads them to the shared information platform in real time. For example: Robot A3 uploads "ground cleaning completion status: 90%, remaining area near the door exists obstacles", and the system updates the space state mapping accordingly.

[0239] During the execution of the task, if the system detects that two robots (such as A2 and A3) have potential path conflicts (predicted intersection distance less than 0.5 meters) in a narrow passage, it will immediately trigger a dynamic adjustment strategy through the shared information platform, analyze the task urgency of each robot, and control the lower priority robot (such as A3) to pause moving, and plan a detour path for it. The adjustment results are issued through the moving instructions for execution. If an agent (such as A1) is automatically offline due to low battery during the execution of the desktop wiping task, the system will immediately judge it as a device failure event, call the idle agent A5 with replacement ability, and determine the remaining task content by comparing the historical task state, and reissue the new task start instruction to ensure that the task is not interrupted.

[0240] After all sub-tasks are completed, the system dispatches the inspection robot into room A101 to perform structured acceptance of the cleaning effect. The inspection device performs three-dimensional space reconstruction, collects current room state data, and generates a three-dimensional environment model. The system retrieves the preset standard model from the acceptance database before the task is completed, including the expected shape of a clean bed, the space template of a desk without clutter, and the ground cleaning degree image sample, etc. Through model registration and spatial feature comparison, it identifies the deviation area. For example: if the corner of the floor detects residual paper scraps, the system will record "there is a cleaning deficiency in the lower right corner of the floor area, the deviation area is 12 cm 2 ", and generate a difference label with coordinates.

[0241] Finally, the acceptance report is written into the acceptance database and serves as an important basis for subsequent task evaluation and ability learning, participating in the process of ability parameter optimization and task allocation strategy adjustment. The whole process realizes the service closed loop of the elderly scene from semantic perception to space scheduling, from execution cooperation to result acceptance, improves the task adaptation accuracy, response efficiency and service quality, and especially shows good adaptability and human-machine cooperation ability in the needs of flexible arrangement, cleaning and maintenance, and abnormal handling in the living space of the elderly.

[0242] In a large intelligent medical institution, the nursing dispatch center system receives a task request from a doctor through a voice terminal: "Arrange a robot to complete bed arrangement, instrument positioning, and floor cleaning in room 303." The task is converted to structured text through a voice recognition system, and the core task goal is identified as "restoring the room environment" through semantic understanding. Based on this, three sub-tasks are divided: ① bed arrangement, ② medical instrument classification and positioning, and ③ floor cleaning. The system further analyzes the spatial elements and operation types involved in the task semantics, defines task type labels and operation priorities.

[0243] After that, the spatial perception system deployed by the medical institution uses ceiling LiDAR and visual cameras to perform three-dimensional scanning of room 303, generating complete spatial point cloud and image fusion data, and identifying spatial structure features such as bed area, instrument rack, and passageway. Based on these structure information, the system divides the three tasks into three task areas and calculates the area, complexity, and criticality indicators of each area, such as the bed area which involves direct contact with patients, with a task weight of 0.5, the instrument area with a weight of 0.3, and the floor area with a weight of 0.2. At the same time, the system queries the agent capability database for all standby robots, including mechanical arm stability, disinfection spraying precision, instrument recognition accuracy, and obstacle avoidance ability.

[0244] Subsequently, the system performs capability matching evaluation between each agent and sub-task, calculates the matching score, and synthesizes the priority allocation matrix based on the task area weight and matching degree to form a global collaboration framework. At this time, combined with the current real-time state of each agent, such as robot C1 being charged, C2 being idle and close to room 303, C3 currently executing a task in room 304 and about to end, etc., the system calculates the comprehensive adaptation score of the current schedulable agent to each sub-task based on its location, idle state, power level, and task history stability.

