Cloud edge terminal hierarchical collaborative system and method for lunar surface multi-agent

Abstract

The application discloses a lunar surface multi-agent cloud-edge-end hierarchical cooperation system and method, and relates to the technical field of space robots and distributed intelligent cooperation. In view of the characteristics of limited communication on the lunar surface, uneven distribution of computing power and multi-working condition operation, a unified model containing tasks, resources and constraints is constructed; through natural language analysis and event template instantiation, standardized task semantics suitable for short-period human intervention and long-period unmanned operation conditions are generated; a cloud-edge-end hierarchical cooperation mechanism is established, global strategies are generated on the cloud side, regional schemes are formulated on the edge side, and states are executed and feedback on the end side; based on real-time feedback, closed-loop optimization is implemented, local adjustment is triggered on the edge side, cross-region correction is performed on the cloud side, and dynamic evolution is formed. The application realizes efficient cloud-edge-end cooperation and continuous evolution of multi-agents in a complex lunar surface environment.

Description

Technical Field

[0001] This invention relates to the fields of aerospace robotics and distributed intelligent collaborative technology, and more specifically to a cloud-edge-device hierarchical collaborative system and method for multi-agent systems on the lunar surface. Background Technology

[0002] With the continuous development of lunar research station construction and lunar surface exploration activities, lunar operations have gradually evolved from single-task rover exploration to a complex operational model involving multiple parallel tasks, coupled facilities, and collaborative intelligence agents. Typical lunar surface tasks include rover exploration, sample collection, facility deployment, equipment maintenance, emergency repair, and material transfer. These tasks are typically not completed independently by a single platform but require the collaborative support of multiple types of lunar surface facilities and various intelligent agent platforms. For example, relay communication nodes maintain cross-regional communication links, energy resupply points ensure continuous operational capability, temporary work stations and fixed equipment support specific task execution, and transport and operation vehicles handle regional mobility, resource transfer, and organizational coordination. Therefore, lunar operations are no longer a problem of independent execution by a single task entity but have gradually evolved into a complex problem of collaborative scheduling and resource allocation involving multiple tasks, facilities, and agents.

[0003] Meanwhile, the International Lunar Research Station is positioned as a comprehensive scientific research and experimental base with long-term autonomous operation capabilities and the ability to conduct multidisciplinary scientific research activities. Its operation mode naturally features alternating periods of short-term manned intervention and long-term unmanned operation. During the short-term manned intervention phase, the system needs to quickly understand and respond to the mission instructions issued by the astronauts to form an executable collaborative operation plan. During the long-term unmanned operation phase, the system needs to maintain the continuous execution of tasks within the region, coordinate and allocate resources, and adapt to abnormal situations under limited communication and computing power. This alternating operation mode means that the autonomous lunar research system not only needs to have global overall planning capabilities, but also must have the ability to autonomously plan and continuously operate under conditions of local environmental changes, communication degradation, or even temporary inaccessibility from the cloud.

[0004] Existing multi-agent collaborative scheduling schemes typically employ centralized, manually rule-driven static reasoning methods, which struggle to achieve real-time collaboration and closed-loop optimization among multiple facilities under conditions of limited communication and insufficient edge computing power, resulting in limited system flexibility and overall execution efficiency. Furthermore, traditional single-layer collaborative methods struggle to simultaneously ensure cross-regional global consistency and regional autonomy: when the cloud is unreachable or communication degrades, the system is prone to task interruptions, escalated resource conflicts, unexecutable plans, or increased security risks; conversely, relying entirely on edge autonomy risks the loss of cross-regional resource coordination and global task priority prioritization capabilities.

[0005] Therefore, to address the challenges of resource allocation difficulties and insufficient collaborative capabilities under limited communication conditions on the lunar surface, a cloud-edge-device hierarchical collaborative system and method for multiple agents on the lunar surface needs to be proposed. Summary of the Invention

[0006] In view of the above problems, this invention is proposed to provide a cloud-edge-device hierarchical collaborative system and method for lunar multi-agent systems that overcomes or at least partially solves the above problems. Based on a unified description of tasks, resources and constraints, it breaks through the traditional manual rules and static reasoning methods in human loops, forming an information closed loop of global planning on the cloud side, autonomous arrangement and operation organization on the edge side, and execution and effect feedback on the device side. This enables dynamic evolution and collaborative optimization of task schemes in multi-task parallel scenarios, supporting the long-term reliable operation of the International Lunar Research Station under short-term manned intervention and long-term unmanned operation conditions.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a cloud-edge-device hierarchical collaborative system for lunar multi-agent systems, comprising: The cloud-side layer is built on the high-performance computing platform of the lunar base or lander. It is used to generate cross-regional mission strategies based on standardized mission packages, aggregate global situational data, and generate cross-regional mission strategies. The edge layer, deployed on mobile work platforms or fixed facility nodes with edge computing capabilities, is used to receive the cross-regional task strategy and generate a regional collaborative solution based on the real-time regional situation; and The edge layer, built upon the lunar multi-agent architecture, is used to receive the regional coordination scheme and execute specific actions, generating a summary package containing task progress, resource consumption, and abnormal events. The edge layer and cloud layer are also configured to receive summary packets from the endpoint layer, which are used to trigger a state feedback closed-loop optimization process based on the content of the summary packets. The edge layer performs real-time evaluation of the task execution status within the region based on the summary packets, and performs local task optimization and resource reallocation when an anomaly is detected or the triggering conditions are met. Under the condition that communication is reachable, the cloud layer performs a global re-judgment of the cross-region task progress based on the summary packets returned by the edge layer, performs cross-regional collaborative correction, and reissues the updated task packets and constraint packets to form a closed-loop evolution.

[0009] Secondly, embodiments of the present invention provide a cloud-edge-device hierarchical collaborative method for multi-agent systems on the lunar surface, comprising: S1 constructs a unified scheduling model for multiple elements: it provides unified modeling and structured description of tasks, agents, facility resources, and execution constraints; S2 is based on multi-condition task semantic generation and standardized encapsulation: for short-cycle manned operation conditions, structured tasks are generated through natural language parsing; for long-cycle unattended operation conditions, structured tasks are generated through event triggering and template instantiation; and the structured tasks of the two types of operation conditions are uniformly semantically mapped and encapsulated based on the unified scheduling model established by S1 to form a standardized task package. S3 establishes a cloud-edge-device hierarchical collaboration mechanism: The cloud-side layer receives the standardized task package and generates a cross-regional task strategy based on global situational data; the edge-side layer receives the cross-regional task strategy, generates a regional collaboration scheme in combination with the real-time situation of the region, and sends an execution instruction to the device-side layer; the device-side layer receives the regional collaboration scheme and executes specific actions, generating a summary package containing task progress, resource consumption, and abnormal events. S4 implements state feedback adaptive closed-loop optimization: continuously receiving summary packets uploaded by the edge layer; the edge layer performs real-time evaluation of the task execution status within the region based on the summary packets, and performs local task optimization and resource reallocation when an anomaly is detected or the triggering condition is met; under the condition of communication reachability, the cloud layer performs a global re-judgment of the cross-region task progress based on the summary packets returned by the edge layer, performs cross-regional collaborative correction, and reissues the updated task packets and constraint packets to form a closed-loop evolution.

