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

Abstract

The invention relates to the technical field of artificial intelligence, can be applied to service scenes such as pension service, financial science and technology and medical health, and discloses a multi-agent cooperative task processing method, device and equipment and a medium, and the method comprises the steps: obtaining a task instruction, analyzing a core target, and decomposing the core target into a plurality of subtasks; obtaining environment information, dividing task areas, and generating a cooperation framework in combination with agent capability and area weight; real-time states of the agents are obtained, and the optimal agents are matched based on the cooperation framework to generate a task allocation table; a task distribution table is issued to control the intelligent agent to execute the task and upload execution information; monitoring an execution process, and performing dynamic adjustment and updating a task allocation table when detecting path conflicts or equipment faults; and after the subtask is completed, obtaining environment completion state data, and comparing the data with a preset standard model for acceptance. According to the method, efficient task decomposition and intelligent distribution are realized by fusing task semantics, environment information and intelligent agent capability, and the cooperation stability is improved by introducing a real-time state perception and self-adaptive mechanism.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device and storage medium for multi-agent collaborative task processing. Background Technology

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

[0003] In elderly care services, robots are often used to assist with tasks such as home organization and cleaning. However, in existing systems, tasks are mostly manually set or triggered by predefined templates, lacking the ability to semantically analyze natural language task instructions. This makes it difficult to flexibly identify specific targets based on the actual verbal commands issued by the elderly. Furthermore, task allocation typically does not consider room structure and area characteristics, resulting in non-targeted task scheduling. The lack of an efficient capability alignment mechanism among multiple robots prevents optimal matching based on their current capabilities during task allocation, leading to some agents being overloaded while others remain idle, resulting in overall low efficiency.

[0004] In the healthcare sector, smart devices are increasingly being used in tasks such as ward rounds, environmental maintenance, and assisted logistics. Current technologies generally suffer from rudimentary modeling of healthcare environments, making it difficult to effectively divide task areas based on spatial distribution. Due to the failure to accurately match device capabilities with regional needs, high-capacity devices are often assigned low-complexity tasks, while low-capacity devices struggle to perform high-precision operations. Furthermore, real-time changes in device status are not promptly incorporated into the task scheduling mechanism, leading to delayed task allocation and disjointed collaboration, impacting the continuity and responsiveness of task execution.

[0005] In the fintech sector, intelligent service robots or multi-agent platforms are used in services such as intelligent branch navigation, customer reception, and back-end data processing. However, existing systems struggle to build stable and efficient task collaboration frameworks for time-sensitive and process-dependent tasks. Systems often fail to effectively utilize work environment data and regional resource characteristics for task decomposition, lacking mechanisms to dynamically match task requirements with equipment capabilities. This results in unscientific task allocation, low resource allocation efficiency, and consequently, impacts service quality and user experience. Summary of the Invention

[0006] The main objective of this invention is to provide a multi-agent collaborative task processing method, apparatus, device, and storage medium, aiming to solve the technical problems of existing multi-agent task processing methods that cannot perform joint modeling and dynamic task allocation based on task semantics, environmental structure, and agent capability status, resulting in low collaboration efficiency, poor resource utilization, and a lack of globally optimized task collaboration mechanisms.

[0007] To achieve the above objectives, the present invention provides a multi-agent cooperative task processing method, comprising:

[0008] 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;

[0009] 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;

[0010] 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.

[0011] 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.

[0012] 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.

[0013] 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.

[0014] Furthermore, to achieve the above objectives, the present invention provides a multi-agent collaborative task processing device, comprising:

[0015] 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.

[0016] 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.

[0017] 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.

[0018] 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.

[0019] 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.

[0020] 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.

[0021] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a multi-agent cooperative task processing program stored in the memory and executable on the processor, wherein when the multi-agent cooperative task processing program is executed by the processor, it implements the steps of the multi-agent cooperative task processing method as described above.

[0022] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a multi-agent cooperative task processing program, wherein the multi-agent cooperative task processing program, when executed by a processor, implements the steps of the multi-agent cooperative task processing method described above.

[0023] Beneficial Effects: This invention relates to the field of artificial intelligence technology and can be applied to business scenarios such as elderly care services, fintech, and healthcare. It discloses a method, apparatus, device, and medium for multi-agent collaborative task processing, including: acquiring task instructions and performing semantic parsing to identify core task objectives, decomposing the task into multiple sub-tasks; acquiring work environment information and dividing task regions, generating a global collaboration framework by combining agent capability parameters and region weights; acquiring real-time status information of agents, generating a task allocation table based on optimal agent matching within the global collaboration framework; issuing the task allocation table to control agents to execute tasks and uploading task execution information; monitoring the execution process based on a shared information platform, triggering a dynamic adjustment strategy when path conflicts or equipment failures are detected, generating and issuing an updated task allocation table; and acquiring environmental completion status data after sub-task execution and comparing it with a preset standard model for acceptance. This invention achieves efficient conversion of task instructions into multi-agent collaborative execution by combining task semantic information, environmental structure information, and agent capability parameters for unified modeling, and introduces a real-time state perception and dynamic allocation mechanism; it also improves the stability and reliability of system collaboration through real-time identification and adaptive adjustment of path conflicts and equipment failures. Attached Figure Description

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0025] Figure 1 This is a schematic diagram of an application environment for a multi-agent collaborative task processing method according to an embodiment of the present invention;

[0026] Figure 2 This is a flowchart illustrating an embodiment of the multi-agent cooperative task processing method of the present invention;

[0027] Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the multi-agent collaborative task processing device of the present invention;

[0028] Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0029] Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0030] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0031] The multi-agent cooperative task processing method provided in this invention can be applied to, for example... Figure 1In this application environment, the user terminal communicates with the server via a network. The server can obtain task instructions from the user terminal and perform semantic parsing to identify the core task objectives, decomposing the task into multiple sub-tasks; obtain working environment information and divide task areas, generating a global collaboration framework by combining agent capability parameters and area weights; obtain real-time status information of agents, and generate a task allocation table based on optimal agent matching according to the global collaboration framework; issue the task allocation table to control agents to execute tasks and upload task execution information; monitor the execution process based on a shared information platform, triggering dynamic adjustment strategies when path conflicts or device failures are detected, generating and issuing updated task allocation tables; and obtain environmental completion status data after sub-tasks are completed and compare it with a preset standard model for acceptance. This invention achieves efficient transformation of task instructions into multi-agent collaborative execution by combining task semantic information, environmental structure information, and agent capability parameters for unified modeling, and introduces real-time status awareness and dynamic allocation mechanisms; it improves the stability and reliability of system collaboration through real-time identification and adaptive adjustment of path conflicts and device failures. The user terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0032] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the multi-agent cooperative task processing method provided by the present invention. It should be noted that although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0033] like Figure 2 As shown, the multi-agent cooperative task processing method proposed in this invention includes the following steps:

[0034] S10, 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;

[0035] In this embodiment, the task instruction acquisition operation typically targets the human-computer interaction input channel, with input formats including but not limited to multimodal input data such as voice, text, and images. The task instructions can originate from human service requests, scheduling information from an autonomous scheduling platform, or task content returned by external system interfaces. For example, in a home service scenario, a user can issue a command like "Please tidy the bedroom" via a voice assistant, or send a list-style text task via a mobile terminal. In financial or medical scenarios, task information can also come from event requests pushed by a business scheduling platform according to rules. Upon receiving the task instruction, the system needs to standardize and translate it, converting the raw task data into a structured format for subsequent processing modules to use. Structured processing includes operations such as unified semantic identification, keyword extraction, standardizing task entity boundaries, and eliminating ambiguities. During the translation process, a word segmentation engine, named entity recognition module, and rule engine can be used to assist in grammatical parsing. Finally, the task content is represented in a unified structure as multiple fields, such as task verbs, target entities, execution objects, and environmental constraints.

[0036] After the instruction format conversion is completed, the semantic parsing stage begins. The goal of this stage is to extract the implicit behavioral intentions and operational objectives within the task. Semantic parsing can leverage pre-trained language models for contextual understanding and combine this with a domain-defined task ontology for matching and reasoning. The language model embeds the instruction text into a vector space semantic representation and then identifies the core action type and target object using a classifier or matching rules. For example, in "tidy up the living room," "tidy up" is identified as the action, and "living room" as the target area. Semantic parsing can also include contextual relevance assessment, such as whether the location of the "desk" mentioned in the preceding instructions affects the current "cleaning" objective, or determining whether there are complex semantic relationships such as time, sequence, or constraints. The semantic tags extracted through these parsing methods can include, but are not limited to, task objectives, operational constraints, priority constraints, and task triggering conditions.

[0037] After identifying the core task objective, it needs to be broken down into multiple subtasks. The decomposition process is based on a hierarchical task type structure defined in the knowledge base, and automatically matches subtask types with the environmental structure and existing task graph. For example, "tidying the living room" can be broken down into three subtasks: "cleaning the coffee table," "putting the remote control back in place," and "sweeping." Task decomposition can also be based on the distribution of entities within the target area. For example, based on spatial layout information, "cleaning the kitchen" can be broken down into sub-objectives such as "cleaning the stove," "cleaning the floor," and "putting kitchen utensils back in place." During the decomposition process, pre-defined task operation trees, task dependency graphs, and device reachability models can be referenced to generate subtasks. Each subtask is tightly bound to subsequent capability matching and task scheduling processes. Task decomposition is not limited to parallel subtasks; it can also handle special task forms such as sequentially dependent tasks, conditionally triggered tasks, and mutually exclusive tasks. By defining the execution order, conditional statements, or execution constraints in the subtask description, the subtasks possess complete schedulability and executability.

[0038] Task instructions can be obtained through input terminals equipped with natural language interaction interfaces, such as edge computing devices with embedded speech recognition chips. These devices can listen to and receive human language instructions in real time, converting the speech signal into a text stream. The converted text is then input into a natural language processing engine via a semantic recognition module. A language model with an encoder-decoder structure is used to extract semantic units, and semantic slots are filled using rule templates to further extract structured targets. For example, upon receiving the instruction "Help me tidy up the living room," the system automatically extracts "tidy up" as the action, "living room" as the target entity, and "help me" as the service request type, storing the structural information in a standard task format.

[0039] The semantic parsing of task objectives can be accomplished by integrating context-aware language models such as BERT with a knowledge graph-based semantic matching engine. For complex synthesized instructions, dependency parsing is first used to perform preliminary semantic analysis, and then a domain semantic database is used to determine the actual meaning of each task unit and eliminate ambiguity. For example, in "Please tidy up the desk and the floor," the desk needs to have its drawers and tabletop tidied, and the floor needs to be cleaned; therefore, the parser splits it into two different types of task objectives.

[0040] During task decomposition, a mapping graph between task types and environmental objects can be constructed, and a set of subtasks can be automatically generated based on the mapping relationship. The task decomposition algorithm can adopt a hierarchical task tree structure for modeling, where the root node represents the overall task objective, child nodes are divided according to operation dimensions or target regions, and leaf nodes represent specific execution instructions. Based on the identified target region characteristics, the system refines the task to the spatial entity granularity, generating a set of subtasks that can be directly assigned to specific execution units. In addition, after subtask decomposition, each task is assigned a logical execution identifier, including task priority, parallel flag, preset execution window, and other information, providing a basis for subsequent allocation and scheduling processes.