[0245] For example, disinfection robot C2 is selected to perform task ③ floor cleaning because it is currently located near room 303 and has high wiping capacity; C4 with high-precision grasping ability is dispatched to perform task ② instrument positioning; and nursing robot C5 is assigned to task ① bed arrangement because it has a bed surface flattening recognition module. The system generates a complete task allocation table and sends it to each agent, and synchronously writes the task content into the task allocation database, recording the dispatch behavior.

[0246] During task execution, each robot automatically goes to the corresponding area according to the task instructions, starts the program, and collects execution data in real time through the vision and tactile sensors carried by itself. For example, C5 uses a tension sensor to perceive the balance of the bed sheet tension and continuously corrects it; C4 identifies the shape label of medical scissors and forceps and classifies and returns them to their places. All robots will structure the dynamic data in the execution process and upload it to the shared information platform, including execution path, abnormal mark, task status and other data content. The platform continuously updates the task completion status diagram and the agent position diagram of the current ward, which provides reference for the scheduling system.

[0247] If a path conflict occurs during execution, for example, C2 and C5 intersect in the right channel of the bed due to unsynchronized coordination sequence, the system will detect that the intersection point of the two moving trajectories is less than the safety threshold of 0.4 meters, and trigger the conflict resolution strategy. At this time, after analyzing the task urgency of both sides, it is determined that the bed arrangement has higher priority, and C2 is instructed to pause and re-plan a detour path to complete the avoidance. If C4 has an abnormal mechanical arm force feedback during execution and interrupts the task, the system determines that it is a device failure, immediately calls C6 in the idle state and with similar capabilities to take over the unfinished task, and issues a follow-up execution instruction based on the recorded task progress.

[0248] When all sub-tasks are completed, the system dispatches a patrol robot into the ward to structure the acceptance of the overall task result. The patrol robot obtains the current state data of the ward through the 3D camera and surface cleanliness sensor, and generates a three-dimensional environment model, which is compared with the standard ward state model preset by the medical institution in terms of spatial structure and functional area. The system identifies the degree of bed sheet wrinkle, the placement coordinates of instruments, the ground cleanliness light reflection value and other dimensions, labels the non-compliant areas and generates a detailed acceptance report, such as "bed edge flatness is less than 8mm", "instrument return deviation exceeds the standard by 3cm", and stores the acceptance report in the acceptance database.

[0249] On this basis, the system extracts the environment completion state data in the acceptance report and the operation log in the task execution process, generates environment quality evaluation indicators and task efficiency indicators, and learns the mapping relationship between the ability parameters of each agent and the task performance through a pre-trained neural network model. For example, it is found that the cleaning efficiency of C2 is better than expected, and the system will increase the "ground cleaning efficiency" value in its ability parameters to 0.87, and update the "disinfection task matching degree weight" in the task allocation weight parameters, to improve its priority in future similar tasks. The optimization results are written into the agent capability database and the task allocation strategy library, respectively, to provide intelligent scheduling support for future similar tasks.

[0250] Through the above process, the system realizes the whole-process closed-loop management and control of multi-task, multi-ability, and multi-state coordination scheduling in a medical environment, significantly improves the efficiency of robot resource allocation, the recovery speed of ward environment, and the reliability of service response, and has good applicability in scenes with high sensitivity to environmental state such as postoperative ward and intensive care area.

[0251] During the operation period of a smart bank branch, the operation management system receives a service optimization request: "Adjust the robot configuration for hall customer guidance, document delivery, and terminal guidance." The system collects voice task instructions through the front-end voice interaction terminal, converts them into structured text relying on the language recognition engine, and inputs them into the natural language understanding model to identify that the task goal is "to reconstruct the service process of the bank hall." The semantic deconstruction result shows that the goal can be divided into three service sub-tasks: customer guidance, document delivery, and self-service terminal guidance, corresponding to spatial areas such as the customer waiting area, the document processing area, and the self-service area, respectively. Task priority parameters are established according to customer flow density and task response time limit.

