Multi-scene collaborative application-oriented ubiquitous intelligent AI agent task scheduling method

Abstract

The present application relates to the technical field of agent scheduling, in particular to a ubiquitous intelligent AI agent task scheduling method for multi-scene collaborative application, comprising the following steps: deploying cloud, edge and terminal three-level ubiquitous AI agents, collecting multi-scene environment data and computing power margin information in real time to construct a production situation atlas; decoupling unmanned aerial vehicle production orders into subtask streams with time sequence dependency, and extracting their computing load characteristics and time constraint parameters; based on this, triggering a multi-level resource allocation mechanism, and dividing the decoupled tasks into local operation tasks and global collaborative subtasks. The present application assigns local tasks to the terminal and edge joint processing domain for execution of in-situ scheduling with edge flexibility unloading mechanism, at the same time, the cloud generates cross-scene collaborative scheduling strategy for the main control global task and decomposes and issues it, effectively breaking the single-point computing power bottleneck and cross-domain collaborative data island, realizing precise matching of computing resources and high space-time consistency of production process.

Description

Technical Field

[0001] This invention relates to the field of intelligent agent scheduling technology, specifically to a ubiquitous intelligent AI agent task scheduling method for multi-scenario collaborative applications. Background Technology

[0002] With the deepening development of intelligent manufacturing, the production mode of high-precision and complex products such as drones is shifting from traditional assembly line operations to cross-scenario, flexible, and intelligent collaboration. In a typical drone assembly process, multiple physical scenarios are often involved, such as warehousing scheduling, parts distribution, and final assembly and debugging. The efficiency of collaboration between these scenarios directly determines the production cycle and product quality. However, existing manufacturing scheduling technologies face the following technical challenges in multi-scenario collaborative applications: Traditional scheduling schemes generally adopt a centralized polling execution mode, requiring all tasks to be reported to the cloud for processing, making it impossible to achieve dynamic collaboration of computing power between terminals and the edge; at the same time, cross-scenarios such as warehousing and final assembly lack a unified time-series alignment mechanism, which easily leads to cross-scenario asynchrony, waiting, and time-series errors; and they cannot dynamically allocate resources based on task load and real-time computing power, resulting in wasted computing resources and high response latency. Most existing scheduling schemes are based on preset static rules, making it difficult to achieve unified real-time modeling of cross-scenario physical environments and ubiquitous cloud-edge-device computing resources. When facing task flows with complex time-series dependencies and strict time constraints, they cannot perform fine decoupling and multi-level resource matching; physical assembly and logical scheduling lack closed-loop evolution, and cannot achieve flexible allocation and adaptive correction of computing power in the event of sudden demands or assembly deviations, which seriously restricts the efficiency and quality stability of cross-scenario collaborative production.

[0003] To address this, a ubiquitous intelligent AI agent task scheduling method for multi-scenario collaborative applications is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a ubiquitous intelligent AI agent task scheduling method for multi-scenario collaborative applications, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A ubiquitous intelligent AI agent task scheduling method for multi-scenario collaborative applications includes: By deploying ubiquitous AI agents at the cloud, edge, and terminal levels in warehousing, assembly, and related scenarios, real-time environmental data from multiple scenarios is collected. Simultaneously, the computing power reserve information of the computing resource node pool composed of hardware carriers at each level of the AI ​​agents is extracted to construct a production status map. When receiving drone production orders, the orders are decoupled into sub-task flows with time-dependent relationships based on the production status map, and the computational load characteristics and time constraint parameters of each sub-task flow are extracted. A multi-level resource allocation mechanism is triggered based on computational load characteristics and time constraints. This mechanism divides the decoupled subtask flow into local task tasks and global collaborative subtask tasks. The local task tasks are assigned to a joint processing domain composed of terminal execution agents and edge coordination agents for in-situ scheduling. The global collaborative sub-tasks are assigned to the cloud-based master control agent to generate a collaborative scheduling strategy. The cloud-based master control agent decomposes the collaborative scheduling strategy into cross-scenario execution instructions and sends them to the edge coordination agent and terminal execution agent in the corresponding scenario for collaborative operation.

[0006] Preferably, the multi-scenario environmental data includes warehousing scenario data, final assembly scenario data, and cross-scenario collaborative data; the warehousing scenario data includes material inventory balance, material storage location, and real-time pose of handling equipment; the final assembly scenario data includes the occupancy status of each workstation, the operating parameters of the assembly robot, and the real-time spatial pose of UAV components and the geometric accuracy of connection points; the cross-scenario collaborative data includes the occupancy status of logistics channels between scenarios, and the communication bandwidth and latency parameters between intelligent agent nodes at all levels, and a production status map is constructed by combining the multi-scenario environmental data; The ubiquitous intelligent AI agent includes a cloud-based master control agent, an edge coordination agent, and a terminal execution agent.

[0007] Preferably, the production status map construction process involves performing unified time-stamped synchronization processing on the multi-scenario environmental data and performing a normalization transformation of the spatial coordinate system to generate a structured environmental dataset; defining physical entity nodes based on the material storage location and workstation occupancy status, and attaching the corresponding material type and equipment number as static attribute parameters to the physical entity nodes; calculating the transmission weight of the logical link based on the communication bandwidth and latency parameters between each level of intelligent agent nodes, and simultaneously establishing physical connection relationships with passage cost weights between physical entity nodes based on the occupancy status of the logistics channel; injecting real-time collected data on the pose of handling equipment, the operation parameters of assembly robots, and the spatial pose and geometric accuracy data of UAV components as dynamic feature quantities into the physical entity nodes, and using the computing power margin information of each level of intelligent agent hardware carrier as the computing power carrying parameters of the node itself.

[0008] Preferably, the specific process of decoupling the order into a subtask flow with temporal dependencies involves parsing the product structure tree and process path in the UAV production order and mapping them to a preset set of atomic work units; establishing temporal connections between work units based on the physical assembly logic and process sequence between each atomic work unit, defining the in-degree constraints and out-degree triggering conditions of each node; retrieving the corresponding material storage location, handling equipment pose, and workstation occupancy status from the production status map, converting physical space distance and equipment availability into the pre-execution cost of each work unit, dynamically adjusting the weights of logical connection edges in the directed acyclic graph, and generating a subtask flow with temporal dependencies.