[0041] Example Explanation: In a smart elderly care service scenario, when an elderly user requests "Please tidy up my room" via a voice assistant, the system receives the task instruction and performs semantic analysis, identifying "tidy up" as the action and "room" as the target area. Based on semantic relationships, the analysis system further confirms the multiple task objectives implicit in the "tidy up the room" instruction, such as "making the bed," "putting items back in their place," and "cleaning the floor." The task decomposition engine, based on room layout information, furniture and item distribution diagrams, and historical service data, refines the task into multiple executable sub-tasks, such as "moving clutter from the floor to the storage cabinet," "cleaning the bedside table surface with a wet wipe," and "arranging the bedding neatly." Each sub-task carries specific spatial positioning information and action instructions, which are then assigned by the subsequent scheduling engine to service robots with corresponding capabilities. For example, robots with strong fine motor skills might put items back in their place, while robots with high cleaning efficiency might handle floor cleaning. This approach not only improves the rationality of task allocation and the accuracy of execution, but also significantly reduces the risk of repetitive tasks and missed areas, thereby providing elderly users with a safer, more comfortable, and dignified living experience.

[0042] In the healthcare field, nursing instructions such as "Please help the patient make the bed and clean the ward" can be semantically analyzed to identify "making the bed" and "cleaning the ward" as two independent objectives, which are further broken down into sub-tasks such as "straightening the pillows," "smoothing the sheets," and "cleaning the floor." These sub-tasks can be precisely issued based on the ward's spatial structure and the capabilities of intelligent nursing equipment, enabling efficient and collaborative nursing operations.

[0043] In the fintech business, work order tasks in the customer service system, such as "review customer points for this month and clear abnormal records," can be parsed to extract two task objectives: "review points" and "clear records." Based on the data table structure and processing logic, the system automatically breaks down the task into sub-tasks such as "query customer points data," "identify points anomalies," and "perform points repair," and then schedules the corresponding data processing engine and log cleanup service module to complete the task collaboratively.

[0044] This embodiment employs a semantic analysis-based task recognition mechanism, enabling task instructions to be accurately parsed into structured task objectives, avoiding the one-sidedness and ambiguity of traditional rule-based matching methods in task understanding. Further decomposing the parsed task results into multiple sub-tasks with clearly defined operational objects and execution constraints helps improve the accuracy and flexibility of task scheduling. This operational chain avoids overlapping responsibilities and unclear division of labor during task execution, improving the clarity, response speed, and scalability of the overall system's task processing.

[0045] S20: Obtain working environment information, divide multiple task areas according to the working environment information and determine the area weight of each task area, and combine the capability parameters of multiple agents and the area weights to generate a global collaboration framework.

[0046] In this embodiment, acquiring work environment information typically relies on devices with spatial awareness capabilities to perform a comprehensive scan of the service area. Environmental information includes, but is not limited to, spatial layout structure, dimensional parameters of each area, obstacle distribution, furniture location, usage frequency indicators, environmental cleanliness status, and functional classification. This information can be collected using LiDAR, depth cameras, inertial navigation devices, or infrared sensors to generate a structured spatial data model. This model digitally represents the geometric composition and functional distribution of the target environment, providing support for subsequent operations.

[0047] After establishing the environmental space model, the workspace needs to be logically divided. The division can be based on the space's purpose (such as living room, kitchen, bedroom, etc.) or can be automatically segmented by combining physical structural boundaries (such as doors, walls, furniture placement lines). Each region will be treated as an independent task processing unit, and its boundaries, functional labels, and processing difficulty will be recorded in the task region definition dataset.

[0048] The determination of region weights is based on a comprehensive calculation of parameters from multiple dimensions. The primary dimension is the region area, which determines the task intensity. Secondly, the frequency of function usage can be extracted from historical behavior logs to determine the usage frequency of each region; for example, kitchens and bathrooms have higher priority in cleaning tasks. Other dimensions include region accessibility coefficients, task complexity scores, and cleaning target density. These indicators are then used to calculate a normalized region weight using a weighting function. This weight determines the region's priority and resource allocation ratio in collaborative scheduling.

[0049] Acquiring capability parameters for multiple agents typically relies on a unified capability parameter database. These parameters include mechanical structural parameters (such as reach, load, and degrees of freedom), environmental adaptability indicators (such as terrain passability), operational accuracy level, task completion success rate, energy efficiency, and operational duration records. These capability parameters can be acquired and updated based on manufacturing specifications, system initialization calibration, historical execution data, and other sources. All capability parameters are encoded into structured vectors for processing by the agent task matching model.

[0050] The key to constructing a global collaborative framework lies in the synergistic integration of environment segmentation results with agent capability parameters. The system establishes capability matching scoring relationships between all task regions and agents by constructing a multi-objective optimization matching matrix. The scoring mechanism combines the required capability vector of a region with the agent's capability vector for similarity measurement and further considers the regulatory effect of region weights on task resource allocation. After completing all matching degree evaluations, an allocation structure covering all task regions can be generated using greedy algorithms, linear programming algorithms, graph matching algorithms, etc. This structure defines the task priority sequence, task region affiliation, and expected execution time window for each agent, ultimately forming a global collaborative framework for controlling execution.

[0051] Environmental information can be collected through a multimodal sensor network fixedly installed on the ceiling or walls, or through inspection by an autonomously navigated mobile robot equipped with sensors. When dividing open spaces into zones, convolutional neural networks can be used to perform functional semantic segmentation of spatial images, thereby achieving automatic spatial function labeling. When dividing complex environments, BIM models or pre-stored CAD drawings can be used to assist in spatial modeling. Regarding zone weight calculation, the coefficients of the parameter weights can be set empirically or optimized through regression analysis of the data-driven model in historical task executions.

[0052] Capability parameters can be managed using a centralized capability database maintenance model, where each agent's capability parameters are dynamically updated after the task is completed. If adaptation to highly dynamic task environments is required, task adaptability feedback metrics can be temporarily retrieved before the collaboration framework is built to achieve real-time capability assessment. The task matching scoring matrix can use cosine similarity as the basic scoring function, or a task adaptation loss function can be introduced for back-evaluation. The output structure of the collaboration framework supports embedding into a real-time scheduling module, enabling continuous correction of the task allocation table.

[0053] Example Description: In the field of elderly care services, the system first uses sensors to comprehensively scan the living space of the elderly, identifying typical living areas such as bedrooms, living rooms, and kitchens. It then analyzes the usage frequency and functional intensity of each area based on the elderly's daily activity data. The system designates the bedroom as a high-weight area, focusing on tidying tasks, especially before and after the elderly's afternoon nap. The agent's capability parameters indicate that robot A1 excels at tidying tasks, while robot A2 is more efficient at floor cleaning. Based on this matching logic, the global collaboration framework prioritizes assigning bedroom tidying tasks to A1 and living room cleaning tasks to A2, effectively improving task completion quality and reducing the elderly's discomfort.

[0054] In the healthcare sector, the system is deployed in medical institution logistics management scenarios, enabling multiple intelligent agents (cleaning, handling, and delivery) to collaboratively handle tasks in the medical environment. Environmental modeling identifies functional areas such as operating rooms, wards, and corridors. The system assigns different weights based on the area's infection risk level, operational frequency, and spatial complexity. For example, operating rooms require frequent disinfection, while wards require targeted medication delivery to specific beds. Capability parameters show that agent B1 possesses high-precision positioning and steady-state navigation capabilities, making it suitable for equipment organization tasks in operating rooms; agent B2 is suitable for cleaning tasks in corridors. This collaborative framework, achieved through joint modeling of weights and capabilities, reduces repetitive entry and exit paths in high-infection-risk areas, improving overall operational efficiency and infection control capabilities.

[0055] In the fintech business, the system is applied to environmental maintenance tasks in large data centers, such as the cleaning and maintenance of critical areas like monitoring rooms, server rooms, and power modules. The system divides cleaning areas and assigns weights based on indicators such as equipment distribution density, ventilation requirements, and the complexity of ground cables. The C1 robot possesses electromagnetic interference resistance and adaptability to complex terrain, while the C2 robot has efficient dust removal and obstacle avoidance capabilities. Based on the matching relationship between capabilities and task requirements, the collaborative framework assigns server room cable management tasks to C1, while C2 is responsible for tasks such as dust removal from air conditioning vents. This allocation method maximizes the efficiency of the intelligent agents and improves the stability and security of operations in highly sensitive areas without affecting data center operations.

[0056] This embodiment effectively improves the distinguishability and allocation accuracy of task units and enhances the adaptability of resource scheduling by structuring the working environment and introducing a task region division and weight evaluation mechanism. Establishing a task matching mechanism based on capability parameters avoids mismatches between tasks and executors, improving agent utilization efficiency and task execution quality. The resulting collaborative framework has strong interpretability and adjustability, providing a data foundation for dynamic scheduling.

[0057] S30: Obtain real-time status information of multiple agents, combine the global cooperation framework and the real-time status information, match the optimal agent for each subtask, and generate a task allocation table;

[0058] In this embodiment, acquiring real-time state information of multiple agents is a fundamental step in realizing the dynamic scheduling mechanism. Its essence lies in capturing the instantaneous capability boundaries and action constraints of each executing entity in the collaborative system. The state of each agent is not merely whether it is currently idle, but a real-time mapping of a multi-dimensional set of attributes, including but not limited to location information, motion state, posture parameters, current task context, system health indicators, remaining energy level, task execution accuracy, module load rate, sensor feedback stability, path reachability, and network communication status. This state information is collected through the agent's local perception system and task control unit, and synchronized to the scheduling center or edge control cluster via a low-latency link by an embedded communication module, forming a real-time state vector matrix based on timestamps. This matrix needs to maintain consistency and availability throughout the scheduling cycle; therefore, it requires a combination of a state sampling window mechanism and data integrity fault tolerance strategies, such as using Kalman filtering to predict and complete intermittent states, or enabling redundant channels to improve communication robustness in high-interference environments.

[0059] Real-time status information is not processed in isolation but must be integrated into the global collaboration framework built in the previous stage. This global collaboration framework establishes an initial matching relationship between agents and subtasks based on task region features, capability parameters, and region weights. Its output is a static priority allocation structure, which only has strategic reference value in real-world dynamic environments. Therefore, during scheduling and execution, this matching structure needs to be adapted and adjusted using real-time status information.

[0060] During the matching process, each subtask is first bound to its corresponding task region, which can be defined by spatial segmentation algorithms, physical boundary annotations, or semantic label mappings. Then, an initial matching score between each subtask and all candidate agents is extracted from the global collaboration framework as a static baseline. Next, the influence coefficient of each candidate agent's current state on the subtask's completion efficiency and success rate is dynamically calculated, including but not limited to the following dimensions:

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

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

[0063] Energy status: Is the remaining battery power greater than the minimum execution requirements of the task (battery power should be reserved for safe return)?

[0064] Task congestion factor: Whether there are any bound tasks causing queuing, delays, or regional conflicts;

[0065] Resource availability: Does the current agent possess the necessary tools, modules, or interface capabilities to complete the task?

[0066] Communication reachability: Are there any communication blind spots in the task area that affect the real-time issuance of task instructions or status reporting?

[0067] Execution history stability: statistical data on the success rate, completion time, and reasons for failure of similar tasks.

[0068] These factors can construct a multidimensional state vector, which is input into a weighted scoring model or a trained scheduling scoring network model (such as a gradient boosting tree or attention neural network) to output the current matching score. This score is the priority estimate for the agent to complete the subtask at the current moment.

[0069] To prevent overfitting or skewed allocation, the system can introduce random perturbation terms (such as softmax temperature control) to achieve scheduling diversity during matching, 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 prioritizes subtasks and selects an execution agent from the set of idle agents with the highest matching scores for each subtask. Resource conflict detection is required during the matching and allocation process to prevent multiple tasks from being bound to the same agent or a single agent from being overloaded. The task allocation table should be a structured data carrier, containing fields such as task identifier, task region, execution time limit, scheduling source, assigning agent identifier, expected state change, and task completion threshold, for use in subsequent task issuance and state tracking.