[0252] The system retrieves data from the multi-source sensing devices deployed in the hall, including ceiling panoramic cameras, laser radars, and RFID landmark tags, to construct a real-time three-dimensional spatial state model of the hall. The system identifies spatial structure units including the consultation desk, VIP area, and self-service area, and then divides them into multiple task areas. Each area constructs an area weight model based on dimensions such as area, flow density, service type, and scene complexity, for example, the customer guidance area is set to a weight of 0.4, the document delivery area is set to 0.35, and the self-service equipment area is set to 0.25. The system then reads data from the agent capability database, including the navigation accuracy, voice broadcast clarity, document recognition ability, and equipment operation familiarity of the service robot cluster.

[0253] Based on the relevance of capability parameters and sub-task types, the system calculates the matching score of all schedulable agents, combines the area weight to generate a priority scheduling order, and constructs a global collaboration framework for the current bank branch service scene accordingly. The real-time state sensing system synchronously collects information such as the current position of each robot, whether it is in a task state, and the remaining power, and calculates the distance parameters to each task area. The system integrates the current capability matching degree, real-time availability, remaining energy, and path cost to form a comprehensive scoring mechanism, and completes the optimal task allocation. For example, Robot-B has the highest score because it has stable voice broadcast frequency, is familiar with the layout of the hall, and is idle, so it is assigned to complete the customer guidance task; Robot-C is assigned to complete the document delivery task; and Robot-D has a terminal indicator light recognition module, which is suitable for the terminal guidance task.

[0254] The task allocation table is generated and then issued to each service robot. Robot-B moves to the bank entrance according to the task start instruction, performs the guiding task for the customers entering the store, and broadcasts the waiting time and queue number of each window in real time; Robot-C departs from the data archiving area, follows the path planning to go to the back office to pick up the contract text, and avoids collision with customers on the way through the sensor; Robot-D patrols the self-service terminal area, actively identifies customers who are stagnating and provides device usage guidance. Each robot collects dynamic data including service response time, task completion, path deviation, customer stay behavior, etc. during the execution process, and uploads them to the shared information platform in real time to build a complete service state diagram.

[0255] If Robot-C has a low battery warning during file transmission, the system automatically triggers the fault response mechanism, extracts the current task information and progress parameters, and reassigns them to Robot-E which is in an available state and has the same transmission capability. The scheduling system calculates the proximity to the target path and the current workload, evaluates the feasibility, and issues a task start instruction to replace the task, realizing uninterrupted file transmission. If Robot-B and Robot-D have path intersection due to the gathering of customers in the self-service device area during peak hours, the system detects that the trajectory intersection is less than the safety threshold of 0.3 meters, calculates the task urgency, and determines that Robot-D needs to have priority, guiding Robot-B to wait and detour, avoiding confusion or congestion in the scene.

[0256] After the task is completed, the system dispatches the acceptance robot to patrol the entire service environment to obtain space usage status and customer distribution change data. Based on visual perception and semantic analysis, a three-dimensional service state model is generated and compared with the preset standard office service state model, including customer flow efficiency, device usage status, file transmission error, etc. The difference area is marked to form a structured acceptance report. The system identifies data such as "average customer stay time in the self-service terminal area is longer" and "file transmission path deviation frequency is higher", and pushes them to the acceptance database.

[0257] The system further extracts the acceptance report and execution log information to generate environment quality evaluation indicators and task efficiency evaluation indicators, such as service response time, path repetition rate, interaction completion rate, etc. The evaluation indicators are input into the neural network analysis model to build a nonlinear relationship between task performance and capability parameters, and output the optimized values of capability parameters and task allocation weights. For example, the system identifies that Robot-D has good adaptability in the terminal guidance task, and adjusts the "device understanding ability" parameter and increases the matching weight of this capability in the terminal scenario in the task scheduling function. These updated data are stored in the agent capability database and the task allocation strategy library respectively, which are used to optimize the scheduling strategy of subsequent bank branch daily operations.