[0009] Preferably, the process of extracting the computational load characteristics and time constraint parameters of each subtask flow involves traversing the subtask flow, quantifying the peak floating-point operation volume and memory throughput requirements of each atomic task unit based on the atomic task unit, generating a load feature vector, analyzing the topology of the subtask flow using the critical path method, and combining the communication latency between agents at all levels in the production status map and the order delivery deadline to calculate the latest start time and time slack of each atomic task unit, thereby generating time constraint parameters.

[0010] Preferably, the multi-level resource allocation mechanism extracts the time slack from the time constraint parameters of each atomic work unit, and obtains the bidirectional link communication latency between the current scene node and the cloud master control agent based on the production status map; performs three-dimensional joint decision: if the time slack is less than or equal to the bidirectional link communication latency, and the corresponding atomic work unit does not involve workstation interaction that crosses physical boundaries, and the peak floating-point operation volume and memory throughput requirement in its load feature vector do not exceed the sum of the real-time computing power margin of the terminal execution agent and the edge coordination agent in the current scene, it is confirmed as a local work task that meets the requirements of immediate response; The remaining atomic task units that do not meet the local task determination conditions are identified as global collaborative subtasks.

[0011] Preferably, the in-situ scheduling involves the terminal execution agent calling its own hardware computing power to parse the local task and comparing the real-time monitored instantaneous available computing power with the peak floating-point operation volume of the local task. When the instantaneous available computing power is less than the peak floating-point operation volume, the computational load of the local task is divided into locally retained load and unloaded load based on the difference in computing power. The terminal execution agent processes the locally retained load and transmits the unloaded load to the edge coordination agent in the same scene via the local area network. The terminal execution agent and the edge coordination agent respectively parse the load they carry and issue control commands to the corresponding physical entity nodes.

[0012] Preferably, the collaborative scheduling strategy involves the cloud-based master control agent mapping the global collaborative sub-tasks to the production status map, extracting the physical connection relationships and logical link transmission weights across warehousing and final assembly scenarios; sorting the global collaborative sub-tasks by spatiotemporal nodes based on the extracted physical connection relationships and logical link transmission weights, generating a collaborative scheduling strategy that includes cross-domain transfer paths and trigger timestamps for each node; segmenting the collaborative scheduling strategy along the scene physical boundaries of the cross-domain transfer paths to generate local execution sequences belonging to the corresponding scenarios; extracting timestamps at the cross-scenario physical handover nodes to generate timing synchronization tokens, encapsulating and combining the local execution sequences and the timing synchronization tokens into the cross-scenario execution instructions, and issuing them to the edge coordination agent of the corresponding scenario.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention addresses the data fragmentation problem inherent in traditional static or centralized architectures by deploying a three-tiered ubiquitous intelligent agent across cloud, edge, and device levels and collecting real-time physical data and hardware computing power reserves from multiple scenarios to construct a production status map. Combined with directed acyclic graph decoupling and multi-level parameter extraction for drone production orders, it unifies the time-stamping of computational load characteristics across physical environments such as equipment status and spatial distances in warehousing and final assembly with logical layers. This provides objective and quantifiable decision support boundaries for task scheduling, avoiding resource mismatch issues caused by opaque environmental conditions and the black-box nature of computing resources during complex order execution across scenarios.

[0014] 2. This invention employs a three-dimensional joint decision-making mechanism based on time constraints, physical boundaries, and computing power capacity limits to precisely divide the decoupled task flow into two categories: local operations and global collaboration. Furthermore, it introduces an edge-based elastic offloading mechanism within the local in-situ scheduling. When the underlying terminal encounters a sudden surge in computing power while processing local control tasks, it slices the excess computing load in real time based on the instantaneous computing power difference and directionally transfers it to the edge-coordinated intelligent agent in the same scenario for parallel parsing and instruction compensation. This elastic mechanism breaks the physical resource constraints of single-point hardware, ensuring the immediate responsiveness of the underlying closed-loop control when dealing with sudden high-concurrency production demands and preventing assembly sequence disruptions caused by local computing power overload.

[0015] 3. This invention uses a cloud-based master control agent to breakpoint the collaborative scheduling strategy along the boundaries of physical transfer paths in multiple scenarios, and extracts the timestamps of cross-domain handover nodes to generate time synchronization tokens. This constructs a decentralized triggering architecture for cross-scenario material and data handover, giving edge and terminal agents at all levels the ability to autonomously wake up and align their actions within a specific time interval. This changes the limitation of traditional scheduling models that must rely on the cloud center for full lifecycle micro-intervention. When dealing with sudden working conditions with assembly accuracy deviations or local latency fluctuations, it ensures the spatiotemporal consistency of the cross-domain collaborative process and the smooth connection of production rhythms between multiple workstations. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a ubiquitous intelligent AI agent task scheduling method for multi-scenario collaborative applications; Figure 2 This is a schematic diagram of the three-dimensional joint decision-making mechanism for multi-level resource allocation in this invention; Figure 3 This is a schematic diagram of the local in-situ scheduling end-edge load elastic splitting mechanism of the present invention. Detailed Implementation

[0017] 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.

[0018] Example 1 Please see Figure 1 This invention provides a ubiquitous intelligent AI agent task scheduling method for multi-scenario collaborative applications, the technical solution of which is as follows: A ubiquitous intelligent AI agent task scheduling method for multi-scenario collaborative applications includes: By deploying ubiquitous AI agents at the cloud, edge, and terminal levels in warehousing, assembly, and related scenarios, real-time environmental data from multiple scenarios is collected. Simultaneously, the computing power reserve information of the computing resource node pool composed of hardware carriers at each level of the AI ​​agents is extracted to construct a production status map. When receiving drone production orders, the orders are decoupled into sub-task flows with time-dependent relationships based on the production status map, and the computational load characteristics and time constraint parameters of each sub-task flow are extracted. Based on the characteristics of the computational load and the time constraint parameters, the matching weight between subtasks and each computing resource node is determined, triggering a multi-level resource allocation mechanism: subtasks that meet the requirements of immediate response are assigned to the joint processing domain composed of terminal execution agents and edge coordination agents for in-situ scheduling; global collaborative subtasks are assigned to the cloud master control agent for cross-scenario planning and to generate collaborative scheduling strategies. For the aforementioned global collaborative sub-tasks, the cloud-based master control agent and the agents in each scenario drive the production status map to perform simulation evolution according to the collaborative scheduling strategy, dynamically reconstruct the sub-task flow based on the evolution results, and re-trigger the multi-level resource allocation mechanism.