[0071] Under different task densities, allocation strategies can adopt different granularities and distribution characteristics. For example, when the task density is low and the agent has abundant resources, a greedy scheduling algorithm can be used to quickly allocate resources; when the task density is high and resources are scarce, near-global optimization algorithms such as genetic algorithms, reinforcement learning policy networks, and simulated annealing algorithms can be used to improve the quality of scheduling solutions.

[0072] The task matching model can incorporate an adaptive weighting mechanism based on different task types. For example, in material handling tasks, the agent's load capacity and path stability are dominant; in delicate operation tasks, the robotic arm's accuracy and stability scores should be the core scoring factors. The task matching model should support plug-and-play strategy switching to adapt to different business logics.

[0073] The periodicity of real-time status acquisition can be dynamically adjusted according to the urgency of the task. Agents in important areas or on critical paths need to report status more frequently, while the sampling frequency can be appropriately reduced in peripheral areas to save communication bandwidth. The status anomaly detection mechanism should also run in real time. When a status breakpoint, parameter mutation, or logical conflict is detected (such as executing a task but the idle flag is true), the status verification mechanism and task reconstruction logic should be triggered to improve system robustness.

[0074] The task assignment table can be designed as a two-layer structure: the first layer is a task-dimensional structure, which includes subtask IDs and assignment targets; the second layer is an agent-dimensional structure, which identifies the current binding status of all tasks, facilitating subsequent conflict detection and anomaly recovery.

[0075] Example Description: In an elderly care scenario, multiple service robots need to complete the task of tidying up and organizing a room. In one task, robot A1 was originally assigned to tidy up the elderly person's bedroom, but the system detected that its current location was too far from the bedroom and its battery was only 15%, insufficient to complete the entire task. At the same time, robot A4 had just finished cleaning the bathroom, was currently idle, and was only 2 meters away from the bedroom. The scheduling system, by calculating its path cost, battery redundancy, and historical tidying accuracy score, decided to temporarily switch the bedroom tidying task to A4, ensuring uninterrupted service and optimal resource utilization.

[0076] In the healthcare field, cleaning robot teams are required to perform alternating disinfection tasks in patient ward corridors and isolation wards. Due to a sudden outbreak alert, the disinfection of high-risk infection areas needs to be accelerated. By acquiring the current task status and regional path traffic conditions of all agents, the system discovered that robot B2, originally assigned to disinfection, was stuck in a path blockage, while robot B5 had completed its task in an adjacent area and possessed high-protection module capabilities. B5 was immediately dispatched to perform the disinfection task, shortening the response delay and ensuring the task was completed on time.

[0077] In a financial data center environment, inspection robots need to identify line anomalies in multiple server rack areas. The scheduling system detected in real-time that robot C1 needed to be urgently removed for maintenance due to an abnormal temperature sensor, and its task could not continue. Meanwhile, robot C3 had just completed an adjacent task and had a stable image recognition module and low interference tolerance. Therefore, the task was immediately switched from C1 to C3 to ensure that the high-priority inspection task was not interrupted and to avoid passive blind spots in important data links.

[0078] This embodiment incorporates real-time state information from multiple agents into the task scheduling logic, dynamically perceiving the load, path, resource, and energy efficiency boundaries of the task execution system. This allows task allocation to accurately reflect the optimal recipient for each subtask, improving task completion efficiency, scheduling response speed, and resource utilization. When faced with dynamic changes in agent states (such as battery depletion, task congestion, or temporary communication interruptions), the system possesses adaptive reallocation capabilities, effectively reducing task failure rates and overall coordination costs due to incorrect matching. The mechanism of real-time state participation in task matching transforms scheduling decisions from static presets to highly flexible responses, adapting to highly dynamic, multi-perturbation, and multi-objective collaborative task execution scenarios.

[0079] S40, 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;

[0080] In this embodiment, distributing the task allocation table to multiple agents relies on a synchronization and coordination mechanism between the task scheduling module and the communication module. The task allocation table is the carrier of structured scheduling output, containing multiple fields such as subtask identifier, task region, allocation object identifier, execution time limit, task priority, task type label, and parameter template. This table is transmitted via a network scheduling bus in wired or wireless mode, typically using MQTT, ROS communication protocols, or custom low-latency data protocols for task push. Before distributing the task allocation table, the scheduling system must perform table structure validity checks and content consistency checks to ensure that the task objective, target region, and agent capabilities are matched consistently, avoiding execution interruptions or conflicts caused by illegal instructions.

[0081] After receiving the task allocation table, the agent parses the task content through the local scheduling interface, extracts the target task area, execution target, and task behavior chain template, and generates an action path and task behavior sequence according to a preset control strategy. Under the scheduling of the control module, the agent executes the sub-task process one by one according to the task behavior sequence. The task control process may involve sub-stages such as path planning, module wake-up, action execution, status acquisition, anomaly perception, and feedback response. The behavior template can load different subroutines according to the task type, such as scheduling the chassis drive module and vacuuming module for cleaning tasks, and scheduling the robotic arm module and image recognition module for tidying tasks.

[0082] During task execution, the intelligent agent needs to continuously collect task execution information, including current location changes, target object processing status, sensor data, execution phase completion markers, error status codes, and module load. The raw data collection format is unstructured, which needs to be converted into a structured format by a local data processing unit, annotating the information type, timestamp, task context, and execution phase labels. Some tasks also require real-time environmental data, such as temperature, humidity, electromagnetic interference intensity, and space occupancy.

[0083] After processing, structured task execution information needs to be uploaded to the shared information platform. The shared information platform serves as the public information interaction center for the multi-agent collaborative system, supporting unified storage, real-time updates, and access authorization for task information, status information, and spatial information. Information uploading employs a multi-channel parallel mechanism to ensure high-frequency, low-latency communication requirements. The system optimizes bandwidth usage through a block-based upload strategy or streaming compression protocol and supports data reporting receipts and anomaly retransmission mechanisms to prevent information loss or mistransmission. Upon receiving task execution information, the shared information platform automatically maps it to the current task space status and updates the corresponding task progress, region completion, and anomaly flag fields for subsequent monitoring and scheduling modules to access.

[0084] The task assignment table can use different field templates depending on the type of task. For example, for a structured space cleaning task, the task table includes the cleaning path point set, action mode parameters, area risk level, and expected completion accuracy level; for an item sorting task, it includes item identification tags, target placement location, operation grasping posture template, and maximum number of retries.

[0085] The generation of task behavior sequences can rely on rule template libraries or policy network models. For example, in elderly care scenarios, a tidying task template includes a five-stage process: "identifying the target → reaching out to grasp → moving to the correct location → taking a photo for verification → marking as complete." The task execution control process is driven by a state machine mechanism, and each stage requires a success flag to proceed to the next stage.

[0086] The task execution information upload strategy supports two modes: periodic upload and event-triggered upload. Periodic upload is suitable for continuous tasks (such as inspection and cleaning), while event upload is suitable for result-oriented tasks (such as switch control and grasping operations). For agents performing tasks in highly dynamic and complex environments, the upload mechanism can also be combined with frame rate compression and anomaly detection mechanisms to trigger information reporting only in key frames or abnormal frames, thereby reducing the platform's data load.

[0087] In the shared information platform, execution information can be stored in hash groups based on task identifiers, supporting task backtracking and visualization of execution trajectories. Simultaneously, the platform supports synchronization interfaces with higher-level management platforms, pushing task execution status to the monitoring end or user interface, achieving transparent task management.

[0088] Example Description: In an elderly care service scenario, home robot A2 is assigned to tidy up the bedroom. The system sends it a task assignment table, including the target area "master bedroom," the task type "item return," and the behavior template "extend arm → recognize → move → place," while also specifying the task completion time limit and environmental interference level. After parsing, A2 begins executing the task. Upon recognizing scattered clothing, it performs a grasping action and moves it to the wardrobe area, recording the state changes and sensor feedback at each stage of the action. After tidying up, it uploads the structured task execution information to a shared information platform, allowing other robots to determine whether the bedroom tidying is complete and avoiding duplicate scheduling.

[0089] In the healthcare field, after receiving a task assignment sheet, disinfection robots perform disinfection tasks according to preset area paths. The system continuously collects data on their movement trajectory, spray release status, and air quality changes, and uploads this information in real time to a shared information platform according to area location tags. This allows the backend monitoring system to remotely monitor the quality of task execution and environmental improvement.

[0090] In financial data centers, inspection robots receive task assignment sheets and proceed to designated server rack areas to perform line inspections. During each inspection, they record key data such as cable connection status images, power indicator light status, and ambient temperature and humidity information, uploading this data in real time to a shared information platform. The backend system uses this task execution information to determine if there are any abnormal fluctuations or if manual intervention is required, enabling remote and efficient management of the inspection process.

[0091] This embodiment achieves high-concurrency and precise multi-task scheduling by issuing a task allocation table and having agents execute tasks according to a behavioral sequence. Task execution information is uploaded to a shared information platform in real time, enabling the scheduling system to comprehensively grasp the execution status of each task, the current operating status of the agents, and environmental change data. This mechanism enhances the perception closed-loop capability and execution stability of the multi-agent system, helps to promptly detect task anomalies and trigger corrective measures, and improves overall collaboration efficiency and execution reliability. The unified collection of data on the shared information platform also provides complete data support for subsequent task optimization and capability learning, enhancing the system's adaptability and evolutionary capabilities.

[0092] S50, 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.

[0093] In this embodiment, monitoring the execution process of multiple intelligent agents relies on the continuous data access capabilities and unified state modeling mechanism of the shared information platform. The shared information platform continuously receives and integrates the task execution information uploaded by each intelligent agent, and constructs a global visual operation map by combining the dynamic state of the environment and spatial location data. This map records key operational parameters of all intelligent agents, such as spatial location, movement trajectory, current task progress, module operation status, and resource consumption status.

[0094] Path conflict detection is based on spatiotemporal cross-analysis of multi-agent spatial trajectories. In the shared information platform, the system performs overlap analysis by comparing the predicted future trajectories of multiple agents with their currently occupied paths. If any two trajectories intersect within a future window and the distance between the intersection point is below a preset safety threshold, it is marked as a path conflict event. Predicted trajectories can be obtained through a trajectory inference model based on historical motion states or through future travel paths provided by the local planning modules of each agent. Once a path conflict event is identified, the system immediately freezes the motion state of low-priority agents, preventing them from entering the conflict area, and initiates the conflict resolution process.

[0095] Equipment fault detection is based on anomaly codes in task execution information, sensor self-test feedback, module operating status, and energy level indicators. If an agent's uploaded status information shows abnormal action responses, execution stagnation, or energy levels below a set threshold, it is determined to be a device fault event. The shared information platform combines historical task execution models and a capability tag library to identify the scope of the fault and the point of task interruption.

[0096] The dynamic adjustment strategy is divided into two paths: conflict resolution paths and task reallocation paths. For path conflict events, the system first assesses the current task urgency of each agent involved in the conflict. Urgency calculation considers factors such as task priority, remaining time, and spatial travel costs. Agents with low urgency are controlled to pause movement and trigger a path replanning process, generating an avoidance path to bypass the conflict area. The avoidance path is generated by the path planning engine in the global map state, optimizing path costs and latency metrics. The result is written into the movement command and awaits issuance.