[0258] The embodiment can restore the spatial state after the task is completed in a structured manner by controlling the patrol equipment to perform spatial scanning on the task execution area and constructing a three-dimensional environment model; the task execution quality can be quantitatively evaluated by comparing the three-dimensional model with a preset standard model and accurately marking the difference area; the digital storage and result tracing of the task completion effect are realized by generating an acceptance report and storing it in an acceptance database. The process not only significantly improves the result verifiability and quality control capability of the multi-agent system in collaborative task execution, but also provides a solid data foundation for subsequent data-driven optimization based on feedback.

[0259] In an embodiment, a multi-agent collaborative task processing device is provided, which corresponds to the multi-agent collaborative task processing method in the above embodiments. Referring to Figure 3 , Figure 3 A functional module schematic diagram of a preferred embodiment of the multi-agent collaborative task processing device is shown. The task analysis module 10, the task planning module 20, the task allocation module 30, the task execution module 40, the dynamic scheduling module 50 and the task acceptance module 60. The detailed description of each functional module is as follows:

[0260] The task analysis module 10 is used to obtain task instructions, perform semantic analysis on the task instructions to identify a core task target, and decompose a task corresponding to the task instructions into a plurality of sub-tasks based on the core task target;

[0261] The task planning module 20 is used to obtain work environment information, divide a plurality of task areas according to the work environment information and determine a region weight of each task area, generate a global collaboration framework in combination with a capability parameter of a plurality of agents and the region weight;

[0262] The task allocation module 30 is used to obtain real-time state information of a plurality of agents, match an optimal agent for each sub-task in combination with the global collaboration framework and the real-time state information, and generate a task allocation table;

[0263] The task execution module 40 is used to issue the task allocation table to the plurality of agents, control the agents to execute the sub-tasks, and upload task execution information collected in the execution process to a shared information platform;

[0264] The dynamic scheduling module 50 is used to monitor the execution process of the agents through the shared information platform, trigger a dynamic adjustment strategy for conflict resolution or task re-allocation when detecting an agent path conflict or equipment failure, generate an updated task allocation table and reissue it;

[0265] The task acceptance module 60 is configured to obtain environment completion state data after all sub-tasks are executed, and compare the environment completion state data with a preset standard model for acceptance.

[0266] In an embodiment, the task analysis module 10 is specifically configured to:

[0267] receive a task instruction in voice input;

[0268] convert the task instruction into structured text information through a voice recognition module;

[0269] input the structured text information into a pre-trained language model for semantic analysis, and extract core semantic elements of the task instruction through the pre-trained language model;

[0270] identify a core task target based on the core semantic elements;

[0271] dissolve the core task target into a plurality of sub-task categories based on task type characteristics;

[0272] define an execution priority parameter based on the region relevance of the sub-task categories;

[0273] store the core task target, the sub-task categories, and the execution priority parameter into a task planning database.

[0274] In an embodiment, the task planning module 20 is specifically configured to:

[0275] obtain three-dimensional space data of a working environment through a space scanning sensor;

[0276] identify environment structure characteristics based on the three-dimensional space data;

[0277] divide a plurality of task regions according to the environment structure characteristics;

[0278] determine an area parameter and a functional demand parameter of each task region;

[0279] determine a region weight based on the area parameter and the functional demand parameter;

[0280] query capability parameters of a plurality of agents from an agent capability database;

[0281] calculate a matching degree score of each agent with each sub-task category based on the capability parameters;

[0282] sort the matching degree scores to generate a priority allocation order;

[0283] generate a global collaboration framework based on the priority allocation order and the region weight.

[0284] In an embodiment, the task allocation module 30 is specifically configured to:

[0285] associate each subtask to a corresponding execution area;

[0286] obtain real-time state information of each agent, including idle state, energy level and current position;

[0287] determine a distance parameter of the current position of each agent and the execution area of the subtask;

[0288] extract a matching score of the agent and the subtask from the global collaboration framework;

[0289] determine a comprehensive score in combination with the matching score, the idle state, the energy level and the distance parameter;

[0290] select an idle agent with the highest comprehensive score for each subtask, and generate a task allocation table containing the subtask, the execution area and the allocated agent;

[0291] store the task allocation table to a task allocation database.