[0010] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a cloud-edge-device hierarchical collaborative system and method for multi-agent systems on the lunar surface, which has the following effects: 1) In response to the mixed operation mode of short-term manned operation and long-term unmanned operation on the lunar surface, this invention achieves adaptive operation by designing a differentiated feedback judgment mechanism; 2) This invention effectively solves the challenges of limited communication bandwidth on the lunar surface and insufficient computing power on the edge by constructing a three-level distributed computing architecture of cloud-edge-device, and balances the contradiction between cross-regional global coordination and local rapid response. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0012] Figure 1 This is a diagram of the overall architecture of a cloud-edge-device hierarchical collaborative system. Figure 2 This is a diagram illustrating the hierarchical mapping relationship of task packages; Figure 3 Here is the overall flowchart of the cloud-edge-device hierarchical collaboration method; Figure 4 Flowchart for constructing a unified scheduling model for multiple elements; Figure 5 Flowchart for planning tasks with varying timeframes; Figure 6 Flowchart for implementing closed-loop optimization of state feedback. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] The first aspect of the invention, such as Figure 1 As shown, the cloud-edge-device hierarchical collaborative system adopts a three-tier architecture of cloud side, edge side, and device side. Its implementation mainly includes the following steps: S1: System Layering and Deployment Methods In some embodiments, the cloud-edge-device hierarchical collaborative system for multi-agent lunar operations is deployed in three layers: (1) Cloud-side layer: The cloud-side layer consists of a high-performance computing and storage platform on the lander or lunar base, responsible for global situational awareness aggregation and assessment, global task planning, and cross-regional collaborative strategy generation and updating under conditions of cross-regional, multi-task, and multi-facility coupling. The cloud-side layer has high computing power and storage capacity to carry data-intensive and computationally intensive workloads, including but not limited to: multi-source perception and operational data storage, global task situational analysis and judgment, cross-regional task segmentation and coupling constraint management, and global priority and strategy feedback. Specifically, the multi-source perception and operational data storage module is used to receive and store task status, resource status, facility status, communication status, environmental perception data, and historical execution data from the cloud side, edge side, and terminal side, providing a unified data foundation for subsequent global analysis and task planning. The global task situational analysis and judgment module is used to comprehensively evaluate the progress of tasks in each region, resource consumption, facility availability, task conflict risks, and execution deviations, forming a situational judgment result oriented towards the global plan. The cross-regional task segmentation and coupling constraint management module is used to divide the total task into multiple regional task packages and establish correlation constraints between tasks in different regions. The global priority and strategy feedback module updates task priorities, resource quotas, and coordination strategies based on task importance, resource changes, and execution feedback, and distributes the update results to the edge layer and the endpoint layer. In some embodiments, the cloud layer also maintains a system-wide task knowledge base, resource status base, facility capability base, and historical execution experience base to support long-term analysis of task evolution trends, regional coordination relationships, and resource supply and demand relationships, and provides the edge layer with phased priority sequences and cross-regional resource coordination basis.

[0015] Furthermore, the cloud-side layer does not directly undertake all real-time control tasks, but mainly focuses on planning problems that are global, low-frequency, strongly coupled, and computationally complex. This allows for centralized management of global consistency constraints, cross-regional collaborative relationships, and multi-facility sharing constraints, in order to avoid global conflicts arising from local information conditions by edge-side or single-region nodes.

[0016] (2) Edge Layer: The edge layer consists of a mobile operation platform and a fixed infrastructure system with medium computing power and regional information service capabilities. Typical components include: edge computing units carried by lunar transport / operation vehicles weighing hundreds of kilograms or more, and fixed lunar infrastructure nodes (such as relay / communication facilities, energy supply points, temporary operation stations, etc.). As the regional coordination hub, the edge layer is responsible for regional situational integration, regional coordinated scheduling, joint allocation of resources and conflict resolution, and maintaining regional autonomous operation under conditions where the cloud side is inaccessible or communication is limited.

[0017] In some embodiments, the edge layer may specifically include a regional task management unit, a resource allocation unit, a conflict detection and resolution unit, a state aggregation unit, and a communication proxy unit. Specifically, the regional task management unit receives task policies and constraints from the cloud and forms a regional task queue; the resource allocation unit allocates energy, communication, relay, workstations, and transportation capacity based on facility capacity, occupancy status, and task requirements; the conflict detection and resolution unit identifies task timing conflicts, resource occupancy conflicts, and path intersection conflicts, and generates local adjustment results; the state aggregation unit receives and summarizes task progress, resource occupancy, equipment status, and anomaly information reported by edge agents and facilities within the region, unifies the multi-source state data, and forms a regional situational awareness result for use in collaborative scheduling, resource allocation, and conflict resolution; the communication proxy unit implements message forwarding and communication management between the edge and cloud, edge, and other edge nodes, distributes and transmits task packages, constraint packages, resource quotas, and execution summaries, and maintains basic communication and data synchronization when communication is restricted or the cloud is unreachable. Compared to the cloud-side layer, the edge layer focuses more on collaborative scheduling under mesoscale, short-cycle, and real-time changing conditions. It not only undertakes regional deployment after the cloud-side strategy is issued, but also undertakes the function of rapid reorganization in the event of environmental changes, local failures, or the insertion of high-priority tasks, so that the system can still have the ability to operate continuously in the communication-restricted lunar environment.

[0018] (3) Edge-side layer: The edge-side layer consists of small lunar surface agents weighing from a few kilograms to tens of kilograms (small rover / exploration vehicle, robot dog, humanoid robot, micro flying / jumping platform, etc.). The edge-side layer is responsible for collecting on-site environment and equipment status, acquiring local situational awareness and basic analysis, executing task actions and implementing safety closed-loop control, and transmitting key events and summary information back to the edge / cloud side to support collaborative optimization. The edge-side layer still has the ability to maintain the execution of the minimum task set and the satisfaction of safety constraints even when communication degrades.

[0019] In some embodiments, the edge layer may include a task parsing and action generation unit, an environment perception unit, a local safety control unit, a fault detection and degraded operation unit, and a status summary generation unit. The task parsing and action generation unit converts regional-level task instructions issued from the edge into executable action sequences; the environment perception unit collects information on the surrounding environment, terrain obstacles, equipment status, and work objects, and generates local situational awareness results; the local safety control unit executes safety obstacle avoidance, stopping, detour, retreat, or degraded operation strategies in cases of obstacle appearance, increased terrain risk, equipment malfunction, or link interruption; the fault detection and degraded operation unit monitors the operating status of sensors, actuators, energy, and communication modules, identifies anomalies or performance degradation, and switches to the corresponding degraded operation mode; the status summary generation unit encapsulates summaries of location status, task progress, resource consumption, environmental changes, and abnormal events, and transmits them back to the edge and / or cloud.

[0020] The edge layer mainly addresses execution problems related to local, high frequency, strong real-time and security. Its protection focus is not on global task optimization, but on maintaining action-level closed-loop execution and minimum task continuity under conditions of limited computing power, limited energy and unstable communication.

[0021] S2: Cross-layer message interface and data structure In some embodiments, the system uses standardized message interfaces to transmit information between the cloud, edge, and device, including at least the following four types of message objects, to support the closed-loop operation of planning-execution-feedback: 1. TaskPacket: Used to describe the task objective and its structured information, including task identifier, objective description, spatial location / region, time window, priority, task dependencies, resource requirements, completion criteria and feedback strategy; 2. ConstraintPacket: Used to describe task execution constraints, including safety constraints, time window constraints, energy constraints, resource constraints, cooperative constraints, and cross-regional coupling constraints; 3. Resource Quota: Used to describe the facility's resource capacity and allocation status, including resource type, capacity, allocated amount, lock table and release rules, and recovery rate if necessary; 4. SummaryPacket: Used to describe execution feedback and key events, including agent identification, task progress, resource consumption, abnormal events, environmental changes, communication status and timestamps, and may include confidence level markers.