[0097] In the event of equipment failure, the system identifies the currently interrupted task and extracts its remaining workload parameters. Simultaneously, it retrieves idle agents with a high degree of matching degree from the capability database. Matching criteria include historical task completion capabilities, geographical location, energy level, and module health status. The selected replacement agent receives the task takeover instruction and registers its new task status on 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 healthcare field, when cleaning robots are operating in ward areas, their path predictions may conflict with those of supply robots. The system avoids collisions by judging the passage order and replanning the path, ensuring unobstructed access for medical equipment. If a cleaning robot experiences a cleaning module failure, the system automatically switches the task to a standby robot and issues an updated task allocation table to complete the task continuation.

[0105] In the financial sector, when a data center inspection robot exits its task due to sensor malfunction during a cabinet inspection, the system immediately dispatches a replacement robot to take over the task based on the task history and capabilities, regenerates the task list, and notifies the management platform to avoid missing critical task nodes and ensure the continuity of the operation and maintenance process.

[0106] This embodiment achieves real-time monitoring of the agent execution process through a shared information platform. It can promptly identify path conflicts and equipment failures during task execution, ensuring the stability and continuity of the multi-agent system's operation. Path conflict events can be effectively prevented from causing physical interference between agents through real-time trajectory prediction and dynamic path adjustment mechanisms, improving task execution safety and spatial resource utilization efficiency. Equipment failure events can be promptly handled through a task reassignment process, ensuring uninterrupted task operation. The dynamic updating and synchronous distribution mechanism of the task allocation table ensures that the entire system maintains the consistency and integrity of overall task coordination even in the event of local anomalies.

[0107] S60. After all subtasks have been completed, obtain the environmental completion status data and compare and verify the environmental completion status data with the preset standard model.

[0108] In this embodiment, after completing all sub-tasks, it is necessary to evaluate the actual changes in the environmental state caused by the task execution, and judge the quality of task completion based on this evaluation result. The environmental completion state data represents a structured expression of the current state of the target area after task execution, reflecting quantifiable information such as spatial layout, object location, cleanliness, and obstacle arrangement. This data is acquired by the agent after execution, triggering a final state acquisition process. It typically combines vision, LiDAR, depth camera, or multimodal sensor scanning of the task execution area, and filters, fuses, and reconstructs the scanned data to generate representations such as 3D point clouds, raster maps, or RGB-D images.

[0109] To ensure comparability, the completed environmental state data needs to be transformed into a unified 3D environment model through data structure standardization and feature extraction algorithms. Feature extraction includes steps such as geometry-based shape reconstruction, semantic-based object label recognition, and topology-based spatial partitioning modeling. The 3D environment model structure should possess high resolution, spatial continuity, and semantic integrity, enabling fine-grained comparisons with reference models.

[0110] A pre-defined standard model indicates the expected state of the target after the task is completed, serving as the criterion for acceptance judgment. This model can be obtained through expert rule setting, sampling of historical best task results, or manual modeling. Its data structure is consistent with the 3D environment model, ensuring that point-to-point, block-to-block, or semantic tag-to-semantic tag correspondence matching can be achieved during the comparison process.

[0111] The alignment process is based on mechanisms such as spatial reconstruction registration, semantic overlap matching, and error calculation. Spatial registration can use the Iterative Closest Point (ICP) algorithm, global feature matching algorithm, or deep learning-driven feature pairing network to align the current 3D environment model to the standard model coordinate system. Semantic matching identifies the offset, missing points, and misplacement of key objects, thereby evaluating the deviation from task completion. Error calculation can output metrics such as global offset metric, key point error range, and semantic matching rate to support acceptance decisions.

[0112] The final comparison results are used to generate acceptance judgment labels. If the error is within the set threshold range, the subtask is considered qualified; if there are abnormal areas exceeding the allowable deviation, these areas are marked as difference areas, and information such as specific spatial coordinates, difference type, and impact level is output. The comparison and judgment results are uniformly written into the acceptance report, which structurally represents the task achievement status and provides data basis for subsequent result archiving, rework scheduling, and system optimization.

[0113] Environmental status data can be collected by an inspection robot equipped with an RGB-D camera. The image data is used to generate a dense point cloud through a 3D reconstruction algorithm, and a consistent 3D model is constructed using spatial meshing. Object recognition can be performed using deep learning object detection algorithms such as YOLO or Mask R-CNN to extract controllable objects in the task area, such as furniture, miscellaneous items, medical devices, or equipment terminals.

[0114] The preset standard model can be derived from different data sources. In scenarios where tasks are executed multiple times, historical task completion samples can be aggregated to generate diverse reference templates. For tasks requiring high precision, engineers can manually configure the 3D spatial structure, object placement status, and environmental cleanliness scoring model to form an expert rule-defined standard model. During the comparison process, a hierarchical error assessment mechanism is introduced. Coarse-layer comparison is used for structural contour checks, fine-layer comparison locates key object offsets, and semantic-layer comparison assesses object category errors or omissions. For example, if the target is "desktop cleaning," the system compares the changes in desktop area occupancy, the types and quantities of remaining items, and whether they are consistent with the expected model. The system will crop and label areas with comparison errors exceeding a threshold, forming a difference area output. The acceptance results are output in the form of a structured report, supporting graphical visual viewing and structured database storage. In addition to error indicators, the report can also include confidence scores, task allocation record indexes, and suggested rework paths.

[0115] After the acceptance and comparison of the task execution results, the system further conducts a comprehensive analysis of the environmental completion status data and task execution information obtained from the comparison to achieve continuous optimization of the capability cognition and scheduling mechanism. Environmental completion status data represents the static spatial results after task completion, reflecting the cleanliness, structural restoration, and functional achievement of the target area. This data originates from environmental scanning records of the inspection agent or terminal equipment after task completion and is typically represented in the form of 3D point clouds, depth images, or semantic models, and is uniformly stored in the acceptance database. Task execution information comes from the dynamic data stream generated throughout the execution process, including the start and end times of each agent's work, path trajectory, state changes, execution anomalies, sensor inputs, and system responses, which are collected and archived in real time by the shared information platform.

[0116] Based on environmental completion status data, structured evaluation indicators oriented towards spatial status can be constructed, such as spatial coverage integrity, residual area area, object positioning accuracy, and interference area frequency, to quantify the quality of results achieved. Based on task execution information, efficiency evaluation indicators oriented towards process behavior can be constructed, such as task load per unit time, path length optimization ratio, energy consumption to completion ratio, human intervention rate, and interruption frequency, to characterize task execution efficiency and stability. These two types of evaluation indicators together constitute the task performance vector, reflecting the entire process performance of the task from process to result.

[0117] The above metrics are input into the machine learning model to establish a mapping relationship between the evaluation metrics, capability parameters, and scheduling weights. Capability parameters describe the 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. Task allocation weights determine the degree of influence of each capability factor in the scheduling function, thereby affecting future task allocation results.

[0118] The goals of model training and analysis include two dimensions: first, to dynamically adjust the agent's capability parameters based on existing evaluation data, forming a capability perception mechanism driven by historical performance; and second, to optimize the scheduling weight settings through reverse optimization of evaluation data, making the scoring function more closely reflect actual performance differences. The model type can be selected based on the system scale and task complexity, choosing from neural networks, graph structures, or hybrid learning architectures, supporting nonlinear correlation modeling between heterogeneous capabilities.

[0119] Finally, the system updates its parameters based on the model output, writes the optimized capability parameters into the agent capability database as the basis for subsequent capability invocation, and writes the optimized task allocation weight parameters into the task allocation strategy library to update the task scheduling strategy. This mechanism constructs a capability evolution closed loop guided by acceptance results, enabling adaptive evolution and accuracy enhancement of the multi-agent collaborative system in long-term operation.

[0120] Environmental completion status data can be derived from semantically segmented 3D spatial images. Key regional features can be extracted using spatial reconstruction software, yielding dimensional indicators such as the number of objects, object position offsets, and remaining stain area. These indicators are then standardized and converted into numerical vectors of uniform dimensions. Task execution information is obtained by reading task lifecycle logs and execution path records. Information such as task duration, trajectory offset, path overlap, and number of anomalies is aggregated, and statistical methods and Bayesian modeling are used to construct an execution evaluation distribution.

[0121] The machine learning model can employ a multi-task learning framework, constructing a dual-output path: one path learns ability parameter estimation, and the other learns weight parameter adjustment. A loss function is used to optimize the model by combining prediction bias and task completion scoring error. Model training can be performed online in real-time or asynchronously after the task cycle ends. A sliding window is used to collect historical data samples to construct the training set, and validation metrics such as MSE and R² are applied. 2 Evaluate the performance of the model.

[0122] When the amount of data is limited, transfer learning can be used to use models trained in other similar scenarios as initialization parameters to reduce model convergence time. In scenarios with diverse task categories, task type encoding vectors can be introduced as additional inputs to the model to improve the heterogeneous adaptation capability of tasks.

[0123] The ability parameter update merges the old and new values ​​using an exponential moving average method to avoid drastic fluctuations in the parameters caused by a single abnormal result; the weight parameter update reconstructs the scoring function in the form of score normalization to ensure the consistency of the total weight of the scoring system.

[0124] Example Description: In elderly care service scenarios, after the elderly care robot completes the task of tidying up the elderly person's room, the inspection robot uses 3D scanning to capture the distribution of items and the cleanliness of the space. The system compares the generated 3D environment model with a standard bedroom cleanliness model, identifies items that have not been cleaned up under the bed, and marks the discrepancies in the acceptance report, providing precise coordinates for subsequent rework.

[0125] In healthcare settings, post-operative wards need to be restored to a clean state. After the cleaning robot completes its task, the system collects spatial data on key nodes such as beds, medical carts, and disinfection areas. By comparing this data with a standard sterile environment model, it identifies a cart misalignment that resulted in an untreated cleaning blind spot. Based on this, the system determines that part of the cleaning task has not met the standards and notifies the robot to perform additional cleaning.

[0126] In unattended data centers in the financial sector, after the maintenance robot completes the tasks of cleaning the passageways between equipment and inspecting the ground, the system obtains the status modeling results of the task area. By comparing it with the baseline security model, it detects that there are still unprocessed messy cable areas on the side of the server, triggering a non-compliance judgment and generating rework suggestions.

[0127] This embodiment acquires environmental completion status data and compares it with a preset standard model, enabling the system to judge task execution effectiveness in a data-driven manner, overcoming the deficiency that relying solely on task execution logs cannot reflect the final spatial state. Through visual modeling and a structured comparison mechanism, incomplete or incorrectly completed areas can be accurately identified and quantified as difference indicators, further automating acceptance testing. The spatial positioning and task mapping capabilities of the difference areas enhance the system's anomaly closed-loop handling capabilities, providing a data foundation for subsequent rework, evaluation, and capability updates.

[0128] This invention relates to the field of artificial intelligence technology and can be applied to business scenarios such as elderly care services, fintech, and healthcare. It discloses a method, apparatus, device, and medium for multi-agent collaborative task processing, comprising: acquiring task instructions and performing semantic parsing to identify core task objectives, and decomposing the task into multiple sub-tasks; acquiring work environment information and dividing task regions, and generating a global collaboration framework by combining agent capability parameters and region weights; acquiring real-time status information of agents, and generating a task allocation table based on optimal agent matching within the global collaboration framework; issuing the task allocation table to control agents to execute tasks and uploading task execution information; monitoring the execution process based on a shared information platform, triggering a dynamic adjustment strategy when path conflicts or equipment failures are detected, and generating and issuing an updated task allocation table; and acquiring environmental completion status data after the sub-tasks are completed and comparing it with a preset standard model for acceptance. This invention achieves efficient conversion of task instructions into multi-agent collaborative execution by combining task semantic information, environmental structure information, and agent capability parameters for unified modeling, and by introducing a real-time state perception and dynamic allocation mechanism; and improves the stability and reliability of system collaboration through real-time identification and adaptive adjustment of path conflicts and equipment failures.