[0292] In an embodiment, the task execution module 40 is specifically configured to:

[0293] parse the subtask content, the execution area and the allocated agent in the task allocation table;

[0294] send a task start instruction containing the execution area to the allocated agent, control the allocated agent to move to the execution area and start the task program;

[0295] collect environmental dynamic data through the sensors carried by the allocated agent;

[0296] convert the environmental dynamic data into structured execution information;

[0297] label the state tag of the structured execution information;

[0298] upload the structured execution information with the state tag to a shared information platform in real time;

[0299] update the state of the work area in the shared information platform.

[0300] In an embodiment, the dynamic scheduling module 50 is specifically configured to:

[0301] monitor the movement trajectory of the agent in the shared information platform;

[0302] when detecting that the movement trajectories of two agents intersect at a distance less than a safety threshold, determine a path conflict event;

[0303] determine a task urgency of an agent causing the path conflict based on a work area status and a task progress in the shared information platform;

[0304] generate a passing order decision according to the task urgency;

[0305] control a low-urgency agent to pause movement based on the passing order decision;

[0306] plan an avoidance path for the paused agent and update a space occupation status of the shared information platform;

[0307] send a movement instruction containing the avoidance path to the paused agent when handling the path conflict event;

[0308] determine a device failure event when detecting that an energy level of an agent is lower than a failure threshold;

[0309] call an idle agent with a capability matching for a failed agent in the device failure event;

[0310] obtain a work progress parameter of a failure task from the shared information platform;

[0311] determine a remaining work amount parameter based on the work progress parameter;

[0312] send a new task start instruction containing the remaining work amount parameter to the idle agent when handling the device failure event;

[0313] generate an updated task allocation table based on the movement instruction and the new task start instruction;

[0314] issue the updated task allocation table to corresponding agents.

[0315] In an embodiment, the task acceptance module 60 is specifically configured to:

[0316] control a patrol device to scan a task execution area and collect environmental space data of the task execution area through the patrol device;

[0317] generate a three-dimensional environmental model based on the environmental space data;

[0318] obtain preset standard model data;

[0319] perform feature comparison between the three-dimensional environmental model and the preset standard model, and mark a difference area in the three-dimensional environmental model that does not meet a standard;

[0320] generate an acceptance report containing coordinates of the difference area, and store the acceptance report to an acceptance database.

[0321] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external user terminals via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a multi-agent cooperative task processing method on the server side.

[0322] In one embodiment, a computer device is provided, which may be a user terminal, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a multi-agent cooperative task processing method on the user side.

[0323] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0324] Obtain task instructions, perform semantic parsing on the task instructions to identify core task objectives, and decompose the task corresponding to the task instructions into multiple sub-tasks based on the core task objectives;

[0325] Obtain work environment information, divide multiple task regions based on the work environment information and determine the region weight of each task region, and combine the capability parameters of multiple agents and the region weights to generate a global collaboration framework;

[0326] Obtain real-time status information of multiple agents, combine the global collaboration framework with the real-time status information, match the optimal agent for each subtask, and generate a task allocation table.

[0327] distribute the task assignment table to the plurality of agents, control the agents to execute the sub-tasks, and upload task execution information collected in the execution process to a shared information platform;

[0328] monitor the execution process of the agents through the shared information platform, trigger a dynamic adjustment strategy to resolve conflicts or reassign tasks when detecting agent path conflicts or device failures, generate an updated task assignment table and re-distribute it;

[0329] After all sub-tasks are executed, obtain environment completion state data, and compare the environment completion state data with a preset standard model for acceptance.