[0019] The multi-scenario environmental data includes warehousing scenario data, final assembly scenario data, and cross-scenario collaborative data. The warehousing scenario data includes material inventory balance, material storage location, and real-time pose of handling equipment. The final assembly scenario data includes the occupancy status of each workstation, the operating parameters of the assembly robot, and the real-time spatial pose of UAV components and the geometric accuracy of connection points. The cross-scenario collaborative data includes the occupancy status of logistics channels between scenarios and the communication bandwidth and latency parameters between intelligent agent nodes at all levels. The multi-scenario environmental data is combined to construct a production status map. Specifically, in the warehousing, final assembly, and logistics channels connecting the two in the drone production base, a three-level ubiquitous AI intelligent agent network consisting of cloud, edge, and terminal components is deployed. Terminal intelligent agents are deployed on material handling AGVs, assembly robots, and drone chassis, responsible for performing perception and basic operations. Edge intelligent agents are deployed on edge servers in each workshop, responsible for logical coordination within the area. Cloud-based master control intelligent agents are deployed in the factory data center, responsible for global optimization. All intelligent agent nodes are connected to a unified industrial internet to ensure real-time data synchronization.

[0020] Data is captured through terminal intelligent agents and sensor networks within the warehousing scenario; the remaining material inventory of various drone parts is obtained in real time through weighing sensors or barcode readers on the material racks; visual sensors and LiDAR on the top of the warehouse are used in combination with SLAM algorithms to determine the specific material storage location of the parts on the shelves; the real-time pose of the handling equipment in the warehouse is transmitted back in real time through the wheeled odometers and inertial measurement units on the AGVs; the above data is aggregated by edge intelligent agents to form static and dynamic data streams in the warehousing scenario. By using pressure sensors and industrial control computer status codes at each workstation, the occupancy status of each workstation (such as idle, in assembly, or equipment failure) is identified; the assembly robot's operating parameters, including the current robotic arm speed, torque output, and real-time trajectory of the end effector, are directly retrieved from the robot's controller; and a high-precision vision tracking system deployed around the workstation is used to capture the real-time spatial pose of the drone components, and a laser rangefinder is used to detect the geometric accuracy of the connection points of each component. To ensure the smooth flow of materials between scenarios, the interaction status between scenarios is monitored in real time. Photoelectric sensors and infrared arrays deployed in the passageways connecting the warehouse and the final assembly workshop are used to monitor the occupancy of logistics channels between scenarios (such as whether there are obstacles or passing vehicles). At the same time, network monitoring tools are used to detect the communication bandwidth and latency parameters between intelligent agent nodes at all levels (terminals, edge, cloud) in real time, and a production status map is constructed by combining multi-scenario environmental data.

[0021] By deeply deploying ubiquitous AI agents at the cloud, edge, and terminal levels across multiple scenarios, multi-dimensional real-time monitoring of warehouse material status, final assembly production equipment, and cross-scenario communication environments has been achieved. The physical layer hardware carrier is directly transformed into intelligent nodes with proactive sensing capabilities, effectively breaking down the perception barriers between different work spaces and equipment levels. This ubiquitous deployment architecture eliminates the fragmentation of underlying data and information silos, and realizes real-time synchronization and unified aggregation of massive heterogeneous data across scenarios.

[0022] The production status map construction process involves uniformly synchronizing the multi-scenario environmental data with a unified time scale and performing a normalization transformation of the spatial coordinate system to generate a structured environmental dataset. Based on the material storage location and workstation occupancy status, physical entity nodes are defined in the map topology, and the corresponding material type and equipment number are attached to the physical entity nodes as static attribute parameters. The transmission weight of the logical link is calculated based on the communication bandwidth and latency parameters between each level of intelligent agent nodes. Simultaneously, physical connections with passage cost weights are established between physical entity nodes based on the occupancy status of the logistics channel. Real-time collected data on the pose of handling equipment, the operation parameters of assembly robots, and the spatial pose and geometric accuracy data of UAV components are injected into the physical entity nodes as dynamic features, and the computing power margin information of each level of intelligent agent hardware carrier is used as the node's own computing power carrying parameters.