[0129] In one embodiment, step S10 above includes:

[0130] S101, receives task commands via voice input;

[0131] S102, the task instructions are converted into structured text information through the speech recognition module;

[0132] S103, 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;

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

[0134] S105, the core task objective is decomposed into multiple sub-task categories based on task type characteristics;

[0135] S106, Define execution priority parameters based on the regional association of subtask categories;

[0136] S107, Store the core task objective, sub-task category and execution priority parameters to the task planning database.

[0137] In this embodiment, the acquisition of task instructions relies on a natural language interaction mechanism. The input can be speech, text, or a multimodal composite expression. Speech input serves as a crucial entry point for issuing unstructured tasks in the adapted service scenario, and is received through a microphone array or speech acquisition module. The speech signal needs to be processed by a speech recognition module that combines acoustic modeling and language modeling. Deep neural network structures (such as CTC or RNN-T) are commonly used to convert the speech stream into structured text information, providing functions such as sentence segmentation, error correction, and context preservation to ensure the accuracy of subsequent semantic analysis.

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

[0139] The process of identifying core task objectives needs to be based on the dependencies between semantic elements. Graph Attention (GAT) or a logic-based intent reasoning module should be used to extract high-semantic-density objective definitions from action verbs, target objects, and task context, such as "tidying the bedroom," "cleaning the living room," and "putting medicines away." This allows for the establishment of a task objective representation oriented towards specific performance behaviors. This objective representation should include not only the task name but also information such as object category, target spatial location, and desired state.

[0140] Once the task objective is established, it will be decomposed based on task templates or rule systems in the existing task library. Task type characteristics may include task operation complexity, number of objects, and degree of spatial coupling. For example, "tidying up the room" can be broken down into sub-task categories such as "storing clothes," "cleaning the floor," and "putting books back in their place." Each sub-task category needs to obtain the corresponding operation chain through an action mapping table. For example, "storing clothes" corresponds to the action sequence of "identifying clothes → grabbing → placing in the wardrobe."

[0141] The definition of execution priority parameters is not based on static rules, but rather needs to be calculated by combining task type characteristics and spatial layout relationships. Priority determination can consider the sequential relationships between nodes in the spatial dependency graph; for example, "cleaning the floor" should be performed after "moving obstacles." It can also be determined based on the importance weight of objects, such as "putting medicines back in their place" taking precedence over "organizing bookshelves." Furthermore, current time, environmental constraints, or user preferences can also be used as weighting factors in priority calculation, forming a dynamic priority vector.

[0142] Finally, the core task objectives, subtask categories, and corresponding execution priority parameters need to be written into the task planning database. This database should have task semantic version control capabilities, execution instance storage capabilities, and task relationship network modeling capabilities, providing a semantic foundation for subsequent task region partitioning, task scheduling, and execution feedback. It is recommended to use a graph database (such as Neo4j) or a document-oriented database with nested structures (such as MongoDB) to facilitate the expression of logical dependencies, shared parameters, and state inheritance relationships between subtasks.

[0143] This embodiment, through structured processing and semantic modeling of natural language task instructions, enables the automatic understanding and decomposition of complex tasks in non-rule, non-programming interactions, enhancing the system's ability to interpret human intentions. By introducing a pre-trained language model, semantic element extraction, and task structure mapping, it achieves the restoration of implicit information in instructions and action-level decomposition, avoiding ambiguity and information loss in instruction understanding. Furthermore, by dynamically generating execution priority parameters based on task type characteristics and spatial logical relationships, the execution order of subtasks no longer depends on preset static templates and can adaptively adjust according to the current task content and environmental conditions. After uniformly storing the task decomposition results in a task planning database, it not only supports the rapid invocation of subsequent task scheduling algorithms but also forms a structured accumulation of task execution knowledge.

[0144] In one embodiment, step S20 above includes:

[0145] S201 acquires three-dimensional spatial data of the working environment through a spatial scanning sensor;

[0146] S202, Identify environmental structural features based on the three-dimensional spatial data;

[0147] S203, Divide the environment into multiple task areas according to the environmental structural characteristics;

[0148] S204, determine the area parameters and functional requirement parameters for each task area;

[0149] S205, determine the region weight based on the area parameter and functional requirement parameter;

[0150] S206, query the capability parameters of multiple agents from the agent capability database;

[0151] S207, Calculate the matching score between each agent and each sub-task category based on the capability parameters;

[0152] S208, Generate a priority allocation order according to the matching degree scores;

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

[0154] In this embodiment, the process of acquiring work environment information is based on spatial perception, aiming to establish a quantifiable and structured spatial representation model for task area division and resource scheduling. This process begins with modeling the work environment using spatial scanning sensors, such as RGB-D cameras, LiDAR, structured light sensors, or multimodal fusion devices. These devices transform the environment into three-dimensional spatial data, in the form of point clouds, voxel maps, depth maps, or semantic meshes. The three-dimensional spatial data undergoes denoising, registration, and stitching processing to form a globally resolvable digital model of the environment, possessing geometric and topological information accessible to algorithms.

[0155] Based on 3D data, spatial structure recognition technology is used to extract environmental structural features, specifically including spatial closure detection, wall and door / window location recognition, planar region segmentation, furniture recognition, and spatial boundary fitting. Implementation methods can employ RANSAC planar extraction, supervoxel clustering, and deep convolutional semantic segmentation networks (such as MinkowskiNet). Output results include spatial boundary coordinates, obstacle distribution maps, and functional area candidate sets, providing semantic support for subsequent region division tasks.

[0156] Once the environmental structural features are identified, the entire work environment needs to be divided into multiple task areas. The division of task areas is not solely based on geometric distance or size, but rather on a fusion of multi-dimensional characteristics such as spatial function, object density, and cleaning complexity. For example, a living room can be divided into three task areas: the "coffee table area," the "sofa area," and the "TV cabinet area," based on spatial accessibility, the number of obstacles, and the suitability of the task type. The division process can construct a spatial similarity map, aggregating spatial sub-blocks into functionally consistent task areas.

[0157] Each task area requires the extraction of area parameters and functional requirement parameters. The area parameter can be accurately calculated using 3D modeling algorithms to determine the projected area or volume of the area. The functional requirement parameters are derived based on factors such as the objects contained within the area, spatial density, and task frequency. These parameters express the task complexity and priority of the area in the task processing. For example, the area around the dining table has high functional requirements due to its dense clutter, while the area in the corridor has low requirements due to its simple cleaning tasks. The functional requirement parameters can be quantified into dimensions such as expected task processing load and task complexity factors, forming an area attribute vector.

[0158] Based on a combination of area and functional requirement parameters, region weights are determined. Region weights not only reflect the size of the task but also comprehensively consider spatial difficulty and resource scheduling priority. Regularization methods can be introduced into the weight calculation to prevent excessively high weights for individual large but low-complexity regions, which could affect the balance of task allocation. Weight values ​​can be normalized and used as multiplication factors in subsequent scheduling calculations.

[0159] The acquisition of capability parameters relies on an agent capability database. Capability parameters include, but are not limited to, cleaning speed, grasping accuracy, recognition accuracy, obstacle avoidance capability, and energy efficiency, and are formed through static parameter tables or dynamic performance backtesting. After retrieving these parameters from the database, a matching degree calculation needs to be performed between each agent and various sub-tasks. The matching degree measures the agent's adaptability in completing a specific sub-task, and the calculation method can be based on multi-dimensional parameter similarity, historical execution performance, or output prediction by a learning model. The matching degree value is usually represented by a scoring matrix, where rows correspond to agents, columns correspond to sub-task categories, and cells represent the corresponding matching scores.

[0160] After computation is completed for each agent, they are sorted according to their matching scores to generate a priority allocation order. The sorting mechanism can employ a weighted average method, a hierarchical scoring mechanism, or a heuristic sorting function, supporting the priority matching of agents with high suitability for specific sub-task categories to construct a task scheduling sequence.

[0161] Building upon this foundation, priority allocation order and regional weights are integrated to generate a global collaboration framework for the entire task scenario. This global collaboration framework should have a three-layer nested structure: task partitioning, capability allocation, and priority level. It should support concurrent scheduling and task reconfiguration by multiple agents and reserve interfaces for dynamic adjustment. The collaboration framework can be represented as a multi-layered graph structure, with task regions as nodes, weights as edge attributes, and agent matching as graph embedding paths. This structure serves as the decision-making basis for subsequent task allocation and dynamic collaboration, providing the system with a structured, dynamic, and data-driven task collaboration model.

[0162] This embodiment divides the task region based on environmental spatial data and structural features, and introduces area parameters and functional requirement parameters as quantification criteria. This maps the spatial environment into discrete task regions with execution attributes, significantly improving the accuracy and operability of task modeling. Through a matching degree scoring mechanism between capability parameters and regional task requirements, dynamic adaptation between the agent and the task is achieved, thereby avoiding resource mismatch and improving task execution efficiency. The fusion of priority allocation order and region weights can balance task urgency and optimal resource utilization, ultimately forming a globally optimal collaborative structure.

[0163] In one embodiment, step S30 above includes:

[0164] S301, associate each subtask with its corresponding execution region;

[0165] S302, acquire real-time status information for each agent, including idle state, energy level and current position;

[0166] S303, determine the distance parameter between the current position of each agent and the subtask execution area;

[0167] S304, Extract the matching score between the agent and the subtask from the global collaboration framework;

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

[0169] S306: Select the idle agent with the highest comprehensive score for each subtask and generate a task allocation table containing subtasks, execution areas and assigned agents;

[0170] S307, Store the task allocation table in the task allocation database.

[0171] In this embodiment, associating each subtask with its corresponding execution region is the first step in achieving accurate task distribution. This process establishes a mapping relationship between task semantic information and spatial topology. Each subtask already has a category label and target behavior description during the semantic planning stage, such as "tidying up the desktop," "cleaning the floor," and "moving clutter." Combining the task region division results generated in the environmental structure analysis, region binding is completed based on the matching degree between task semantics and region functions. For example, a one-to-one mapping is established between the "tidying up the desktop" task and the "desktop area in the study." This process can be automatically completed through similarity calculation between task labels and region semantic labels, or mapping rules can be preset in the rule base to ensure the accuracy of task spatial positioning.

[0172] Obtaining real-time state information for each agent is a prerequisite for dynamic scheduling. 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; this information typically comes from the control command response system or the execution thread management module, indicating whether the number of current task execution threads is zero. The energy level reflects the agent's endurance and comes from the energy monitoring module, providing information such as remaining battery percentage and battery health status. The current position represents the agent's real-time coordinates in the environmental space, derived from the localization module (such as visual SLAM, laser SLAM, or UWB localization systems), and is typically represented as a three-dimensional position vector in the world coordinate system or two-dimensional planar coordinates.

[0173] Determining the distance parameter between each agent's current position and the subtask execution region is a crucial method for quantifying scheduling costs. The task execution region can be modeled as a spatial boundary or center point, while the agent's current position is a dynamic point. Based on this, the distance between the two can be calculated using methods such as Euclidean distance, A* path length, or navigation cost. The specific method depends on the environmental complexity and computational resource constraints. In high-degree-of-freedom environments, path cost distance should be prioritized to reflect the true navigation complexity. In obstacle-dense or multi-layered structural scenarios, graph search or topological graph cost function calculations need to be introduced.