[0022] The aforementioned task packages, constraint packages, resource quotas, and summary packages together constitute a standardized message interface between the cloud, edge, and endpoint sides, used to uniformly describe task content, execution constraints, resource status, and feedback information. Through this standardized interface, cloud-side planning results, edge-side planning results, and endpoint execution status can be incorporated into a unified data representation framework, reducing the complexity of data parsing across heterogeneous platforms.

[0023] In some embodiments, the system defines a hierarchical mapping relationship between cross-layer message objects, such as... Figure 2 As shown, the system includes at least three types of task packages: cloud-side task packages, edge-side task packages, and endpoint task packages. The specific implementation process is as follows: The system first receives the overall task, constraints, and resource quotas, and forms a global planning task on the cloud side. Subsequently, the cloud-side task is sent to the edge side, which refines and reorganizes it based on the real-time regional situation and available resources to form a regional collaboration plan. Next, the edge-side regional collaboration plan is transmitted to the endpoint to form an action execution task. During execution on the endpoint, a state summary is generated and sent back to the edge side, which then summarizes it and sends it back to the cloud side, thereby realizing cross-layer task flow and closed-loop optimization. Among them, the cloud-side task package is used to represent global planning and cross-regional collaboration semantics, the edge-side task package is used to represent regional decomposition and collaborative scheduling semantics, and the endpoint task package is used to represent action execution and security control semantics. The same task can be inherited and transformed between different levels, and constraint packages and resource quotas can also be refined or trimmed as they are transmitted down the layers. Execution feedback can be transmitted back layer by layer according to the task identifier and associated with the upper-level task object.

[0024] S3: System Operating Modes and Responsibility Boundaries In one embodiment, the system defines two typical operating modes: 1. Short-cycle manned intervention mode: Astronauts issue mission instructions through natural language, the cloud side completes mission intent parsing and mission decomposition, and sends the mission constraints to the edge side and end side after consistency verification; In some embodiments, the short-cycle manned intervention mode corresponds to the human-machine collaborative operation state formed during the short-term stay of astronauts at the lunar base during the operational phase, involving activities such as experiments, inspections, deployments, sampling, and maintenance. This mode is characterized by its relatively short duration, relatively concentrated mission activities, high degree of collaboration between humans and multiple intelligent agents, and frequent mission adjustment needs. In this mode, the system mainly focuses on mission understanding, planning, organization, and collaborative execution under human-machine collaborative conditions. The cloud side emphasizes mission intent parsing, global mission planning, and cross-regional collaborative judgment; the edge side emphasizes regional mission organization, collaborative scheduling, and joint resource allocation; and the terminal side emphasizes action execution, safety closure, and status feedback.

[0025] 2. Long-term unattended operation mode: The system enters an autonomous operation mode of "edge-side dominance, cloud-side judgment, and end-side autonomy". The edge side acts as a permanent collaborative hub to optimize and jointly allocate resources based on end-side feedback and regional situation changes. The end side maintains a minimum task set and a security closed loop. When the cloud side is available, it performs a global re-judgment of the stage data and injects back the cross-regional correction strategy.

[0026] In some embodiments, the long-term unattended mode corresponds to the state of long-term autonomous operation of the lunar base and multi-agent system after the astronaut crew returns. In this mode, the system is mainly oriented towards continuous operation and autonomous collaboration under unattended conditions. The edge side is responsible for organizing and dynamically adjusting tasks within the region, the terminal side is responsible for local execution, safety control, and maintaining minimum tasks under limited conditions, and the cloud side is responsible for phased global judgment and cross-regional collaborative correction when communication is available.

[0027] The second aspect of the invention, namely the cloud-edge-device hierarchical collaboration method, is applicable to both short-cycle manned operation and long-cycle unattended operation. Both types of operation uniformly follow the following four core steps: S1 Constructs a unified scheduling model for multiple elements; S2 Task semantic generation and standardized encapsulation based on multiple working conditions; S3 establishes a hierarchical collaboration mechanism between cloud, edge, and endpoint; S4 implements closed-loop optimization of state feedback.

[0028] The overall flowchart of the method is as follows: Figure 3 As shown: S1 is used to establish a unified scheduling model oriented towards tasks, agents, facilities and constraints; S2 is used to generate standardized task semantics for different working conditions and form task packages and constraint packages; S3 is used to complete the hierarchical mapping, distribution and solution of task packages, constraint packages and resource quotas based on the cloud-edge-device hierarchical architecture; S4 is used to trigger edge-side local optimization and cloud-side global correction based on edge-side summary package feedback, thereby forming a closed-loop adaptive dynamic optimization process.

[0029] For ease of explanation and to maintain consistency with the system object in the first aspect of the invention, some general symbols are uniformly agreed upon in this method: Let the task set be:

[0030] Let the set of multi-agent agents on the lunar surface be:

[0031] Let the set of facility resources be:

[0032] Let the set of regions be:

[0033] Among them, the task ( The unified descriptive form of ) can be expressed as:

[0034] In the formula, For task identification, For task type, For the task location or region, For execution time window, As a task priority, For task dependencies, For the set of task capability requirements, This describes the resource requirements for the task. For the mission objective, This serves as a criterion for mission completion. The significance of this unified description format lies in the fact that, regardless of whether the mission originates from astronaut input or is automatically triggered by the system, its core attributes can be expressed within the same data structure, thus providing a unified mission object for subsequent solutions.

[0035] intelligent agent ( The unified descriptive form of ) can be expressed as:

[0036] In the formula, For a collection of capabilities, Current position For the remaining energy, For a healthy state, In communication status, This represents the current task load. This descriptive form enables different types of agents to perform capability matching, load allocation, and path organization within the same planning framework.

[0037] Facilities and resources ( The unified descriptive form of ) can be expressed as:

[0038] In the formula, For facility type, For resource capacity, It is in an available state. It is in an occupied state. This is for restoring or supplementing the rate. Through this approach, various fixed facilities such as relay nodes, power replenishment points, workstations, and docking points can be uniformly incorporated into the resource constraint and allocation framework.

[0039] The specific implementation process of the method of the present invention is described in detail below: S1: Construct a unified scheduling model for multiple elements This step serves as the foundational modeling step for the two types of operating conditions. It is used to construct a unified scheduling model applicable to both types of operating conditions, integrating task objects, agent objects, facility resource objects, and execution constraints into the same planning model, forming a common modeling foundation corresponding to task packages, constraint packages, and resource quotas. For example... Figure 4 As shown, in S1, the system first inputs tasks from different sources into a unified object modeling entry point, and then sequentially executes task and dependency modeling, capability and resource modeling, constraint model construction, and comprehensive goal construction to form a unified scheduling model. Based on this model, the task package base, constraint package base, and resource quota base are further obtained and used as the unified input for cloud-edge-device hierarchical solution.

[0040] S1-1 Task Set and Dependency Modeling In some embodiments, lunar surface operations are uniformly represented as a set of tasks:

[0041] Here, M represents the total number of tasks in the task set. Tasks can include inspection, detection, sampling, deployment, transportation, maintenance, patrol, environmental monitoring, status maintenance, and fault handling, etc. Tasks originate from different sources under different operating conditions, but they must all be mapped to a unified task object when entering the planning process. For tasks with multi-stage execution relationships, a task dependency graph is established:

[0042] in, This is a set of dependency edges, including sequential dependencies, synchronization constraints, and parallel relationships. If the task... Must precede the task If executed, then:

[0043] If task With the task If simultaneous execution is required, then define:

[0044] If tasks can be executed in parallel, then define:

[0045] in, This is a function for determining synchronization relationships. These are parallel relation determination functions, used to characterize the tasks. With the task Synchronization constraints and parallel execution relationships between tasks. Task dependency graph. It is a set of tasks A graph-structured representation of the tasks and their interdependencies, wherein the task set The individual task objects in the graph constitute the task dependency graph. The nodes, therefore, the task in the text ,Task All belong to the task set The specific tasks in the graph are represented by a task dependency graph. The nodes in.