[0330] In one embodiment, a computer-readable storage medium is provided, which stores a computer program that is executed by a processor to implement the following steps:

[0331] Obtain task instructions, perform semantic analysis on the task instructions to identify core task targets, and decompose tasks corresponding to the task instructions into a plurality of sub-tasks based on the core task targets;

[0332] Obtain work environment information, divide a plurality of task areas according to the work environment information and determine the area weight of each task area, combine the ability parameters of a plurality of agents and the area weight to generate a global collaboration framework;

[0333] Obtain real-time state information of a plurality of agents, combine the global collaboration framework and the real-time state information to match the optimal agent for each sub-task, and generate a task assignment table;

[0334] Distribute the task assignment table to the plurality of agents, control the agents to execute the sub-tasks, and upload task execution information collected in the execution process to a shared information platform;

[0335] Monitor the execution process of the agents through the shared information platform, trigger a dynamic adjustment strategy to resolve conflicts or reassign tasks when detecting agent path conflicts or device failures, generate an updated task assignment table and re-distribute it;

[0336] After all sub-tasks are executed, obtain environment completion state data, and compare the environment completion state data with a preset standard model for acceptance.

[0337] It should be noted that the functions or steps that the above computer-readable storage medium or computer device can implement can be referred to the related descriptions of the server side and the user side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0338] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0339] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is exemplified, and in actual application, the above-mentioned functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0340] It should be noted that if non-company software tools or components appear in the embodiments of the present application, they are only used for example introduction and do not represent actual use. The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for multi-agent cooperative task processing, characterized in that, Includes the following steps: Obtain task instructions, perform semantic parsing on the task instructions to identify core task objectives, and decompose the task corresponding to the task instructions into multiple sub-tasks based on the core task objectives; Obtain work environment information, divide multiple task regions based on the work environment information and determine the region weight of each task region, and combine the capability parameters of multiple agents and the region weights to generate a global collaboration framework; The process of acquiring work environment information, dividing the work environment information into multiple task regions and determining the region weight of each task region, and combining the capability parameters of multiple agents with the region weights to generate a global collaboration framework includes: Three-dimensional spatial data of the working environment are acquired through spatial scanning sensors; Identify environmental structural features based on the aforementioned three-dimensional spatial data; The environmental structural features specifically include spatial layout structure, size parameters of each area, distribution of obstacles, furniture location, usage frequency indicators, environmental cleanliness status, and functional classification. Multiple task areas are divided based on the aforementioned environmental structural characteristics; The process of dividing multiple task regions based on the environmental structural characteristics specifically includes: Divide the space into multiple mission areas based on spatial purpose and / or physical structure; Determine the area parameters and functional requirement parameters for each task area; The region weights are determined based on the area parameters and functional requirement parameters. The functional requirement parameters specifically include function usage frequency, regional accessibility coefficient, task complexity score, and cleaning target density; The area parameters, function usage frequency, regional accessibility coefficient, task complexity score, and cleaning target density are calculated using a weighting function to generate regional weights. Query the capability parameters of multiple agents from the agent capability database; The specific capability parameters include mechanical structure parameters, environmental adaptability indicators, operational accuracy level, task completion success rate, energy consumption efficiency, and operational time records. Based on the aforementioned capability parameters, a matching score between each agent and each sub-task category is calculated. The priority allocation order is generated by sorting the matching scores. A global collaboration framework is generated based on the aforementioned priority allocation order and regional weights; The generation of the global collaboration framework based on the priority allocation order and regional weights specifically includes: A distribution structure covering all task regions is generated using a greedy algorithm, a linear programming algorithm, and a graph matching algorithm. The allocation structure is analyzed to obtain the task priority sequence, task region affiliation, and expected execution time window for each agent; The task priority sequence, task region affiliation, and expected execution time window of each intelligent agent are integrated to generate a global collaborative framework; Obtain real-time status information of multiple agents, combine the global collaboration framework with the real-time status information, match the optimal agent for each subtask, and generate a task allocation table. The task allocation table is sent to the multiple intelligent agents to control the intelligent agents to execute the sub-tasks, and the task execution information collected during the execution process is uploaded to the shared information platform. The execution process of the intelligent agent is monitored through the shared information platform. When a path conflict or device failure is detected, a dynamic adjustment strategy is triggered to resolve the conflict or reassign the task, and an updated task allocation table is generated and redistributed. After all subtasks have been completed, environmental completion status data is obtained, and the environmental completion status data is compared and verified with a preset standard model.