[0023] Specifically, the system performs unified time-stamp synchronization processing on real-time collected warehouse scene data, final assembly scene data, and cross-scene collaborative data. Using the industrial Ethernet clock synchronization protocol (precision time protocol in this embodiment), the sampling times of all sensors, handling equipment, and assembly robots are aligned to ensure consistency of multi-source data in the time dimension. Subsequently, the system performs a spatial coordinate system normalization transformation, converting the warehouse local coordinate system, the final assembly local coordinate system, and the coordinate systems of each mobile device to the factory's global three-dimensional geographic coordinate system, generating a structured environmental dataset. Based on the material storage locations and workstation occupancy status in the structured environment dataset, physical entity nodes are defined in the graph topology. In this embodiment, each determined storage location coordinate and each workstation that is in operation or waiting is abstracted as a logical node; the corresponding material type (such as drone wings, fuselage, power modules, etc.) and equipment number (such as assembly robot arm No. 1, inspection instrument No. 2) are used as static attribute parameters and are attached to the corresponding physical entity nodes in the form of key-value pairs, thus initially forming a topology framework that reflects the static layout and resource distribution of the factory; Each terminal intelligent agent exchanges 512-byte probe data packets with the edge intelligent agent and with the cloud intelligent agent periodically (every 100ms in this embodiment). The communication latency parameters between intelligent agent nodes at each level are obtained by recording the time difference between the sending and receiving of data packets. At the same time, each intelligent agent node calls the traffic monitoring interface of its network adapter in real time to read the current real-time throughput and obtain the communication bandwidth. Based on the acquired communication bandwidth and latency parameters between intelligent agent nodes at each level, the transmission weight of the logical link is calculated using a normalized weighted algorithm. In this embodiment, the bandwidth allocation ratio is preset to 0.3, and the latency allocation ratio is preset to 0.7. The emphasis is placed on the latency parameter because the control commands for UAV assembly tasks are typically lightweight messages, requiring less bandwidth, but demanding extremely high synchronization and real-time performance. Latency fluctuations directly affect assembly accuracy and collaboration efficiency. Specifically, the real-time collected bandwidth value is divided by the baseline bandwidth limit (set to 1000 Mbps in this embodiment) to obtain the normalized bandwidth score. Simultaneously, the difference between the baseline latency limit (set to 100 ms in this embodiment) and the real-time latency value is subtracted and divided by this baseline limit to obtain the normalized latency score. The product of the bandwidth score multiplied by 0.3 and the latency score multiplied by 0.7 is then weighted and summed to calculate a quantized weight value between 0 and 1. Simultaneously, physical connections with passage cost weights are established between physical entity nodes based on the occupancy status of logistics channels. Specifically, the ratio of the area of ​​moving obstacles within the channel to the total area of ​​the channel, as reported by the factory's visual monitoring equipment in the cross-scenario collaborative data, is extracted to obtain the channel congestion percentage as the occupancy status of the logistics channel. It is then determined whether there is a physical passage path between any two physical entity nodes in the map. If so, a physical connection edge (i.e., a physical connection relationship) is established between these two physical entity nodes, and the preset base physical distance value (set to a base distance of 5 meters in this embodiment) is multiplied by the channel congestion percentage. The result is used as the passage cost weight for that road segment and attached to the corresponding physical connection edge. The real-time collected pose data of the handling equipment, the operation parameters of the assembly robot, and the spatial pose and geometric accuracy data of the UAV components are injected into the physical entity nodes as dynamic feature quantities. In this embodiment, warehousing and final assembly scene data are extracted, the three-dimensional coordinates and real-time yaw angle of the automated guided vehicle are used as the pose data of the handling equipment, the real-time output torque of each joint of the assembly robot and the currently executed action command number are used as the operation parameters of the assembly robot, and the six-degree-of-freedom coordinate parameters of the UAV component to be assembled and the assembly seam deviation value fed back by the sensor scan are used as the spatial pose and geometric accuracy data of the UAV component; (this embodiment refreshes every 50ms), and the above dynamic data is written to the dynamic attribute fields of the corresponding physical entity nodes in the graph in real time; Artificial general-purpose agents deployed at various levels in the cloud, edge, and terminal periodically call the system interface of the underlying hardware resource manager to read in real time the current number of available idle cores and utilization rate of the central processing unit, the remaining available video memory space of the graphics processor, and the current available megabytes of system memory. This information is encapsulated into a structured set of computing power reserve information, bound to and written into the attribute pool of the physical entity node representing the corresponding agent in the graph, and used as the computing power carrying parameter of the node to quantify the upper limit of the task concurrent processing capability of each node at the current moment. By constructing a production status map, not only is the factory layout statically reconstructed, but the real-time boundaries of communication networks, logistics channels, and computing resources are also dynamically quantified, providing a solid and panoramic data foundation for the precise decoupling of tasks, conflict-free scheduling across scenarios, and optimal allocation of resources for multi-level AI agents.

[0024] Upon receiving a drone production order, the order is decoupled into a sub-task flow with temporal dependencies based on the production status map. The specific process of decoupling the order into a sub-task flow with temporal dependencies involves parsing the product structure tree and process path in the drone production order and mapping them to a preset set of atomic work units. Based on the physical assembly logic and process sequence between each atomic work unit, a directed acyclic graph model is used to establish the temporal connection relationship between the work units, and the in-degree constraints and out-degree triggering conditions of each node are defined. The system retrieves the corresponding material storage locations, handling equipment positions, and workstation occupancy status from the production status map, transforms physical space distance and equipment availability into the pre-execution cost of each work unit, dynamically adjusts the weights of logical connection edges in the directed acyclic graph, and generates subtask flows with temporal dependencies.

[0025] After receiving a production order, the cloud-based main control intelligent agent extracts the hierarchical component details from the order's product structure tree. In this embodiment, the hierarchical component details include the bottom layer fuselage chassis, the middle layer power motor, and the top layer rotor blades. Simultaneously, it extracts the manufacturing process flow standards specified in the process path. Then, it retrieves a preset set of atomic operation units from the system. In this embodiment, the preset set of atomic operation units is set to include five basic operation types: material gripping, cross-area handling, vision alignment, bolt tightening, and power-on detection. Based on the parsed component details and manufacturing process flow standards, the overall macro-manufacturing requirements of the order are decomposed one by one and mapped to the aforementioned five basic operation types, thereby generating a specific set of atomic operation units for this particular UAV order. Based on the physical assembly logic and process sequence between each atomic work unit, a directed acyclic graph (DAG) model is used to establish the temporal connection relationship between work units, and the in-degree constraints and out-degree triggering conditions of each node are defined. Specifically, in constructing the DAG model, each atomic work unit generated by the mapping is defined as a network node. The interdependencies in physical assembly (e.g., the chassis must be moved before the power motor can be installed) and the sequence of operations specified in the process documents are transformed into connection segments with specific directions between network nodes. At the same time, in-degree constraints and out-degree triggering conditions are defined for each network node. In this embodiment, the in-degree constraint adopts a multi-condition AND logic judgment, that is, all predecessor nodes of the current atomic work unit must feed back completion status signals to the scheduling system (status signal types include: execution success status code = 1, execution failure status code = 0, waiting timeout status code = -1), and the start condition of the current node can only be triggered when the status codes of all predecessor nodes are 1. The out-degree trigger condition automatically broadcasts an activation command to all its successor nodes after the current atomic job unit has been executed; if the execution result status code is 0 or -1, the activation command to subsequent nodes is skipped, and an abnormal alarm is reported to the cloud master control agent. The system accesses the production status map in real time, accurately extracting the three-dimensional physical coordinates of the target material involved in the current atomic work unit, the real-time three-dimensional spatial coordinates and yaw angle of the assigned handling equipment, and the current real-time status of the target assembly station (in this embodiment, the station status is divided into idle, occupied, or stopped) from physical entity nodes and dynamic feature quantities. It calculates the straight-line physical distance between the material storage location and the position of the handling equipment, as well as the path distance from the handling equipment to the target station corresponding to the station's occupied state. In this embodiment, the physical spatial distance is converted into a time cost by multiplying the sum of the above two physical distance values ​​by a preset unit distance time base (set to 3m / s in this embodiment) to obtain the basic spatial time cost value. Simultaneously, equipment availability is converted into a waiting cost. If the station's occupied state is idle, the waiting cost value is recorded as zero; if it is occupied, the estimated remaining operation time value of the current task being executed at that station is extracted as the waiting cost. The basic spatial time cost value and the waiting cost value are added together to calculate the execution pre-cost of each atomic work unit in the current real workshop state. The execution pre-cost is treated as an edge attribute in the graph topology and overlaid on the preceding directed connection segment of the corresponding atomic task node in the directed acyclic graph, thus serving as the new weight of the logical connection edge. After modification, the directed acyclic graph model not only reflects the theoretical logical sequence of the product process but also deeply integrates the actual spatiotemporal losses and resistances caused by physical distance and equipment queuing in the current workshop. Based on this, a topology graph data structure containing specific edge weights and node triggering conditions is output, thereby decoupling the original order into a subtask flow with temporal dependencies.