[0174] Extracting the matching score between the agent and subtask from the global collaboration framework is a quantitative inheritance of the historical adaptability between the agent's capabilities and task requirements. The matching score is modeled in the previous stage using capability parameters, task category attributes, and historical success rates, in the form of a bidirectional mapping matrix. The agent is represented as a row, and the task category as a column, with the score indicating the expected adaptability to complete a certain type of task. During extraction, the scheduling and scoring module in the collaboration framework structure can be directly accessed to quickly return the current agent's evaluation value for the target task.

[0175] Matching score, idle state, energy level, and distance parameters collectively form the basis of scheduling decisions. Calculating the overall score requires constructing a scheduling scoring function, which can employ a weighted linear combination or a nonlinear evaluation model. In a simple strategy, the following function can be defined:

[0176] S=α×M+β×E+γ×(1-D)+δ×I

[0177] Where S is the overall score, M is the matching score, E is the normalized energy level, D is the normalized distance cost, I is the idle state identifier (binary), and α, β, γ, and δ are scheduling weight parameters. The weight values ​​can be obtained through empirical configuration, regression of historical task performance, or dynamic learning using a machine learning model. In more complex scenarios, a neural network structure can also be used, with the above parameters as input, to train and output the optimal scheduling value.

[0178] After the overall score is calculated, for each subtask, the system selects the agent with the highest overall score from the set of currently idle agents as the executor, ensuring that the optimal agent resources are allocated preferentially. To ensure balanced resource distribution, a task load recording or cooldown time mechanism can be introduced to suppress the number of consecutive schedulings of the same agent, thereby preventing resource overload.

[0179] The final task allocation result consists of a triple: (subtask ID, execution region ID, and allocating agent ID). The task allocation table summarizes the allocation information for all subtasks, typically structured as a task item list or mapping table, supporting subsequent instruction issuance and dynamic adjustments. The task allocation table is stored in the task allocation database, ensuring the traceability and auditability of scheduling results and providing fundamental data support for fault recovery and task reassignment. The database supports asynchronous update mechanisms to maintain data consistency with the shared information platform.

[0180] This embodiment employs a spatial binding mechanism between task execution regions and subtasks, providing clear spatial constraints for the task scheduling process and avoiding location ambiguity and path redundancy during task execution. By constructing an agent state profile using multi-dimensional state information, it achieves accurate identification of schedulable resources, effectively mitigating the risk of task failure due to insufficient energy or excessively long paths. The comprehensive scoring mechanism integrates multiple factors such as matching degree, energy consumption, distance, and availability during the scheduling process, exhibiting good policy 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, step S40 above includes:

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

[0183] S402, 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;

[0184] S403 collects dynamic environmental data by assigning sensors to the intelligent agent;

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

[0186] S405, mark the status label of the structured execution information;

[0187] S406, Upload structured execution information with status tags to the shared information platform in real time;

[0188] S407, Update the work area status in the shared information platform.

[0189] In this embodiment, parsing the task allocation table is not merely reading data, but the primary step in deconstructing the scheduling intent. The task allocation table encapsulates information such as subtask identifiers, target execution regions, expected target states, assigned agents, task priorities, and time constraints. The parsing module needs to support complex data structures, including multi-table joins, enumeration type identification, field integrity checks, and syntax and semantic consistency verification. It also needs to verify historical scheduling records and boundary conditions, such as whether the current agent has already executed such tasks. Abstract Syntax Tree (AST) parsing and data constraint checking mechanisms are used to ensure parsing reliability throughout this process.

[0190] Sending a task initiation command to a designated agent involves a high degree of coordination between the task distribution channel and the control interface. The initiation command encapsulates region information (such as geometric boundary points, work altitude, and suggested task path), behavioral templates (action sequence, execution module calls), fault tolerance mechanisms (retry count, rollback conditions), and priority tags. The system securely transmits the command to the agent control unit using REST, DDS, or ROS2 service call protocols. Upon receiving the initiation command, the control module triggers local path planning, safe entry logic, work module activation logic, and execution monitoring submodules to ensure a coordinated chain from command issuance to execution initiation.

[0191] During execution, the agent, equipped with a sensor array, continuously collects dynamic environmental data. This dynamic data includes visual information (images, depth maps), laser scan information, tactile feedback, sound events, force sensor readings, etc. Data acquisition follows a high-frequency sampling strategy and an asynchronous event triggering mechanism to ensure that key environmental changes are accurately captured. The raw sensor data is in continuous time series format and has high-dimensional characteristics, requiring initial processing through filtering (such as Kalman filtering and time window moving average) and noise reduction (such as median filtering and Fourier filtering).

[0192] Structured environmental dynamics data into execution information involves mapping from perceived conditions to execution semantics. The structured information includes fields such as object location, state changes (e.g., whether a door is closed), completed sub-steps (e.g., an object is grasped), and abnormal events (e.g., grasping failure, obstacle blockage). The transformation method combines image semantic segmentation (e.g., Mask R-CNN), object recognition models, and state machine annotation mechanisms to construct an event log table and state change timestamps.

[0193] The status tags that mark structured execution information primarily reflect task progress and exception types. Status tags are categorized into several types: preparing, executing, stage complete, exception pending, retrying, and completed. The agent needs to write the status tags into the packet header through the execution feedback loop, along with the status code and detailed execution data, so that the scheduling and monitoring modules can interpret them in real time.

[0194] Real-time uploading of structured execution information with status tags requires support for high-bandwidth, low-latency information channels. A parallel batch upload and incremental event upload mechanism is employed: regular data is uploaded in batches over time segments, while exceptions or status change events trigger immediate single-item uploads. The communication framework may use WebSocket or MQTT QoS2 guarantees, combined with a local transmission buffering module and upload acknowledgment mechanism to ensure upload reliability.

[0195] After receiving data, the shared information platform needs to map it to the "work area status model," updating task completion status, abnormal nodes, resource occupancy status, and area security status within the space. It also needs to provide an interface in real-time for the scheduling, monitoring, and learning modules to access. Updating the model ensures state consistency and supports multi-view mapping, such as two-dimensional floor plans, three-dimensional environment diagrams, and timeline views.

[0196] This embodiment achieves high-fidelity transmission of scheduling decisions to the agent operation layer by explicitly parsing the task allocation table and standardizing the task initiation mechanism, significantly improving task execution accuracy. The introduction of multi-source dynamic data acquisition and structured processing during execution ensures that task status is not only controllable but also verifiable, guaranteeing global consistency and continuity of the system under asynchronous multi-agent collaboration. Standardized uploading and unified tagging of task execution information enable high traceability and automated evaluation capabilities at each stage of task fulfillment, providing reliable support for subsequent parameter optimization, fault response, and scheduling reallocation of the scheduling model.

[0197] In one embodiment, step S50 above includes:

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

[0199] S502, when the intersection distance of the movement trajectories of two intelligent agents is less than the safety threshold, it is determined to be a path conflict event;

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

[0201] S504, Generate a passage order decision based on the task urgency;

[0202] S505, based on the passage order decision, the low-urgency agent suspends movement;

[0203] S506, Plan an avoidance path for the paused intelligent agent and update the space occupancy status of the shared information platform;

[0204] S507, When processing a path conflict event, send a movement command containing the avoidance path to the suspended agent;

[0205] S508, when the energy level of the intelligent agent is detected to be lower than the fault threshold, it is determined as a device fault event;

[0206] S509, for the faulty intelligent agent in the equipment fault event, an idle intelligent agent whose capabilities are matched is retrieved;

[0207] S510, Obtain the work progress parameters of the fault task from the shared information platform;

[0208] S511, Determine the remaining workload parameters based on the work progress parameters;

[0209] S512, when processing a device failure event, send a new task start command containing the remaining workload parameters to the idle agent;

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

[0211] S514, the updated task allocation table is sent to the corresponding intelligent agent.

[0212] In this embodiment, the continuous monitoring mechanism of the shared information platform is designed for the entire task execution process of the multi-agent system. Its core features include real-time data collection, centralized synchronization, dynamic analysis, and immediate response driven by task behavior data. The agent's movement trajectory can be constructed from odometry, visual SLAM modules, IMU sensors, UWB modules, etc., at each agent's end, and uploaded to the shared information platform via a message middleware, forming a continuously trackable multi-agent position dataset in both spatial and temporal dimensions. This dataset supports millisecond-level status updates, ensuring the platform has real-time control over the overall motion situation.

[0213] Based on this, the system calculates the set of intersection points and the nearest distance between the movement trajectories of any two agents, and compares this value with a set safety threshold to automatically detect potential path conflicts. The safety threshold is determined by comprehensively considering parameters such as robot size, navigation accuracy, and task density in the specific scenario; it essentially defines the behavioral safety boundary. When the trajectory intersection area is smaller than this threshold, it is immediately identified as a path conflict event.

[0214] The handling of path conflict events introduces the concept of task urgency, which is calculated based on the current task progress and area status within the work area. Task progress reflects the degree of task completion in the current area and can be determined through task markers, completion marker counts, or task feedback logs. Area status can be constructed based on information such as spatial cleanliness and spatial coverage fed back by the environmental perception module. Each conflicting agent is assigned a priority score through an urgency calculation function.

[0215] A passage order is determined based on urgency scores, pausing the movement of low-priority agents to ensure the smooth flow of high-priority paths. The pause operation is achieved by sending motion commands to suppress the execution of the navigation controller, thus maintaining local state awareness while avoiding path overlap.

[0216] For paused 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 the trajectories of other agents based on the current space occupancy graph, and updating the space occupancy status of the area involved in the path in the shared information platform in real time.

[0217] Afterward, the platform sends specific movement instructions for the obstacle avoidance path to the paused agent. The instruction format can include path point sequence, destination coordinates, speed parameters, etc., to ensure that its behavior runs independently under the obstacle avoidance logic.

[0218] When the monitoring module detects that the energy level of a certain intelligent agent (provided by the battery management system or power consumption prediction module) is lower than a preset threshold, it triggers device fault event identification. Such events require dynamically retrieving idle intelligent agents with high compatibility with the fault task. This is done by querying the capability database and filtering currently idle entities, and selecting them based on their capability matching degree with the task type.

[0219] Subsequent operations require obtaining the work progress parameters corresponding to the faulty task from the shared information platform, including the percentage of task completion, the remaining path of the task, the amount of data collected, etc., and calculating the remaining workload of the task based on these parameters, such as the cleaning area, the number of items moved, or the number of scan frames.

[0220] Based on the remaining workload, a new task initiation instruction is generated for the selected idle agent, including the task objective, execution area, and remaining subtask identifiers. The new instruction directly initiates the task execution process, ensuring that the task continues uninterrupted.

[0221] Ultimately, the platform integrates all issued movement commands and new task commands, and reconstructs the task allocation table to cover changes in current task status and executing entities. Before being issued, the new task allocation table undergoes consistency checks and state mapping confirmation to ensure no execution conflicts or resource reuse issues are introduced. This table is then synchronously distributed to all relevant agents via message queues or command buses, enabling overall adjustment and coordinated recovery of distributed tasks.

[0222] This embodiment enables rapid identification of path conflicts and equipment malfunctions by continuously monitoring the movement trajectory and energy status of agents in a shared information platform; it can dynamically adjust the behavior of low-priority agents by generating a passage order based on task urgency and planning avoidance paths, thereby effectively alleviating local path congestion; it can achieve agent replacement without task interruption by extracting the work progress of faulty tasks and generating replacement instructions for idle agents; and it can ensure the continuity of task collaboration and execution consistency of the entire multi-agent system by integrating movement instructions and replacement task instructions to generate an updated task allocation table and redistribute it.