[0046] S1-2 Modeling of Agent Capabilities and Facility Resources In some embodiments, for any intelligent agent Define its relationship to the task The capability matching function is expressed as follows:

[0047] in, Indicates task Required skill set Represents intelligent agents The set of abilities possessed, when When, it indicates an intelligent agent Capable of performing tasks The conditions; when When, it indicates a task. The set of executable candidate tasks that are not included in the agent's set.

[0048] Regarding facilities and resources, the task is set facilities The occupancy is Then the facility capacity constraint is:

[0049] in Indication facilities The maximum resource capacity available per unit of time. Resource capacity can correspond to the number of tasks that can be supported simultaneously, the available power output, the allocable communication bandwidth, or the number of available workstations.

[0050] S1-3 Constraint Model Construction In some embodiments, the task The relevant execution constraints are uniformly constructed as a constraint package. And adopt the following unified structure:

[0051] in, Indicates safety constraints. Indicates time window constraints. Indicates energy constraint. Indicates resource constraints. Represents collaborative constraints. This indicates cross-regional coupling constraints.

[0052] Safety constraints are satisfied:

[0053] in, Represents intelligent agents At any moment The state vector, This indicates the safe and feasible domain for task execution. This safety constraint corresponds to the constraint package... Used to define intelligent agents Execute the task The safe operating boundary at that time.

[0054] The time window constraint is satisfied as follows:

[0055] in, and Representing tasks The start and end times of execution. and These represent the allowed start and end times for task execution, respectively. This time window constraint corresponds to the constraint package... , Depend on , Composition, used to define the task Start-up and end-up times.

[0056] Energy constraints are satisfied:

[0057] in, Represents intelligent agents At any moment The remaining energy, This represents the minimum energy threshold required for the agent to maintain current task execution and return safely. This energy constraint corresponds to the constraint package... Used to determine intelligent agents Are they capable of performing the task? and conditions for safe return.

[0058] Resource constraints are satisfied:

[0059] in, Indicates task facilities The amount of space occupied Indication facilities The maximum resource capacity. This resource constraint corresponds to the constraint package in... Used to define tasks Boundaries for the occupation of facility resources.

[0060] Collaboration constraints are used to define the participating entities, collaborative roles, and temporal boundaries when multiple agents collaborate on a task. These collaboration constraints correspond to the constraints in the constraint package. , It consists of a set of participating agents, cooperative roles, and cooperative temporal boundaries, and is used to determine the task. A multi-agent collaborative execution method.

[0061] Cross-regional coupling constraints: if the task and Shared facilities If a cross-regional continuation relationship exists, then the following definition applies:

[0062] Otherwise, it is 0. This cross-regional coupling constraint corresponds to the constraint package in... , It consists of cross-regional task succession relationships and is used to define the coupled execution relationship between tasks in different regions.

[0063] By unifying the above constraints into a constraint package Subsequently, in S2, the constraint package is bound to the task package and generated. In S3, it serves as the unified constraint basis for cloud-side planning, edge-side scheduling, and end-side execution checks. In S4, it serves as the constraint condition for maintaining the consistency of task logic and the legality of execution during task reordering, resource reallocation, and execution chain repair.

[0064] S1-4 Comprehensive Optimization Objective Construction In some embodiments, the unified synthesis objective function can be expressed as:

[0065] in, Indicates the cost of timeliness, This represents the energy consumption cost, among which, Indicates the first An intelligent agent at time Instantaneous power, This indicates that the agent is in its operating cycle. Internal energy consumption Indicates the cost of resource or facility conflicts. Indicates the risk and cost. This represents the cost of stability in the planning evolution. Indicates task The risk level value is determined by the environmental risk of the task execution area, the health status of the agent, the status of the communication link, and the task security level. Indicates task Priority; Indicates task Adjust the flag variable when the task When the current planning round involves reordering, reallocation, time window shifts, or changes in resource allocation compared to the previous planning round, ,otherwise Under different operating conditions, the weights of each item can be adaptively set according to different requirements: In short-cycle, human-assisted scenarios, response speed and execution efficiency are of paramount importance.

[0066] In long-term unattended operation, greater emphasis is placed on continuity, resource sustainability, and long-term stability.

[0067] At this point, the multi-factor unified scheduling model has been established. This model provides the content source for the generation of subsequent task packages and constraint packages, provides the capacity boundary for the calculation of resource quotas, and provides a unified input object for hierarchical solution at the cloud, edge, and endpoint levels.

[0068] S2: Task semantic generation and standardized encapsulation based on multi-condition operation Although both types of working conditions will eventually generate a unified task package in this step. Constraint Package While both require relevant resources, the original sources of the tasks differ: in short-term manned operations, the tasks primarily originate from the astronauts' natural language commands, while in long-term unmanned operations, they mainly stem from event triggers and template instantiation. Therefore, this step can be implemented in two branches, as follows: Figure 5 As shown, differentiated task generation processes are constructed for short-cycle manned and long-cycle unattended operation scenarios, respectively, and then uniformly enter a hierarchical collaborative planning chain. Specifically: "Determine if there is manned intervention": If there is manned intervention, proceed to the short-cycle manned intervention branch, executing S2-A1 to S2-A4 sequentially; if there is no manned intervention, proceed to the long-cycle unattended operation branch, executing S2-B1 to S2-B4 sequentially. In the short-cycle branch, "Is it plannable?": If plannable, proceed to S2-C; if not plannable, perform constraint adjustments and return to the previous task generation chain. In the long-cycle branch, "Is the threshold met?": If met, proceed to S2-B2; if not, enter continuous monitoring and return to S2-B1. In the long-cycle branch, "Is it plannable?": If plannable, proceed to S2-C; if not plannable, wait for resources. In both branches, "Is replanning?": If replanning is required, proceed to the subsequent replanning processing flow; if not, enter the convergence node and transition to a unified hierarchical planning or replanning closed loop.

[0069] S2-A Short-Cycle Task Semantics Generation under Man-Intervened Conditions In short-cycle manned operations, the mission mainly originates from natural language commands input by astronauts via voice, text, or human-computer interaction terminals.

[0070] S2-A1 Natural Language Task Intent Analysis Let the set of input statements be:

[0071] in, Represents a set of natural language instructions. Indicates the first Input natural language commands, This represents the total number of input natural language commands. The natural language parsing function is defined as:

[0072] in, This represents a natural language parsing function. It is a set of structured semantic elements. For any input statement After analysis, we get:

[0073] in, Represents the task object. Indicates the task action type. Indicates the target location or target area. This indicates time requirements or time window information. Indicates priority requirements. This indicates information about collaborative needs.

[0074] S2-A2 Composite Task Decomposition and Structure Generation The task is decomposed into a set of structured semantic elements, and the decomposition function is defined as follows:

[0075] in, This represents the semantic decomposition function for high-level tasks. This represents a set of structured semantic elements obtained from natural language parsing. This represents the structured task set corresponding to the high-level task obtained after decomposition. The decomposition function divides the high-level task into stages, generates subtasks, and completes dependencies based on the action type, target object, target area, time requirements, and collaboration needs in the task intent. If a task contains multiple stages, it can be represented as:

[0076] in, Indicates the first A high-level task Indicates the first Multiple sub-tasks are obtained after breaking down a high-level task. This indicates the number of subtasks or execution stages corresponding to the high-level task. Based on the generated set of subtasks, a local task dependency structure is constructed according to pre- and post-order, synchronous, and parallel relationships, thereby transforming the semantics of the natural language task into a schedulable structured task object.