2. The multi-agent cooperative task processing method as described in claim 1, characterized in that, Obtain task instructions, perform semantic parsing on the task instructions to identify core task objectives, and decompose the task corresponding to the task instructions into multiple sub-tasks based on the core task objectives, including: Receive task instructions via voice input; The task instructions are converted into structured text information using a speech recognition module; The structured text information is input into a pre-trained language model for semantic analysis, and the core semantic elements of the task instructions are extracted through the pre-trained language model. Identify core task objectives based on the aforementioned core semantic elements; The core task objective is broken down into multiple sub-task categories based on task type characteristics; Execution priority parameters are defined based on the regional association of subtask categories; Store the core task objectives, sub-task categories, and execution priority parameters in the task planning database.

3. The multi-agent cooperative task processing method as described in claim 1, characterized in that, Obtain real-time state information of multiple agents, combine the global collaboration framework with the real-time state information, match the optimal agent for each subtask, and generate a task allocation table, including: Associate each subtask with its corresponding execution region; Acquire real-time state information for each agent, including idle state, energy level, and current location; Determine the distance parameter between the current position of each agent and the subtask execution area; Extract the matching score between the agent and the subtask from the global collaboration framework; A comprehensive score is determined by combining the matching score, idle state, energy level, and distance parameters. For each subtask, select the idle agent with the highest overall score and generate a task allocation table containing the subtask, execution area, and assigned agent. The task allocation table is stored in the task allocation database.

4. The multi-agent cooperative task processing method as described in claim 1, characterized in that, The task allocation table is distributed to the multiple intelligent agents, controlling the agents to execute the sub-tasks, and the task execution information collected during the execution process is uploaded to the shared information platform, including: Parse the subtask content, execution area, and allocation agent in the task allocation table; Send a task start command containing the execution area to the allocation agent, and control the allocation agent to move to the execution area and start the task program; Dynamic environmental data is collected by assigning sensors to intelligent agents; The environmental dynamic data is converted into structured execution information; The status labels of the structured execution information are marked; Upload structured execution information with status tags to the shared information platform in real time; Update the work area status in the shared information platform.

5. The multi-agent cooperative task processing method as described in claim 1, characterized in that, The execution process of the intelligent agent is monitored through the shared information platform. When a path conflict or device failure is detected, a dynamic adjustment strategy is triggered to resolve the conflict or reassign tasks, generating an updated task allocation table and redistributing it, including: Monitor the movement trajectory of intelligent agents in the shared information platform; When the intersection distance of the movement trajectories of two intelligent agents is less than a safety threshold, it is determined to be a path conflict event; The task urgency of agents experiencing path conflicts is determined based on the work area status and task progress in the shared information platform. A passage order decision is generated based on the task urgency. Based on the passage order decision, low-urgency agents are controlled to suspend movement; Plan an avoidance path for the paused intelligent agent and update the space occupancy status of the shared information platform; When handling a path conflict event, a movement command containing the avoidance path is sent to the paused agent; When the energy level of the agent is detected to be lower than the fault threshold, it is determined to be a device malfunction event; For the faulty intelligent agent in the device failure event, retrieve the idle intelligent agent whose capabilities are matched; Obtain the work progress parameters of the faulty task from the shared information platform; Determine the remaining workload parameters based on the work progress parameters; When handling a device failure event, a new task start command containing the remaining workload parameters is sent to the idle agent; An updated task allocation table is generated based on the movement command and the new task initiation command; The updated task allocation table is then sent to the corresponding agents.