[0223] In one embodiment, step S60 above includes:

[0224] S601, control the inspection equipment to scan the task execution area, and collect environmental spatial data of the task execution area through the inspection equipment;

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

[0226] S603, Obtain preset standard model data;

[0227] S604, compare the features of the three-dimensional environment model with the preset standard model, and mark the difference areas in the three-dimensional environment model that do not meet the standard;

[0228] S605, generate an acceptance report containing the coordinates of the difference region, and store the acceptance report in the acceptance database.

[0229] In this embodiment, after the task is completed, in order to achieve systematic verification of the execution results, it is necessary to obtain the spatial state expression information of the task execution area after the task is completed, and compare and analyze this state with the predefined expected standards, thereby realizing the structured acceptance of task quality. First, the inspection equipment is controlled to scan the task execution area. This equipment is usually a mobile intelligent agent equipped with spatial perception components such as LiDAR, RGB-D camera, TOF depth camera, and multi-band vision sensor. It can autonomously plan the scanning path and perform full-coverage data collection of the designated area without interfering with the on-site layout.

[0230] The raw environmental spatial data collected during the inspection process reflects the actual spatial state after the task is completed. Data types include dense point clouds, depth images, structured light reconstruction fragments, and local mesh maps. Timestamps, pose information, and scene labels can be included during acquisition to support subsequent temporal and spatial reconstruction. This raw data undergoes a preprocessing workflow, including registration, filtering, denoising, coordinate unification, and viewpoint fusion, to generate a spatial representation with topological continuity and semantic consistency—that is, a 3D environmental model.

[0231] The constructed 3D environment model should possess characteristics such as complete spatial coverage, clear geometric boundaries, and accurate semantic partitioning, for comparison with a preset standard model. The preset standard model represents the spatial state under ideal task completion conditions. It is generated based on task category, regional function, historical high-quality completion samples, or user-defined expectations. Its data format can be a 3D point cloud template, semantic occupancy map, CAD modeling file, or multi-scale spatial schematic diagram, and is compared with the current environment model through feature comparison.

[0232] During the comparison process, a multi-level spatial feature matching strategy was employed. First, geometric-level spatial alignment was performed using the Iterative Closest Point (ICP) algorithm, a global registration network, and feature point matching to achieve coarse and fine registration between the two models. Subsequently, at the semantic level, the conformity of key structures such as object category, spatial boundaries, and item location was analyzed to identify abnormal placement, residual foreign objects, and inadequate cleaning, which violated the standards. All locations or regions that did not meet the standard conditions were marked as discrepancy areas, and their corresponding spatial coordinates, physical dimensions, deviation type, and other multi-dimensional attributes were extracted.

[0233] The aforementioned discrepancies are treated as acceptance results to be corrected. An acceptance report is generated in a structured format, detailing the spatial location information, problem category labels, and quantified deviation values ​​for each discrepancy area, forming interpretable data for the execution system, users, or regulators. The acceptance report is generated as a data table, labeled graph, or model annotation and written to the acceptance database. This database supports task-level indexing, region-level comparison, and time-series evolution analysis, ensuring the traceability and sustainable optimization capabilities of the acceptance results.

[0234] Example Description: In a smart elderly care facility with multi-agent service capabilities, the system receives a voice task instruction from the administrator: "Please arrange for the robot to clean room A101 for the elderly this morning, making the bed, tidying up the floor, and wiping the table." This task is first converted into structured text information by the speech recognition module, and then semantically parsed by a pre-trained language model to identify key actions and objects such as "cleaning," "making the bed," "tidying up the floor," and "wiping the table." The system further divides the task into sub-tasks: ① making the bed, ② cleaning the floor, and ③ wiping the table, and determines the required operational capabilities and equipment support capabilities based on the task type.

[0235] Subsequently, the system uses spatial scanning sensors to collect 3D spatial data of room A101, identifying structural features such as room layout, furniture distribution, and passageways. Based on this, it divides the room into task areas, such as the bed area, floor area, and desktop area, labeling each area with parameters such as area, purpose label, and complexity level. This further determines the task weight for each area; for example, the bed area, directly related to the elderly's sleep, has a weight of 0.45, the floor area 0.35, and the desktop area 0.2. Simultaneously, the system retrieves capability parameter data for all currently idle service robots in the facility from the intelligent agent capability database, including indicators such as grasping accuracy, surface wiping ability, navigation ability, and obstacle avoidance stability.

[0236] The system calculates the matching degree between each sub-task category and the agent's capability parameters, generates a matching degree score, and generates a global collaboration framework based on the weight of each task region, forming a priority ranking of agent allocation to task regions. At this time, the system does not immediately assign tasks, but simultaneously retrieves the agent's real-time status information, such as location information, battery level, idle status, execution history, etc., quantifies the distance parameter between its current location and the target task region, and combines it with factors such as task matching degree score, energy reserve, and executability to comprehensively evaluate the current adaptability of each agent to each sub-task.

[0237] For example, robot A3 is currently closest to the ground area, has a high floor cleaning ability score, sufficient energy, and is idle; therefore, the system assigns it to subtask ②. Robot A1, with its high-precision operation capabilities, although its navigation speed is slightly lower, is assigned to subtask ① because of its strongest grasping and detail-oriented abilities. Ultimately, the system generates a task allocation table, including the name of each subtask, its corresponding execution area, the specific robot number assigned, and the expected execution window time. This table is stored in the database and then distributed to the relevant agents.

[0238] Upon receiving a task, each service robot moves to its designated task area and initiates the corresponding task program. During this process, each robot continuously collects dynamic data through its onboard sensors (such as infrared cameras, pressure-sensitive modules, and surface smoothness detectors), including environmental changes, item status, and completion status. This data is then structured into execution information with status tags and uploaded to the shared information platform in real time. For example, robot A3 uploads "Floor cleaning completion status: 90%, remaining area near the doorway has obstacles," and the system updates the spatial status mapping accordingly.

[0239] During task execution, if the system detects a potential path conflict between two robots (e.g., A2 and A3) in a narrow passage (predicted intersection distance less than 0.5 meters), it will immediately trigger a dynamic adjustment strategy through the shared information platform. After analyzing the urgency of their respective tasks, it will control the robot with lower priority (e.g., A3) to pause its movement and plan an alternative path for it. The adjustment results will be sent out for execution via movement commands. If an agent (e.g., A1) automatically goes offline due to battery depletion while performing a desktop wiping task, the system will immediately determine this as a device failure event, call upon an idle agent A5 with replacement capabilities, and determine the remaining task content by comparing historical task states. A new task start command will then be issued to ensure that the task is not interrupted.

[0240] After all sub-tasks are completed, the system dispatches an inspection robot to room A101 to conduct a structured acceptance inspection of the cleaning results. The inspection equipment performs 3D spatial reconstruction, collects current room status data, and generates a 3D environment model. The system retrieves preset standard models set before task completion from the acceptance database, including the expected shape of a clean bed, a space template of a clutter-free desktop, and floor cleanliness image samples. It identifies deviation areas through model registration and spatial feature comparison. For example, if residual paper scraps are detected in a corner of the floor, the system will record in the acceptance report, "There is a cleaning deficiency in the lower right corner of the floor area, with a deviation area of ​​12cm²." 2 ", and generate difference labels with coordinates.

[0241] Ultimately, the acceptance report was written into the acceptance database and served as an important basis for subsequent task evaluation and capability learning, participating in the process of capability parameter optimization and task allocation strategy adjustment. The entire process realized a closed loop of elderly care service, from semantic perception to spatial scheduling, from execution collaboration to result acceptance, improving task adaptation accuracy, response efficiency, and service quality. In particular, it demonstrated good adaptability and human-machine collaboration capabilities in meeting the needs of elderly people in flexible organization, cleaning and maintenance, and anomaly handling in their living spaces.

[0242] In a large smart healthcare facility, the nursing dispatch center system receives a task request from a doctor via a voice terminal: "Arrange for a robot to tidy up the bed, return medical equipment to its proper place, and clean the floor in ward 303." This task is converted into structured text by a speech recognition system. After being parsed by a semantic understanding model, the core task objective is identified as "ward environment restoration," and based on this, it is divided into three sub-tasks: ① bed tidying, ② medical equipment sorting and returning, and ③ floor cleaning. The system further analyzes the spatial elements and operation types involved in the task semantics, defining task type labels and operation priorities.

[0243] Subsequently, the spatial perception system deployed by the medical institution used ceiling LiDAR and visual cameras to perform a 3D scan of ward 303, generating complete spatial point cloud and image fusion data, and identifying spatial structural features such as bed areas, instrument racks, and passageways. Based on this structural information, the system divided the ward into three task areas and calculated the area, complexity, and key indicators for each area. For example, the task weight for the bed area was set to 0.5 because it involves direct patient contact, 0.3 for the instrument area, and 0.2 for the ground area. Simultaneously, the system queried data from all standby robots in the intelligent agent capability database, including capability parameters such as robotic arm stability, disinfection spraying accuracy, instrument recognition accuracy, and obstacle avoidance ability.

[0244] Subsequently, the system assesses the capabilities of each agent and subtask, calculates a matching score, and synthesizes a priority allocation matrix based on task area weights and matching scores to form a global collaborative framework. Then, considering the current real-time status of each agent (e.g., robot C1 is charging, C2 is idle and close to room 303, C3 is currently completing a task in room 304 and is about to finish), the system calculates the comprehensive suitability score of the currently schedulable agents for each subtask, taking into account their location, idle status, battery level, and historical task stability.

[0245] For example, disinfection robot C2, located near ward 303 and possessing efficient wiping capabilities, was assigned to task ③, floor cleaning; C4, with its high-precision grasping ability, was scheduled to perform task ②, instrument return; and nursing robot C5, equipped with a bed surface flatness recognition module, was assigned to task ①, bed tidying. The system generates a complete task allocation table and distributes it to each agent, while simultaneously writing the task details into the task allocation database and recording the scheduling behavior.

[0246] During task execution, each robot automatically moves to its corresponding area according to task instructions, initiates the program, and collects execution data in real time through its onboard vision and tactile sensors. For example, C5 uses a tension sensor to sense the balance of the bed sheet tension and continuously corrects itself; C4 identifies the shape labels of medical scissors and tweezers and categorizes them. All robots upload structured dynamic data during execution to a shared information platform, including execution paths, anomaly markers, and task status. The platform continuously updates the task completion status map and agent location map of the current ward for the scheduling system's reference.

[0247] If a path conflict occurs during execution, such as C2 and C5 intersecting in the right-side passageway of the bed due to asynchronous coordination, the system will detect that the intersection of their movement trajectories is less than the safety threshold of 0.4 meters and trigger a conflict resolution strategy. After analyzing the urgency of both tasks, the system determines that bed preparation has a higher priority, and C2 is instructed to pause and replan its detour path to avoid the conflict. If C4 experiences abnormal robotic arm force feedback and interrupts its task during execution, the system determines it to be a device malfunction and immediately calls upon C6, which is idle and has similar capabilities, to take over the unfinished task, issuing a continuation execution instruction based on the recorded task progress.

[0248] Once all sub-tasks are completed, the system dispatches a patrol robot to the ward to conduct a structured acceptance review of the overall task results. The patrol robot uses a 3D camera and surface cleanliness sensors to acquire current ward status data and generate a 3D environment model. This model is then compared with a pre-set standard ward status model from the medical institution to determine the spatial structure and functional areas. The system identifies dimensions such as the degree of bed sheet wrinkles, instrument placement coordinates, and floor cleanliness light reflectance, marking areas that do not meet standards and generating detailed acceptance reports, such as "bed edge flatness less than 8 mm" or "instrument placement deviation exceeds the standard by 3 cm." These acceptance reports are then stored in the acceptance database.