[0026] The computational load characteristics and time constraint parameters of each subtask flow are extracted. The process of extracting the computational load characteristics and time constraint parameters of each subtask flow involves traversing the subtask flow, quantifying the peak floating-point operation volume and memory throughput requirements of each atomic task unit, and generating a load feature vector; using the critical path method to analyze the topology of the subtask flow, and combining the communication latency between agents at all levels in the production status map and the order delivery deadline, calculating the latest start time and time slack of each atomic task unit, and generating time constraint parameters.

[0027] Specifically, the cloud-based master control agent reads each atomic task unit in the subtask flow one by one, starting from the initial node with an in-degree of zero, according to the topological hierarchy of the directed acyclic graph. In this embodiment, a global calibration mapping table is stored in the database of the cloud-based master control agent. This mapping table establishes a mapping relationship between standard task types and extreme computing power consumption. For the currently read atomic task unit, the task type label is parsed, and the upper limit of the basic number of calculations per second that must be achieved to complete the visual positioning and high-precision grasping task under the most complex ambient lighting and highest frame rate image processing conditions is queried by calling the global calibration mapping table. That is, the peak floating-point operation volume required for the execution of this atomic task unit is determined. At the same time, the maximum limit value of the data cache read / write throughput per second when this task processes high-definition image streams and 3D point cloud data is queried and assigned as the memory throughput requirement. The extracted peak floating-point operation volume and the memory throughput requirement are numerically concatenated according to a fixed one-dimensional array structure with the computation requirement first and the storage requirement second, to construct the load feature vector specific to this atomic task unit and attach it to the corresponding node. The critical path method involves reading the directed acyclic graph of the generated subtask flows, performing a full path search from the starting node to the ending node, and extracting all possible serial and parallel execution sequences. It then calls the real-time updated production status map to extract the network transmission time (i.e., communication latency) between the agent nodes at each level assigned or to be assigned to adjacent atomic work units in each sequence. The baseline execution time of each atomic work unit is numerically accumulated with the communication latency between its predecessor and successor nodes. The cumulative time span of each complete sequence from the starting point to the end point is calculated, and the sequence with the largest cumulative time span is defined as the critical path. The mandatory delivery timestamp explicitly marked in the externally input drone production order file is obtained and established as the order delivery deadline. Using this order delivery deadline as the hard endpoint of the timeline, a reverse time calculation is performed from back to front along the topology of the directed acyclic graph. For any given atomic work unit, the latest start time of all its direct successor work nodes is found. The communication latency of the given atomic work unit transmitting data to the successor nodes is subtracted from the latest start time of the successor work nodes, and then the baseline execution time of the given atomic work unit itself is subtracted to obtain a reverse time point. If the atomic work unit has multiple direct successor work nodes, multiple reverse time points are calculated, and the earliest time value is set as the latest start time of the atomic work unit. Starting from the initial node of the directed acyclic graph and proceeding forward in time, under the ideal condition that all preceding resources do not need to wait, the earliest possible start time of the given atomic work unit is calculated. The time difference between the latest start time and the earliest possible start time is then determined and established as the time slack for the atomic work unit. In this embodiment, the time slack represents the maximum waiting redundancy time that the node can tolerate for agent resource scheduling or logistics handling without causing the entire drone production order to be overdue. The latest start time calculated for each atomic work unit is numerically bound and formatted with the time slack. On the one hand, by quantifying the computation and memory requirements of each work unit to construct a load feature vector, data support is provided for the matching and on-demand allocation of computing power for multi-level intelligent agents, effectively avoiding node computing power bottlenecks or resource waste. On the other hand, by comprehensively considering communication latency and order delivery constraints, the time slack and latest start time are dynamically calculated, clarifying the urgency of each sub-task and the scheduling fault tolerance boundary, ensuring from a temporal logic perspective that local delays will not lead to global order overdueness.

[0028] See Figure 2 A multi-level resource allocation mechanism is triggered based on computational load characteristics and time constraint parameters. The multi-level resource allocation mechanism divides the decoupled subtask flow into local operation tasks and global collaborative subtasks. The multi-level resource allocation mechanism extracts the time slack from the time constraint parameters of each atomic work unit and obtains the bidirectional link communication latency between the current scene node and the cloud master control agent based on the production status map; it then performs a three-dimensional joint decision: if the time slack is less than or equal to the bidirectional link communication latency, and the corresponding atomic work unit does not involve workstation interaction that crosses physical boundaries, and the peak floating-point operation volume and memory throughput requirement in its load feature vector do not exceed the sum of the real-time computing power margin of the terminal execution agent and the edge coordination agent in the current scene, it is confirmed as a local work task that meets the requirements for immediate response; The remaining atomic task units that do not meet the local task determination conditions are identified as global collaborative subtasks.