6. The multi-agent cooperative task processing method as described in claim 1, characterized in that, After all subtasks have been completed, acquire the environment completion status data, and compare and verify the environment completion status data with a preset standard model, including: The inspection equipment is controlled to scan the task execution area, and environmental spatial data of the task execution area is collected through the inspection equipment. A three-dimensional environment model is generated based on the environmental spatial data; Obtain preset standard model data; The 3D environment model is compared with a preset standard model in terms of features, and the difference areas in the 3D environment model that do not meet the standard are marked. An acceptance report containing the coordinates of the discrepancy areas is generated, and the acceptance report is stored in the acceptance database.

7. A multi-agent collaborative task processing device, characterized in that, The multi-agent collaborative task processing device includes: The task parsing module is used to obtain task instructions, perform semantic parsing on the task instructions to identify the core task objectives, and decompose the task corresponding to the task instructions into multiple sub-tasks based on the core task objectives. The task planning module is used to acquire work environment information, divide multiple task regions according to the work environment information and determine the region weight of each task region, and combine the capability parameters of multiple agents and the region weights to generate a global collaboration framework. The task allocation module is used to obtain the real-time status information of multiple agents, combine the global collaboration framework and the real-time status information, match the optimal agent for each subtask, and generate a task allocation table. The task execution module is used to send the task allocation table to the multiple intelligent agents, control the intelligent agents to execute the sub-tasks, and upload the task execution information collected during the execution process to the shared information platform. The dynamic scheduling module is used to monitor the execution process of the intelligent agent through the shared information platform. When a path conflict or device failure is detected, a dynamic adjustment strategy is triggered to resolve the conflict or reassign the task, and an updated task allocation table is generated and redistributed. The task acceptance module is used to obtain environmental completion status data after all sub-tasks have been completed, and to compare and accept the environmental completion status data with a preset standard model. The task planning module is specifically used for: Three-dimensional spatial data of the working environment are acquired through spatial scanning sensors; Identify environmental structural features based on the aforementioned three-dimensional spatial data; The environmental structural features specifically include spatial layout structure, size parameters of each area, distribution of obstacles, furniture location, usage frequency indicators, environmental cleanliness status, and functional classification. Multiple task areas are divided based on the aforementioned environmental structural characteristics; The process of dividing multiple task regions based on the environmental structural characteristics specifically includes: Divide the space into multiple mission areas based on spatial purpose and / or physical structure; Determine the area parameters and functional requirement parameters for each task area; The region weights are determined based on the area parameters and functional requirement parameters. The functional requirement parameters specifically include function usage frequency, regional accessibility coefficient, task complexity score, and cleaning target density; The area parameters, function usage frequency, regional accessibility coefficient, task complexity score, and cleaning target density are calculated using a weighting function to generate regional weights. Query the capability parameters of multiple agents from the agent capability database; The specific capability parameters include mechanical structure parameters, environmental adaptability indicators, operational accuracy level, task completion success rate, energy consumption efficiency, and operational time records. Based on the aforementioned capability parameters, a matching score between each agent and each sub-task category is calculated. The priority allocation order is generated by sorting the matching scores. A global collaboration framework is generated based on the aforementioned priority allocation order and regional weights; The generation of the global collaboration framework based on the priority allocation order and regional weights specifically includes: A distribution structure covering all task regions is generated using a greedy algorithm, a linear programming algorithm, and a graph matching algorithm. The allocation structure is analyzed to obtain the task priority sequence, task region affiliation, and expected execution time window for each agent; The task priority sequence, task region affiliation, and expected execution time window of each agent are integrated to generate a global collaborative framework.

8. A computer device, characterized in that, The computer device includes a memory, a processor, and a multi-agent cooperative task processing program stored in the memory and executable on the processor. When executed by the processor, the multi-agent cooperative task processing program implements the steps of the multi-agent cooperative task processing method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a multi-agent cooperative task processing program, which, when executed by a processor, implements the steps of the multi-agent cooperative task processing method as described in any one of claims 1-6.

Images