[0077] S2-A3 Task Package and Constraint Package Generation In some embodiments, structured tasks are encapsulated as task packages:

[0078] in, Indicates the operating mode of the task. Indicates the feedback and return strategy. Generates task packages. At the same time, not only are task identifiers, task types, task regions, time windows, priorities, resource requirements, and completion criteria written, but the task relationship model established in S1 is also written into the task dependency description field, the capability requirements formed in S1 are written into the capability field, and the task-related resource requirements are written into the resource requirement field. Specifically, with the task... The relevant sets of predecessor tasks, successor tasks, synchronous task identifiers, and parallel task identifiers are written into the task dependency field to form a structured task description that can directly participate in scheduling and solving.

[0079] At the same time, generate the corresponding constraint package in S1-3 for each task. In this step, task packages and constraint packages are associated one by one to form paired encapsulation results of task semantic objects and execution boundary objects. This allows natural language tasks under human intervention conditions to enter the subsequent hierarchical planning chain after semantic parsing, task decomposition and constraint binding.

[0080] S2-A4 Dynamic Priority and Initial Constraint Strategy Generation In generating task packages and its associated constraint package Then, further calculation tasks are performed. At any moment Dynamic priority:

[0081] in, Indicates the basic priority of the task. Indicates the time of the task The urgency Indicates the expected benefits of the task. Indicates the level of risk of the task. , , , These are the corresponding weighting coefficients. It's important to note that the task package... The task priority field is used to carry the task. The priority value at the current planning moment; the dynamic priority calculated in step S2-A4. This is the update result for that field, and it comes with the task package. The hierarchical collaborative planning process of S3 is then initiated. Dynamic priorities are used to sort, compare, and select the best candidate tasks during the global cloud-side planning and local edge-side scheduling processes in S3.

[0082] Meanwhile, based on the constraint package in S1-3 , forming a task The effective constraint state at the initial planning moment. The effective constraint state includes the executable boundaries of safety constraints, time window constraints, energy constraints, resource constraints, coordination constraints, and cross-regional coupling constraints at the current moment, which are used to determine whether the task meets the start-up conditions, resource conditions, and coordination conditions when entering the planning phase.

[0083] Therefore, the task package Before entering S3, in addition to having a unified task semantic description, it also has a sortable dynamic priority attribute and an executable initial constraint boundary, thus providing a unified input for subsequent hierarchical planning and dynamic scheduling.

[0084] S2-B: Task semantic generation under long-term unattended operation conditions In long-term unattended operation, tasks are no longer proposed by astronauts in real time, but mainly come from event triggers and template instantiation.

[0085] S2-B1 Event Detection and Task Triggering Let the event set be:

[0086] in, This represents a set of events occurring under long-term unattended operating conditions. Indicates the first One event, This represents the total number of events in the event set.

[0087] Regarding the event Its eigenvectors are defined as follows:

[0088] Its eigenvectors Indicates the event type, Indicates the location or target area where the event occurred. Indicates the severity of the incident. Indicates the time when the event occurred. Indicates the object associated with the event.

[0089] Define the trigger function:

[0090] Trigger function Used to determine whether an event meets the task triggering conditions, when When, it indicates an event. The severity reaches the corresponding trigger threshold and triggers task generation; when When this occurs, it indicates that the event will not trigger task generation for the time being; among them, Representation and event The corresponding trigger threshold.

[0091] S2-B2 Task Template Matching and Instantiation Let the task template library be:

[0092] in, This represents a task template library for long-term unattended operation conditions. This indicates the first one in the template library. Task template, This represents the total number of task templates. Define the matching function:

[0093] Matching function Used based on event collection With task template library The correspondence between them is used to generate a task set for unattended operation. ,in, This represents the set of tasks generated under long-term unattended operation conditions.

[0094] Regarding the event Generate the corresponding task:

[0095] Indicates the event The first match obtained Task Templates Perform instantiation to generate the corresponding task. ;in, Indicates by the first A task is generated triggered by an event, indicated by the superscript. This indicates that the task falls under the category of long-term unattended operation.

[0096] S2-B3 Task Package and Constraint Package Generation Generate a task package for unattended operation:

[0097] in, Indicates a long-term unattended operating condition. This indicates the feedback and return strategy. This task package inherits the unified task description structure in S1 and includes the task field, resource requirement field, and task relationship field obtained from event triggering and template instantiation.

[0098] Simultaneously generate the corresponding constraint package in S1-3.

[0099] Therefore, the tasks automatically generated under unattended conditions and the tasks parsed under manned conditions are unified into the same task package-constraint package structure at the representation level, and both serve as direct inputs for S3 hierarchical programming.

[0100] S2-B4 Dynamic Priority and Constraint Policy Update In generating task packages and its associated constraint package Then, further calculations were performed on the task under unattended operating conditions. At any moment Dynamic priority:

[0101] in, Indicates the basic priority of the task. Indicates the severity of the incident. Indicates the degree of sustained impact of the event. Indicates the degree of resource impact. These are the corresponding weighting coefficients. It's important to note that the task package... The task priority field is used to carry the task. The priority value at the current planning moment; the dynamic priority calculated in step S2-B4. This is the update result for that field, and it comes with the task package. The hierarchical collaborative planning process in S3 is then initiated. Dynamic priorities are used in S3 to sort, compare, and schedule automatically generated tasks under unattended operating conditions.

[0102] Meanwhile, based on the constraint package in S1-3 , forming a task The effective constraint state at the initial moment of planning.

[0103] Through the above steps, tasks automatically generated under unattended conditions possess unified task semantics, dynamic priority attributes, and initial executable constraint boundaries before entering S3, thus entering the same hierarchical planning chain as task objects under manned conditions.

[0104] S2-C: Hierarchical Semantic Mapping and Unified Encapsulation In both types of working conditions, regardless of the source of the task, a unified task package is generated first. and its associated constraint package Among them, the task package It represents a complete task semantic object abstracted from the task source side, containing all fields such as task objectives, region information, dependencies, resource requirements, operating conditions, and feedback strategies. It is the parent object of subsequent hierarchical mappings.

[0105] Based on this, and considering the differences in responsibilities between the cloud side, edge side, and endpoint side, the task package is... Layered semantic projection and field pruning are performed to further generate three types of hierarchical semantic objects: cloud-side task packages, edge-side task packages, and terminal-side task packages. Cloud-side task packages focus on preserving cross-regional coordination, task coupling relationships, global priorities, and regional resource boundaries; edge-side task packages focus on preserving regional task decomposition results, local scheduling information, resource locking relationships, and collaborative execution conditions; and terminal-side task packages focus on preserving action execution objectives, local security constraints, and execution feedback strategies.