[0029] Specifically, each atomic task unit in the decoupled subtask flow is read sequentially; for each atomic task unit, the time constraint parameter data packet attached to it is accessed by the extraction program, and the time slack is extracted from a specific field of the data packet; in a specific scheduling scenario of this embodiment, the time slack of the fastening task atom in a certain UAV assembly scenario is extracted to be 50ms. Access the synchronized and updated production status map, and based on the physical entity node location currently assigned to the atomic work unit, retrieve the logical link from the current scene node to the cloud master control agent node in the production status map, and read the transmission weight attribute mounted on the logical link to obtain the bidirectional link communication latency between the current scene node and the cloud master control agent; in this embodiment, by reading the network transmission test data recorded in the map at the current moment, the specific value of the bidirectional link communication latency is obtained as 80ms; After obtaining the above parameters, the first dimension of the three-dimensional joint decision is executed, that is, the extracted time relaxation amount is compared with the obtained bidirectional link communication delay. Since the extracted 50ms time relaxation amount in this embodiment is less than the 80ms bidirectional link communication delay, it is determined that if the task is reported to the cloud to generate a strategy and then sent down, its communication time will exceed the waiting time allowed by the task itself. Therefore, it is confirmed that the atomic operation unit meets the strict delay constraint of the first dimension. The second dimension of the three-dimensional joint decision is to check whether the corresponding atomic work unit involves workstation interactions that cross physical boundaries; the process path attributes of the atomic work unit are analyzed, and the spatial coordinate information of the material source node and assembly execution node required for the operation is extracted; in this embodiment, it is read that the material source and assembly execution are both located in the same physical area of ​​the same final assembly workshop, and the scene identification codes of the two nodes are compared to confirm that the identification codes are completely consistent, and it is determined that the corresponding atomic work unit does not involve workstation interactions that cross physical boundaries, thus satisfying the second dimension of spatial local closed loop conditions; The third dimension of the three-dimensional joint decision-making process involves real-time matching calculations of computing power supply and demand. From the load feature vector of the atomic work unit, the peak floating-point operation volume and memory throughput requirements for executing the task are read. Simultaneously, the production status map is retrieved again to identify all online terminal execution agents and edge coordination agents in the current scenario, and the real-time computing power reserve information of these nodes is extracted and summed. In this embodiment, the sum of the real-time computing power reserve and available memory throughput in the current scenario is obtained. Through difference comparison, it is confirmed that the peak floating-point operation volume and memory throughput requirements for executing the task do not exceed the total real-time computing power reserve of the terminal execution agents and edge coordination agents in the current scenario, thus satisfying the third dimension's computing power resource carrying capacity conditions. When the system detects that the decision conditions of time constraints, spatial boundaries, and computing resources are all met simultaneously, the attribute label of the atomic job unit is rewritten, it is confirmed as a local job task that meets the requirements of immediate response, and it is prepared to be assigned to the local joint processing domain for in-situ scheduling. The three-dimensional joint decision is executed sequentially according to the time constraint dimension, spatial boundary dimension, and computing power resource dimension. When the time constraint dimension is not satisfied (i.e., the time slack is greater than the bidirectional link communication latency), the atomic job unit is directly classified as a global collaborative subtask, without further dimension determination. When the time constraint dimension is satisfied but the spatial boundary dimension is not satisfied, it is also classified as a global collaborative subtask. When both the time constraint dimension and the spatial boundary dimension are satisfied but the computing power resource dimension is not satisfied, a computing power expansion waiting strategy is executed. That is, the terminal execution agent and the edge coordination agent in the current scenario are polled to find nodes that meet the computing power requirements, and the three-dimensional joint decision is retried after a node that meets the conditions is detected. If the waiting time exceeds a preset threshold (set to 30 seconds in this embodiment), the atomic job unit is forcibly classified as a global collaborative subtask.

[0030] The remaining atomic task units that do not meet the local task determination conditions are identified as global collaborative subtasks. For example, if the time slack of a certain task unit is 100ms (a communication delay greater than 80ms), or if it needs to call materials from the warehousing scenario to the final assembly scenario (crossing physical boundaries), it is determined that it does not meet the immediate response requirements and is classified as a global collaborative subtask.

[0031] By introducing a three-dimensional joint decision-making mechanism based on time constraints, spatial boundaries, and computing resources, precise allocation of computing resources for multi-level intelligent agents and efficient cloud-edge-device collaboration are achieved. First, tasks with time slack less than communication latency and not spanning multiple regions are executed locally, eliminating the overhead of round-trip data transmission to the cloud and ensuring the extreme real-time requirements of high-frequency or high-precision assembly stations for immediate response. Second, computing power supply and demand matching is performed simultaneously during the decision-making process, effectively avoiding the risk of blindly assigning high-load tasks, which could lead to overload and paralysis of edge or terminal nodes.

[0032] See Figure 3 The local task is assigned to a joint processing domain composed of terminal execution agents and edge coordination agents for in-situ scheduling; The in-situ scheduling involves the terminal execution agent calling its own hardware computing power to parse the local task and comparing the real-time monitored instantaneous available computing power with the peak floating-point operation volume of the local task. When the instantaneous available computing power is less than the peak floating-point operation volume, the computational load of the local task is divided into locally retained load and unloaded load based on the difference in computing power. The terminal execution agent processes the locally retained load and transmits the unloaded load to the edge coordination agent in the same scene through the local area network. The terminal execution agent and the edge coordination agent respectively parse the carried load and issue control commands to the corresponding physical entity nodes.

[0033] Specifically, after a job instruction is confirmed as a local job task, the terminal executing agent that receives the task activates its control program, calls upon its configured hardware computing power, and parses the data packet content of the local job task; the terminal executing agent starts its internal resource status monitoring component to monitor and extract its instantaneous available computing power in real time; simultaneously, it extracts the pre-quantized peak floating-point operation volume from the attribute description file of the local job task; and compares the real-time monitored instantaneous available computing power with the peak floating-point operation volume of the local job task; in this embodiment, the preset real-time monitored instantaneous available computing power is 5 billion floating-point operations per second, and the extracted peak floating-point operation volume is 8 billion floating-point operations per second; Based on the numerical comparison results, when it is determined that the instantaneous available computing power of the terminal executing the intelligent agent is less than the peak floating-point operation volume required for the task, a load splitting mechanism is immediately triggered. The computing power difference between the two is calculated, i.e., there is a computing power gap of 3 billion floating-point operations per second. The load splitting mechanism divides the overall computing load of the local task into two parts according to the proportional distribution of the computing power difference: a locally retained load adapted to its current available processing capacity, and a load to be unloaded corresponding to the excess computing power difference. In this embodiment, the load module with a computing power matching 5 billion floating-point operations per second is split into the locally retained load, and the load module with a computing power matching 3 billion floating-point operations per second is split into the load to be unloaded. The terminal execution agent's underlying computing unit takes over and directly processes the segmented locally retained payload, independently completing the corresponding local data calculations. At the same time, the terminal execution agent calls the network communication interface to encapsulate the segmented payload to be unloaded, and transmits the encapsulated payload to be unloaded to the edge coordination agent with sufficient computing power in the same scene through the local area network in the same scene.