[0249] Based on this, the system extracts environmental completion status data from the acceptance report and operation logs during task execution, generating environmental quality assessment indicators and task efficiency indicators. It then learns the mapping relationship between the capability parameters of each agent and task performance through a pre-trained neural network model. For example, if C2's cleaning efficiency is found to be better than expected, the system increases its "ground cleaning efficiency" value to 0.87 and updates the "disinfection task matching weight" parameter in the task allocation weights, increasing its priority in future similar tasks. The optimization results are written into the agent capability database and the task allocation strategy library, providing intelligent scheduling support for similar tasks in the future.

[0250] Through the above process, the system realizes closed-loop control of multi-task, multi-capability, and multi-state collaborative scheduling in the medical environment, which significantly improves the efficiency of robot resource allocation, the speed of ward environment recovery, and the reliability of service response. It has good applicability in scenarios with high sensitivity to environmental conditions, such as postoperative wards and intensive care units.

[0251] During the operation of a smart bank branch, the operations management system received a business optimization request: "Adjust the configuration of the robots for customer guidance, document delivery, and self-service terminal guidance in the lobby." The system collected voice task instructions through the front-end voice interaction terminal, converted them into structured text using a language recognition engine, and input them into a natural language understanding model. The system identified the task objective as "reconstructing the bank branch service process." Semantic deconstruction results showed that this objective could be broken down into three service sub-tasks: customer guidance, document delivery, and self-service terminal guidance, corresponding to spatial areas such as the customer waiting area, document processing area, and self-service area, respectively. Task priority parameters were established based on customer traffic density and task response time limits.

[0252] The system retrieves data from multi-source sensing devices deployed throughout the hall, including ceiling panoramic cameras, LiDAR, and RFID landmark tags, to construct a real-time 3D spatial state model of the hall. The system identifies spatial structural units, including the information desk, VIP area, and self-service area, and further divides them into multiple task zones. Each zone is weighted based on dimensions such as area, pedestrian density, service type, and scene complexity; for example, the customer guidance area has a weight of 0.4, the document delivery area 0.35, and the self-service equipment area 0.25. The system then reads data from the intelligent agent capability database of the service robot cluster, including capability parameters such as navigation accuracy, voice broadcast clarity, document recognition ability, and familiarity with equipment operation.

[0253] Based on the correlation between capability parameters and sub-task types, the system calculates the matching score of all schedulable agents, combines it with regional weights to generate a priority scheduling order, and constructs a global collaborative framework for the current bank branch service scenario accordingly. The real-time status perception system synchronously collects information such as the current position of each robot, whether it is currently in a task state, and its remaining battery power, and calculates its distance parameters from each task area. The system integrates current capability matching degree, real-time availability, remaining energy, and path cost to form a comprehensive scoring mechanism to achieve optimal task allocation. For example, Robot-B, due to its stable voice broadcast frequency, familiarity with the lobby layout, and current idle status, receives the highest score and is assigned to complete the customer guidance task; Robot-C is assigned to complete the document delivery task; Robot-D has a terminal indicator light recognition module and is adapted for terminal guidance tasks.

[0254] After the task allocation table is generated, it is distributed to each service robot. Robot-B moves to the bank entrance according to the task start command, performs guidance tasks for customers entering the store, and broadcasts the waiting time and queue number of each window in real time; Robot-C starts from the document archiving area, follows the planned path to the back office to retrieve contract texts, and avoids collisions with customers along the way using sensors; Robot-D patrols the self-service terminal area, proactively identifies customers who are not actively using the equipment, and provides guidance on how to use the equipment. During the execution process, each robot collects dynamic data including service response time, task completion status, path deviation, and customer dwell behavior, and uploads it to the shared information platform in real time to build a complete service status map.

[0255] If Robot-C experiences a low battery alarm during file transfer, the system automatically triggers a fault response mechanism. It extracts the current task information and progress parameters and reassigns it to Robot-E, which is available and has the same transfer capability. The scheduling system calculates its proximity to the target path and current workload, assesses feasibility, and issues a takeover command to ensure uninterrupted file transfer. If Robot-B and Robot-D intersect due to customer congestion in the self-service area during peak hours, the system detects that their paths intersect within a safety threshold of 0.3 meters. After calculating the task urgency, it determines that Robot-D must pass first, guiding Robot-B to wait and detour to avoid chaos or congestion in the environment.

[0256] After the task is completed, the system dispatches an acceptance robot to patrol the entire service environment, acquiring data on space usage and changes in customer distribution. Based on visual perception and semantic analysis, a 3D service status model is generated and compared with a preset standard service hall service status model, including feature dimensions such as customer flow efficiency, equipment usage status, and file transfer errors. Areas of difference are marked, and a structured acceptance report is generated. The system identifies data such as "the average customer dwell time in the self-service terminal area is too long" and "the frequency of file transfer path deviation is high," and pushes this data to the acceptance database.

[0257] The system further extracts the acceptance report and execution log information to generate environmental quality assessment indicators and task efficiency assessment indicators, such as service response time, path duplication rate, and interaction completion rate. These assessment indicators are input into a neural network analysis model to model the non-linear relationship between task performance and capability parameters, outputting optimized values ​​for capability parameters and optimized values ​​for task allocation weights. For example, recognizing Robot-D's good adaptability in terminal guidance tasks, the system increases its "device understanding capability" parameter and enhances the matching weight of this capability in the terminal scenario within the task scheduling function. This updated data is stored in the agent capability database and task allocation strategy library, respectively, for optimizing subsequent scheduling strategies for the daily operations of bank branches.

[0258] This embodiment uses a controlled inspection device to spatially scan the task execution area and construct a 3D environment model, enabling a structured reconstruction of the spatial state after task completion. By comparing the 3D model with a preset standard model and precisely marking the differences, the quality of task execution can be quantitatively evaluated. By generating an acceptance report and storing it in an acceptance database, the digital archiving and traceability of task completion are achieved. This process not only significantly improves the verifiability and quality control capabilities of multi-agent systems in collaborative task execution but also provides a solid data foundation for subsequent data-driven optimization based on feedback.

[0259] In one embodiment, a multi-agent collaborative task processing device is provided, which corresponds one-to-one with the multi-agent collaborative task processing method described in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the multi-agent collaborative task processing device of the present invention. The modules include a task parsing module 10, a task planning module 20, a task allocation module 30, a task execution module 40, a dynamic scheduling module 50, and a task acceptance module 60. Detailed descriptions of each functional module are as follows:

[0260] Task parsing module 10 is used to acquire 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.

[0261] Task planning module 20 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 generate a global collaboration framework by combining the capability parameters of multiple agents and the region weights.

[0262] Task allocation module 30 is used to obtain real-time status information of multiple agents, combine the global cooperation framework and the real-time status information, match the optimal agent for each sub-task, and generate a task allocation table.

[0263] Task execution module 40 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;

[0264] The dynamic scheduling module 50 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.

[0265] The task acceptance module 60 is used to obtain environmental completion status data after all sub-tasks have been executed, and to compare and accept the environmental completion status data with a preset standard model.

[0266] In one embodiment, the task parsing module 10 is specifically used for:

[0267] Receive task instructions via voice input;

[0268] The task instructions are converted into structured text information using a speech recognition module;

[0269] 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.

[0270] Identify core task objectives based on the aforementioned core semantic elements;

[0271] The core task objective is broken down into multiple sub-task categories based on task type characteristics;

[0272] Execution priority parameters are defined based on the regional association of subtask categories;

[0273] Store the core task objectives, sub-task categories, and execution priority parameters in the task planning database.

[0274] In one embodiment, the task planning module 20 is specifically used for:

[0275] Three-dimensional spatial data of the working environment are acquired through spatial scanning sensors;

[0276] Identify environmental structural features based on the aforementioned three-dimensional spatial data;

[0277] Multiple task areas are divided based on the aforementioned environmental structural characteristics;

[0278] Determine the area parameters and functional requirement parameters for each task area;

[0279] The region weights are determined based on the area parameters and functional requirement parameters.

[0280] Query the capability parameters of multiple agents from the agent capability database;

[0281] Based on the aforementioned capability parameters, a matching score between each agent and each sub-task category is calculated.

[0282] The priority allocation order is generated by sorting the matching scores.

[0283] A global collaboration framework is generated based on the priority allocation order and regional weights.

[0284] In one embodiment, the task allocation module 30 is specifically used for:

[0285] Associate each subtask with its corresponding execution region;

[0286] Acquire real-time state information for each agent, including idle state, energy level, and current location;

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

[0288] Extract the matching score between the agent and the subtask from the global collaboration framework;

[0289] A comprehensive score is determined by combining the matching score, idle state, energy level, and distance parameters.

[0290] 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.

[0291] The task allocation table is stored in the task allocation database.

[0292] In one embodiment, the task execution module 40 is specifically used for:

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

[0294] 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;

[0295] Dynamic environmental data is collected by assigning sensors to intelligent agents;

[0296] The environmental dynamic data is converted into structured execution information;

[0297] The status labels of the structured execution information are marked;

[0298] Upload structured execution information with status tags to the shared information platform in real time;

[0299] Update the work area status in the shared information platform.

[0300] In one embodiment, the dynamic scheduling module 50 is specifically used for:

[0301] Monitor the movement trajectory of intelligent agents in the shared information platform;

[0302] 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;

[0303] The task urgency of agents experiencing path conflicts is determined based on the work area status and task progress in the shared information platform.

[0304] A passage order decision is generated based on the task urgency.

[0305] Based on the passage order decision, low-urgency agents are controlled to suspend movement;

[0306] Plan an avoidance path for the paused intelligent agent and update the space occupancy status of the shared information platform;

[0307] When handling a path conflict event, a movement command containing the avoidance path is sent to the paused agent;

[0308] 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;

[0309] For the faulty intelligent agent in the device failure event, retrieve the idle intelligent agent whose capabilities are matched;

[0310] Obtain the work progress parameters of the faulty task from the shared information platform;

[0311] Determine the remaining workload parameters based on the work progress parameters;

[0312] When handling a device failure event, a new task start command containing the remaining workload parameters is sent to the idle agent;

[0313] An updated task allocation table is generated based on the movement command and the new task initiation command;

[0314] The updated task allocation table is then sent to the corresponding agents.

[0315] In one embodiment, the task acceptance module 60 is specifically used for:

[0316] 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.

[0317] A three-dimensional environment model is generated based on the environmental spatial data;

[0318] Obtain preset standard model data;

[0319] 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.

[0320] An acceptance report containing the coordinates of the discrepancy areas is generated, and the acceptance report is stored in the 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] 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.

[0328] 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.

[0329] 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.

[0330] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0331] 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;

[0332] 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;

[0333] 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.

[0334] 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.

[0335] 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.

[0336] 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.

[0337] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and 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 will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this 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. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0339] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0340] It should be noted that if any software tools or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The above-described embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

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; 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, Acquire work environment information, divide the work environment information into multiple task regions 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, including: 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; Multiple task areas are divided based on the aforementioned environmental structural characteristics; 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. Query the capability parameters of multiple agents from the agent capability database; 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 priority allocation order and regional weights.

4. 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.

5. 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.

6. 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.

7. 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.

8. 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.

9. 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 1-7.

10. 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 1-7.

Images