[0106] Cloud-based task package:

[0107] Side mission pack:

[0108] End-side task package:

[0109] satisfy:

[0110] in, Indicates the first The task package corresponding to each task serves as a complete task semantic object abstracted from the task source side; Indicates that it is a task package The cloud-side task package generated by the mapping is used to carry cloud-side semantics such as global planning, cross-regional collaboration, task coupling relationships, global priority, and regional resource boundaries; Indicates a cloud-side task package The generated edge task package is further mapped to carry edge semantics such as regional task decomposition, local scheduling, resource locking relationships, and cooperative execution conditions. Indicates the task package from the side. The generated endpoint task package is further mapped to carry endpoint semantics such as action execution targets, local safety constraints, and execution feedback strategies. Indicates the relationship with the first The constraint package corresponding to each task; Indicates the first associated with this task Description of resource quotas for each region; This represents the mapping function from the overall task package to the cloud-side task package. This represents the mapping function from cloud-side task packets to edge-side task packets. This represents the mapping function from the side-side task packet to the end-side task packet. In the formula, the symbol " This indicates that the lower-level task package is a subset of the upper-level task package, resulting from field inheritance, trimming, refinement, or reorganization. Specifically, the client-side task package is contained within the edge-side task package, the edge-side task package is contained within the cloud-side task package, and the cloud-side task package is contained within the overall task package. Through this unified encapsulation and hierarchical mapping, the overall task package generated by S2... Layered task packages (cloud-side task packages, edge-side task packages, and endpoint-side task packages), constraint packages, and resource quotas together constitute the direct inputs for S3 hierarchical planning and execution control.

[0111] S3: Establish a hierarchical collaboration mechanism between cloud, edge, and endpoint. S3 serves as the main chain shared by the two types of operating conditions, used to realize hierarchical mapping, planning, solving, and execution control of the total task package, hierarchical task package, constraint package, and resource quotas between the cloud side, edge side, and terminal side.

[0112] S3-1 Cloud-based Global Planning Define the cloud-side global planning function as follows:

[0113] in, Indicates the region The results of cloud-side planning, This represents the cloud-side global planning function. This represents a set of cloud-side task packages. This represents the set of constraint packages corresponding to the task. Represents the set of regional resource quotas. This represents the set of cross-regional constraints. The purpose of this planning function is to generate task baselines, resource boundaries, and collaborative constraints for each region based on the global task situation, cross-regional task coupling relationships, and shared facility boundaries. During the cloud-side global planning process, the cloud side receives the overall task package generated by S2. And map it to a cloud-side task package. Then read The dynamic priority field in the task allocation mechanism is used as one of the bases for task sorting and generating regional task baselines. Specifically, the cloud side first excludes tasks that do not yet meet the start conditions based on task dependencies and cross-regional coupling relationships, and then sorts the planarable tasks based on their dynamic priorities. For tasks with higher priorities that meet resource boundaries and cross-regional collaboration conditions, they are prioritized and written into the task baseline set of the corresponding region, and are allocated the corresponding regional resource quotas and collaboration constraints. For tasks with lower priorities or that do not yet meet the constraints, they are retained in the set of tasks to be planned, awaiting subsequent feedback or resource release before entering the next round of planning.

[0114] Set up a region The corresponding cloud-side output is:

[0115] in, Indicates cloud-side output. This indicates that the cloud side is based on the cloud side task package. Generate and distribute to the region The edge task package set is used to carry regional task baselines, dynamic priorities, task dependencies and regional scheduling constraints; Indicates the first Description of resource quotas for each region Indicates the region Cross-regional coordination constraints with other regions. Cloud-side output does not directly generate edge-side actions, but rather provides edge-side task packages, resource upper bounds, and coordination boundaries, enabling edge-side to perform local scheduling without compromising global consistency.

[0116] S3-2 Border Area Planning and Coordinated Scheduling Define the lateral region programming function as follows:

[0117] in, Indicates the region The results of the side-planning Represents the lateral region programming function. Indicates the region Internal cloud-side task package The refined set of edge task packages, Represents the set of associated constraint packages. Indicates the region Resource allocation, The real-time situational awareness of the region includes at least the states of agents, facilities, environment, and communication within the region. The role of the edge-side planning function is to solve for the task execution order, task-agent matching relationships, local resource locking relationships, and conflict resolution results based on the regional task baseline, task constraints, and resource quotas issued by the cloud side, combined with the real-time situational awareness of the region.

[0118] After receiving the edge task package, constraint package, and resource quota from the cloud side, the edge first determines whether the task meets the start conditions based on task dependencies and time windows; then it uses the capability matching function. Candidate agents are screened; resource locking and conflict detection are performed again based on resource constraints and facility capacity constraints; finally, edge-side scheduling results and edge-side executable task packages are generated. During edge-side region planning and collaborative scheduling, the edge side further reads the edge-side task package. The edge uses a dynamic priority field to sort candidate task queues, select conflicting tasks, and determine the scheduling insertion order. Specifically, for candidate tasks that meet task dependencies, time window constraints, capability matching conditions, and resource constraints, the edge forms a regional execution queue according to dynamic priority from high to low. When multiple tasks compete for the same agent, the same facility resource, or the same time window, the edge prioritizes the execution of tasks with higher dynamic priority and adjusts the execution of lower priority tasks by delaying execution, replacing candidate agents, or waiting for resource release. For tasks generated by events in unattended operation, the edge can also insert them into the existing task queue based on the severity of the event and the degree of resource impact, thereby achieving unified scheduling between automatically generated tasks and existing tasks. The edge is also responsible for allocating regional resource quotas. Perform local updates and reallocations to ensure that the local scheduling results remain consistent with the regional resource boundaries.

[0119] S3-3 End-side task execution and action-level closed-loop control The end-side control law can be expressed as:

[0120] in, Represents intelligent agents At any moment The control output, This represents the end-side action execution and safety closed-loop control function. This indicates that the intelligent agent is at time [time]. The local observation state. The role of the edge control law is to transform the edge task packets and associated constraint packets issued by the edge into action-level execution instructions, and to complete path tracking, action execution, safe obstacle avoidance and anomaly degradation handling under local observation conditions.

[0121] End-side reception of task packets from the side The converted end-side task packet After the associated constraint package, only the task fields, local security boundaries and feedback strategy fields directly related to the execution of the ontology are retained, thereby maintaining the action-level closed-loop execution capability under conditions of limited computing power, limited energy and unstable communication.

[0122] S4: Implement state feedback adaptive closed-loop optimization S4 is used to transform the summary packet feedback during edge execution into a replanning basis usable by both the edge and cloud sides, thus forming a closed-loop correction to S3. Its core is not to regenerate an independent task system, but to dynamically correct the local task sequence, task allocation results, resource usage relationships, and cross-regional collaboration boundaries while maintaining the logical consistency of the overall task package, constraint package, and resource quota. For example... Figure 6 As shown, S4, based on the execution process of S3, first generates a summary package containing progress, resources, faults, environment, links, and health information from the edge side, and then aggregates the status to form a regional feedback status from the edge side. Then, depending on the working condition type, it enters either the feedback trigger judgment, edge side rapid replanning, and cloud side global correction process for short-cycle manned working conditions, or the autonomous trigger judgment, edge side autonomous replanning, and cloud side phase correction process for long-cycle unattended working conditions. Subsequently, it adjusts the task order, task allocation, resource consumption, and collaboration boundaries, and injects the updated results back into the task package, constraint package, and resource quota, thus forming a continuously evolving closed-loop replanning process.

[0123] S4-1: End-side summary packet generation and state aggregation In both operating conditions, a summary packet is generated on the endpoint:

[0124] in, Indicates the first The endpoint execution summary package generated by the agent This indicates the progress of the task execution. Indicates resource status information. Indicates fault status information. Indicates environmental status information. Indicates link status information, Indicates health status information, Indicates the time when the summary packet was generated.