[0034] The terminal execution agent and the edge coordination agent that received the data perform parallel parsing of the payload they carry in their respective computing environments. After the edge coordination agent completes the parsing of the payload to be unloaded, it generates execution parameters for the coordination end. Simultaneously, after the terminal execution agent completes the parsing of the locally retained payload, it generates execution parameters for the terminal. Based on their respective parsed execution parameters, the edge coordination agent and the terminal execution agent issue specific control commands to the physical entity nodes corresponding to the local task through the underlying control link, thereby completing the in-situ collaborative execution of the local task. When the available computing power of the terminal is insufficient, this application can quantify the computing power gap and realize the adaptive splitting and parallel processing of computing tasks between the terminal and edge nodes in the same scene. It digests the complex computing load within the local area network, which not only avoids the transmission latency and bandwidth consumption caused by massive amounts of underlying data crossing the Internet to the cloud, but also ensures extremely low latency and high synchronization of field-level physical entity control, and realizes efficient collaboration between edge computing power pool and terminal node computing power.

[0035] The global collaborative sub-tasks are assigned to the cloud-based master control agent to generate a collaborative scheduling strategy. The cloud-based master control agent decomposes the collaborative scheduling strategy into cross-scenario execution instructions and sends them to the edge coordination agent and terminal execution agent in the corresponding scenario for collaborative operation.

[0036] The collaborative scheduling strategy involves the cloud-based master control agent mapping the global collaborative sub-tasks to the production status map, extracting the physical connection relationships and logical link transmission weights across warehousing and final assembly scenarios; sorting the global collaborative sub-tasks by spatiotemporal nodes based on the extracted physical connection relationships and logical link transmission weights, generating a collaborative scheduling strategy that includes cross-domain transfer paths and trigger timestamps for each node; segmenting the collaborative scheduling strategy along the scene physical boundaries of the cross-domain transfer paths to generate local execution sequences belonging to the corresponding scenarios; extracting timestamps at the physical handover nodes across scenarios to generate timing synchronization tokens, and encapsulating and combining the local execution sequences and the timing synchronization tokens into the cross-scenario execution instructions, which are then sent to the edge coordination agent of the corresponding scenario.

[0037] Specifically, when the cloud-based master control agent receives a global collaborative sub-task, it maps the global collaborative sub-task to the already constructed production status map; based on the starting and ending nodes of the task, it extracts the physical connection relationships and logical link transmission weights that span the warehousing and final assembly scenarios in the map topology network.

[0038] Based on the extracted physical connection relationships and logical link transmission weights, the cloud-based master control agent sorts the global collaborative sub-tasks by spatiotemporal nodes. It then comprehensively calculates the obtained passage cost and network latency values, decomposing the cross-domain operation into a continuous sequence of action nodes according to temporal and spatial order, generating a collaborative scheduling strategy that includes cross-domain transfer paths and trigger timestamps for each node. In this embodiment, the pre-generated cross-domain transfer path sequentially includes a starting scene extraction node, a cross-scene physical handover node, and a target scene receiving node; simultaneously, trigger timestamps are calculated and bound to each node. The cloud-based master control agent executes a path decoupling program, which segments the collaborative scheduling strategy along the physical boundary of the cross-domain transfer path. It retrieves the regional boundary attributes from the production status map, using the boundary line between the identified warehousing and final assembly scenarios as the physical boundary of the scenario. Using this physical boundary as a breakpoint, the complete and continuous collaborative scheduling strategy is severed at the breakpoint, dividing it into two independent control flows, generating local execution sequences belonging to the corresponding scenarios. In this embodiment, after segmentation, a first local execution sequence belonging to the warehousing scenario and a second local execution sequence belonging to the final assembly scenario are generated. The two sequences respectively contain spatiotemporal control logic before and after the breakpoint. To ensure absolute time coordination between agents in different scenarios at physical boundary intersections, timestamps at cross-scenario physical intersection nodes are extracted to generate timing synchronization tokens. In this embodiment, the timestamps of cross-scenario physical intersection nodes are extracted and converted into timing synchronization tokens with cross-network segment mandatory verification attributes. The cloud-based master control agent encapsulates and combines the generated local execution sequence with the timing synchronization tokens to generate the cross-scenario execution instructions. In this embodiment, the first local execution sequence and the timing synchronization token are encapsulated and combined into a first cross-scenario execution instruction sent to the warehousing scenario, and the second local execution sequence and the same timing synchronization token are encapsulated and combined into a second cross-scenario execution instruction sent to the final assembly scenario. The cloud-based master control agent sends the encapsulated cross-scenario execution instructions to the corresponding edge coordination agents through the downlink communication network. To address the timing misalignment problem that is prone to occur in cross-scenario collaborative operations, a collaborative scheduling strategy generation method based on physical boundary breakpoint segmentation and cross-scenario timing synchronization tokens is proposed. The cloud-based master control agent no longer issues complex global instructions, but instead decouples the strategy into independent local sequences along physical boundaries. By extracting the timestamps of the handover nodes to generate mandatory verification synchronization tokens, it ensures that edge agents in the two independent local area networks of warehousing and final assembly can achieve precise handshakes and time alignment at physical handover points when facing network fluctuations or local task delays, relying on a unified synchronization token. This eliminates the risks of blind spot collisions and waiting deadlocks in cross-domain logistics and assembly processes.