[0125] The edge receives digest packets uploaded by each end agent within its region. Subsequently, similar information in the summary package is aggregated and combined with the current execution status within the region to form a regional feedback status. The regional feedback status no longer focuses on the local state of a single edge agent, but reflects the execution deviation, resource changes, fault distribution, environmental disturbances, link degradation, and health risks of the entire region, providing a basis for subsequent edge-side planning and cloud-side correction.

[0126] S4-2A: Feedback Determination and Rapid Replanning under Short-Cycle Manned Operations In short-cycle, human-intervention scenarios, the edge layer determines whether rapid correction is needed based on the regional-level feedback status. (When the summary package...) When issues such as task insertion, execution delays, local conflicts, link fluctuations, changes in high-priority instructions, or abnormal resource usage are reflected, the edge identifies the area as one that needs rapid replanning.

[0127] For regions requiring rapid replanning, the edge agent adjusts the local execution chain based on the current edge task package, relevant constraint packages, regional resource quotas, and the latest summary package. Adjustments include correcting the local task order, adjusting task allocation results, updating resource occupancy relationships, and correcting the collaborative relationships between neighboring agents. During the adjustment process, the edge agent maintains the task dependencies, capability matching relationships, and constraint package boundaries established in S1, preventing local adjustments from affecting the overall execution logic.

[0128] When accessible from the cloud, the edge nodes transmit the accumulated summary packets, regional feedback status, and local correction results accumulated during the phase back to the cloud. Based on the feedback status from multiple regions, the cloud reassesses cross-regional task relationships, cross-regional collaborative constraints, shared resource boundaries, and global collaborative strategies. For content requiring global coordination, the cloud generates correction results and injects them back to the relevant edge nodes, ensuring that the edge's rapid replanning results are consistent with the global execution objectives.

[0129] S4-2B: Autonomous Feedback Judgment and Continuous Replanning under Long-Term Unattended Operations In long-term unattended operation, the side mainly relies on the summary packet. The persistent state changes reflected in the data provide regional-level feedback and determine whether autonomous replanning is necessary. When the summary packet consistently reflects disruptions in execution continuity, degradation of agent health, long-term over-occupancy of shared facilities, risks to resource sustainability, continuous deterioration of the environmental state, or prolonged link unavailability, the edge identifies the region as requiring autonomous replanning.

[0130] For regions requiring autonomous replanning, the edge performs autonomous adjustments based on the current edge task package, relevant constraint packages, regional resource quotas, the most recently issued policy cache from the cloud, and the latest summary package. This autonomous adjustment is used to maintain the continuity and security of operations within the region when the cloud is temporarily unavailable or communication is restricted. Adjustments include lowering the priority of high-risk execution items, adjusting the order of agent operations, reallocating available resources, avoiding faulty or low-health agents, reducing the consumption of non-critical resources, and adjusting internal regional collaboration relationships.

[0131] When the cloud side becomes reachable or enters a phased synchronization window, the edge side uploads the accumulated summary packets, regional feedback status, and autonomous adjustment results to the cloud side. Based on the phased operational status of each region, the cloud side makes phased corrections to cross-regional continuity relationships, shared facility occupancy boundaries, regional resource quotas, and global collaboration relationships. For content that has already been autonomously adjusted by the edge side and has not violated global constraints, the cloud side confirms and synchronizes it; for content that affects cross-regional collaboration or global resource allocation, the cloud side makes unified corrections before distributing them to the relevant edge side nodes.

[0132] S4-3: Closed-Loop Reinjection and Continuous Evolution After edge-focused planning or cloud-side corrections are completed, the update results are injected back into the corresponding task packages, constraint packages, and resource quotas. Task packages contain the objectives and execution objects, constraint packages define execution boundaries, resource quotas limit the scope of shared resource access, and summary packages... Used to carry out feedback status on the end side.

[0133] This closed-loop process does not involve reconstructing a new task system from the original planning object. Instead, it inherits the original task package, constraint package, and resource quota, and continuously modifies the task sequence, task allocation relationships, resource occupancy relationships, and collaborative boundaries based on the execution status reflected in the edge summary package. Thus, the edge summary package, edge region feedback status, edge replanning results, and cloud-side correction results form a continuous information loop, ensuring a clear correspondence between the core modules in the first aspect system and the execution flow in the second aspect method, and enabling the method of this invention to possess the engineering implementation capability for continuous evolution.

Claims

1. A cloud-edge-device hierarchical collaborative method for multi-agent systems on the lunar surface, characterized in that, include: S1 constructs a unified scheduling model for multiple elements: it provides unified modeling and structured description of tasks, agents, facility resources, and execution constraints; S2 is based on multi-condition task semantic generation and standardized encapsulation: for short-cycle manned operation conditions, it generates structured tasks through natural language parsing; for long-cycle unattended operation conditions, it generates structured tasks through event triggering and template instantiation. The structured tasks of the two types of working conditions are then subjected to unified semantic mapping and encapsulation based on the multi-element unified scheduling model established by S1, forming a standardized task package; S3 establishes a hierarchical collaboration mechanism between cloud, edge, and terminal: the cloud side layer receives the standardized task package and generates cross-regional task strategies based on global situational data; The edge layer receives the cross-regional task strategy, generates a regional coordination scheme based on the real-time regional situation, and sends an execution instruction to the endpoint layer; the endpoint layer receives the regional coordination scheme and executes specific actions, generating a summary packet containing task progress, resource consumption, and abnormal events. The steps for establishing a cloud-edge-device hierarchical collaboration mechanism specifically include: The cloud-side layer receives the task package, defines the cloud-side global planning function, and generates task baselines, resource quotas and collaborative constraints for each region based on the global task status, cross-regional task coupling relationships and shared facility boundaries, and then distributes them to the corresponding edge layer. The edge layer receives the task baseline and resource quota from the cloud layer, defines the edge region planning function, takes the task baseline, resource quota and real-time regional situation as input, judges whether the task has the conditions to start based on task dependencies and time windows, sorts and schedules the tasks according to dynamic priority, filters candidate agents according to the capability matching function, performs resource locking and conflict detection according to resource constraints and facility capacity constraints, and generates edge scheduling results and end-side task packages. S4 implements closed-loop optimization of state feedback: continuously receiving summary packets uploaded by the edge layer; the edge layer performs real-time evaluation of the task execution status within the region based on the summary packets, and performs local task optimization and resource reallocation when an anomaly is detected or the triggering condition is met; under the condition of communication reachability, the cloud layer performs a global re-judgment of cross-regional task progress based on the summary packets returned by the edge layer, performs cross-regional collaborative correction, and reissues the updated task packets and constraint packets to form a closed-loop evolution.

2. The method as described in claim 1, characterized in that, The construction of the multi-element unified scheduling model includes: Task Set and Dependency Modeling: Representing Lunar Operations as a Task Set Establish a task dependency graph ,in Define the sequential dependencies, synchronization constraints, and parallel relationships between tasks as a set of dependency edges; Modeling of agent capabilities and infrastructure resources: Defining agents For the task The capacity matching function is defined, and facility resource capacity constraints are also defined. Constraint model construction: This involves designing the task... The execution constraints are uniformly constructed as a constraint package, including security constraints, time window constraints, energy constraints, resource constraints, coordination constraints, and cross-regional coupling constraints; Establish a unified comprehensive objective function: in, Indicates the cost of timeliness, This represents the energy consumption cost, among which, Indicates the first An intelligent agent at time Instantaneous power, This indicates that the agent is in its operating cycle. Internal energy consumption Indicates the cost of resource or facility conflicts. Indicates the risk and cost. This represents the cost of stability in the planning evolution. Indicates task facilities The amount of space occupied Indication facilities Maximum resource capacity, and Representing tasks The start and end times of execution. This represents the risk level value. As a task priority, This indicates a task adjustment flag variable. , , , , These represent the weighting coefficients corresponding to the timeliness cost, energy consumption cost, resource or facility conflict cost, risk cost, and planning evolution stability cost, respectively.

3. The method as described in claim 2, characterized in that, intelligent agent For the task Ability matching function for: in, Indicates task Required skill set Represents intelligent agents The set of abilities possessed, when When, it indicates an intelligent agent Capable of performing tasks The conditions; when When, it indicates a task. The set of executable candidate tasks that are not included in the agent's set.

4. The method as described in claim 2, characterized in that, The steps of generating and standardizing task semantics based on multiple operating conditions, specifically for short-cycle, human-intervention operating conditions, include: Natural Language Task Intent Parsing: Receives natural language instructions and parses them into a set of structured semantic elements based on the natural language parsing function. The semantic elements include task object, task action type, target area, time requirements, priority requirements, and collaborative requirements. Composite task decomposition and structured generation: The set of structured semantic elements is decomposed into tasks to generate a set of structured tasks containing multiple sub-tasks, and the dependencies between the sub-tasks are completed. Task package and constraint package generation: The structured tasks are encapsulated into task packages, and an associated constraint package is generated for each task. Dynamic priority and initial constraint strategy generation: After generating the task package and its associated constraint package, calculate the dynamic priority of the task at the current moment under short-cycle manned operation conditions. in, Indicates the basic priority of the task. Indicates the time of the task The urgency Indicates the expected benefits of the task. Indicates the level of risk of the task. , , , The corresponding weight coefficients are used; simultaneously, based on the constraint model established in S1 and the constraint package, a task is formed. The effective constraint state at the initial moment of planning.

5. The method as described in claim 4, characterized in that, The steps of generating and standardizing task semantics based on multiple operating conditions, specifically for long-term unattended operating conditions, include: Event detection and task triggering: When the detected event feature vector satisfies the preset triggering function, the task generation process is triggered; Matching the event with the pre-set task template library based on the event characteristics, filling the event parameters into the matched task template, and generating the corresponding task instance; Task package and constraint package generation: Generate task packages for unattended operation and generate corresponding constraint packages based on preset rules; Dynamic Priority and Constraint Policy Update: After generating the task package and its associated constraint package, calculate the dynamic priority of the task at the current moment under unattended operation conditions. in, Indicates the basic priority of the task. Indicates the severity of the incident. Indicates the degree of sustained impact of the event. Indicates the degree of resource impact. The corresponding weight coefficients; simultaneously, based on the constraint model established in S1 and the constraint package, a task is formed. The effective constraint state at the initial moment of planning.

6. The method as described in claim 1, characterized in that, The structured tasks for the two types of working conditions are semantically mapped and encapsulated based on the multi-element unified scheduling model established by S1, forming a standardized task package including: Perform hierarchical semantic projection and field pruning on the task packages to generate cloud-side task packages, edge-side task packages, and terminal-side task packages: Cloud-based task package: Side mission pack: End-side task package: satisfy: in, Indicates the first The task package corresponding to each task Indicates that it is a task package The cloud-side task package generated by mapping, Indicates a cloud-side task package Further mapping generates the edge task package, Indicates the task package from the side. Further mapping of the generated edge task packets, Indicates the relationship with the first The constraint package corresponding to each task; Indicates the first associated with this task Description of resource quotas for each region; This represents the mapping function from the overall task package to the cloud-side task package. This represents the mapping function from cloud-side task packets to edge-side task packets. This represents the mapping function from the side-side task packet to the end-side task packet.

7. The method as described in claim 6, characterized in that, The steps for establishing a cloud-edge-device hierarchical collaboration mechanism specifically include: Define the cloud-side global planning function as follows: in, Indicates the region The results of cloud-side planning, This represents the cloud-side global planning function. This represents a set of cloud-side task packages. This represents the set of constraint packages corresponding to the task. Represents the set of regional resource quotas. Represents a cross-region constraint set; Define the lateral region programming function as follows: in, Indicates the region The results of the side-planning This represents the edge-side region programming function. Indicates the region Internal cloud-side task package The refined set of edge task packages, Represents the set of associated constraint packages. Indicates the region Resource allocation, It represents the real-time situation of the region, including at least the status of agents, facilities, environment, and communication within the region; The edge layer receives the edge task packet, defines the edge control law, executes the action-level control law according to the associated constraint packet, and completes path tracking, action execution, safety obstacle avoidance and anomaly degradation processing under local observation conditions. The end-side control law is expressed as: in, Represents intelligent agents At any moment The control output, This represents the end-side action execution and safety closed-loop control function. This indicates that the intelligent agent is at time [time]. The local observation status, Indicates the relationship with the first The constraint package corresponding to each task Indicates the task package from the side. The generated endpoint task package is further mapped.

8. The method as described in claim 1, characterized in that, The steps for implementing state feedback closed-loop optimization, specifically under short-cycle manned intervention conditions, include: The edge-side agent generates an edge-side execution summary package: in, Indicates the first The endpoint execution summary package generated by the agent This indicates the progress of task execution. Indicates resource status information. Indicates fault status information. Indicates environmental status information. Indicates link status information, Indicates health status information, Indicates the time when the digest packet was generated; The summary packets of all agents within the edge-side layer convergence region receive regional-level feedback status. In short-cycle scenarios with human intervention, the side layer determines whether rapid correction is needed based on the regional feedback status. Under the condition that communication is accessible, the summary packets accumulated during the phase, the regional feedback status, and the local correction results are transmitted back to the cloud side layer.

9. The method as described in claim 1, characterized in that, The steps of implementing state feedback closed-loop optimization specifically include the following under long-term unattended operation conditions: The edge-side agent generates an edge-side execution summary package: in, Indicates the first The endpoint execution summary package generated by the agent This indicates the progress of task execution. Indicates resource status information. Indicates fault status information. Indicates environmental status information. Indicates link status information, Indicates health status information, Indicates the time when the digest packet was generated; The summary packets of all agents within the edge-side layer convergence region receive regional-level feedback status. In long-term unattended operation, the side layer determines whether autonomous replanning is needed based on the regional feedback status. Under conditions where communication is available, the summary packets accumulated during the phase, regional feedback status, and autonomous adjustment results will be uploaded to the cloud side layer.

10. A cloud-edge-device hierarchical collaborative system for lunar multi-agent systems, used to implement the method described in any one of claims 1-9, characterized in that, include: The cloud-side layer, built upon the high-performance computing platform of the lunar base or lander, is used to aggregate global situational data based on standardized mission packages and generate cross-regional mission strategies. The edge layer is deployed on mobile work platforms or fixed facility nodes with edge computing capabilities to receive the cross-regional task strategy and generate regional collaborative solutions in combination with the real-time regional situation. as well as The edge layer, built upon the lunar multi-agent architecture, is used to receive the regional coordination scheme and execute specific actions, generating a summary package containing task progress, resource consumption, and abnormal events. The edge layer and cloud layer are also configured to receive summary packets from the endpoint layer, which are used to trigger a state feedback closed-loop optimization process based on the content of the summary packets. The edge layer performs real-time evaluation of the task execution status within the region based on the summary packets, and performs local task optimization and resource reallocation when an anomaly is detected or the triggering conditions are met. Under the condition that communication is reachable, the cloud layer performs a global re-judgment of the cross-region task progress based on the summary packets returned by the edge layer, performs cross-regional collaborative correction, and reissues the updated task packets and constraint packets to form a closed-loop evolution.

Images