[0039] Example 2 This implementation applies a ubiquitous intelligent AI agent task scheduling method for multi-scenario collaborative applications to the cross-regional delivery and joint calibration tasks of UAV visual navigation modules. The process includes: By deploying three levels of ubiquitous AI agents in warehousing, final assembly, and related scenarios, real-time environmental data from multiple scenarios is collected and a production status map is constructed. In this embodiment, the warehousing terminal agent collects real-time data showing that the visual navigation module is located in the electronic silo at coordinates (X: 15, Y: 30, Z: 2), and the current yaw angle of the transport AGV is 15° and it is in an idle state. The final assembly terminal agent collects data showing that assembly station No. 3 is occupied, and the estimated remaining operation time is 120 seconds. At the same time, the local area network communication bandwidth from the terminal to the edge is measured to be 800Mbps, and the bidirectional link communication latency from the edge to the cloud master control agent is 50ms. The real-time computing power margin of the assembly robot (terminal) at station No. 3 is extracted to be 80GFLOPS, and the real-time computing power margin of its corresponding edge server (edge) is 200GFLOPS. After synchronizing the above environmental data and computing power margin with time stamps and normalizing the coordinate system, the data is injected into physical entity nodes to construct the production status map at the current moment. The cloud-based main control agent receives the production order for the UAV's visual navigation module and decouples the order into a sub-task flow with temporal dependencies based on the production status map. It analyzes the process path of the order and maps it to four atomic operation units: "A. Material grabbing," "B. Cross-regional logistics handling," "C. Visual alignment and precision bonding," and "D. Power-on joint calibration." A directed acyclic graph (DAG) model is used to establish the temporal connection relationship between these four nodes A->B->C->D, and the physical distance in the production status map is extracted. In this embodiment, the physical path distance from the silo to assembly station 3 is 30 meters. Combined with the AGV's preset travel speed of 2 m / s, the 15-second space consumption cost is added to the 120-second station waiting cost, calculating the pre-execution cost of node B as 135 seconds. This dynamically adjusts the logical connection edge weights in the DAG to generate the sub-task flow. The subtask flow is traversed, and the computational load characteristics and time constraint parameters of each subtask flow are extracted. In this embodiment, for the "visual alignment and precise bonding" operation of node C, the peak floating-point operation required for performing binocular visual point cloud calculation is quantified to be 120 GFLOPS, and the memory throughput requirement is 4 GB / s, generating the corresponding load feature vector. At the same time, the directed acyclic graph is analyzed using the critical path method, and combined with the final delivery deadline of the order (preset to 15:00:00 on the same day), the latest start time of node C is deduced in reverse, and the time slack allowed for this node without causing the order to be overdue is calculated to be 30 ms, generating the time constraint parameters. After obtaining the above parameters, the multi-level resource allocation mechanism performs a three-dimensional joint decision based on the extracted load characteristics and time constraint parameters. In the decision process of this embodiment, for node C, it is found that its time relaxation (30ms) is less than the cloud bidirectional link communication latency (50ms), and the bonding operation is only completed within the No. 3 final assembly station and does not involve crossing physical boundaries. At the same time, the peak floating-point operation required by node C (120GFLOPS) is less than the sum of the real-time computing power margin of the terminal and edge in the current scenario (80+200=280GFLOPS). All three conditions are met, and node C is identified as a local operation task that meets the requirements of immediate response. As for node B "cross-regional logistics handling", analysis shows that it involves physical space transfer from the warehousing scenario to the final assembly scenario, which does not meet the spatial local closed loop condition. Therefore, it is identified as a global collaborative sub-task. For node C, which is identified as a localized task, the terminal execution agent (assembly robot at workstation 3) triggers the in-situ scheduling mechanism and performs computational load splitting. The real-time available computing power of the terminal (80 GFLOPS) is compared with the peak floating-point computation required for the task (120 GFLOPS), determining a 40 GFLOPS computing power gap. The terminal execution agent immediately splits the computational load of node C into two parts: the basic robotic arm servo control and 2D image denoising, matching 80 GFLOPS, are designated as locally retained loads and processed directly using its own hardware; simultaneously, the high-precision 3D point cloud spatial pose calculation, matching 40 GFLOPS, is designated as a load to be unloaded and transmitted with low latency via the local area network to the edge coordination agent within the same scene for parallel computation. After the computation is completed, the terminal and edge agents respectively issue control commands to the robotic arm joints and the end-effector to complete the in-situ collaborative operation. For node B, which is identified as a global collaborative sub-task, the cloud-based master control agent generates a cross-scenario collaborative scheduling strategy. Node B is mapped to the production status map, and spatiotemporal node sorting is performed by considering both passage cost and network latency to generate a cross-domain transfer path from the "warehouse exit" to the "final assembly entrance". The path is then segmented at the physical boundary between the warehouse and final assembly scenarios (i.e., the No. 2 cross-workshop isolation door), generating an AGV driving sequence for the warehouse and a workstation receiving preparation sequence for the final assembly. Simultaneously, the estimated physical handover timestamp of the AGV arriving at the No. 2 cross-workshop isolation door (preset to 14:30:00) is extracted to generate a timing synchronization token. Finally, the cloud encapsulates and sends the warehouse driving sequence and the token to the warehouse edge, and encapsulates and sends the final assembly receiving sequence and the token to the final assembly edge, ensuring that the two ends achieve seamless cross-scenario logistics handover at 14:30:00 using the token.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A ubiquitous intelligent AI agent task scheduling method for multi-scenario collaborative applications, characterized in that, include: By deploying ubiquitous AI agents at the cloud, edge, and terminal levels in warehousing, assembly, and related scenarios, real-time environmental data from multiple scenarios is collected. Simultaneously, the computing power reserve information of the computing resource node pool composed of hardware carriers at each level of the AI ​​agents is extracted to construct a production status map. The production status map construction process involves uniformly time-stamping the multi-scenario environmental data and performing a normalization transformation of the spatial coordinate system to generate a structured environmental dataset. Physical entity nodes are defined based on material storage locations and workstation occupancy status, and the corresponding material types and equipment numbers are attached to these physical entity nodes as static attribute parameters. Based on the AI ​​agent nodes at each level… The system calculates the transmission weight of the logical link based on the communication bandwidth and latency parameters between nodes, and establishes physical connections with passage cost weights between physical entity nodes based on the occupancy of logistics channels. It injects real-time collected data on the pose of handling equipment, the operation parameters of assembly robots, and the spatial pose and geometric accuracy data of UAV components as dynamic features into the physical entity nodes, and uses the computing power margin information of each level of intelligent agent hardware carrier as the node's own computing power carrying parameters. It receives UAV production orders, decouples the orders into time-dependent sub-task flows based on the production status map, and extracts the computational load characteristics and time constraint parameters of each sub-task flow. A multi-level resource allocation mechanism is triggered based on computational load characteristics and time constraint parameters. This mechanism divides the decoupled subtask flow into local tasks and global collaborative subtasks. Local tasks are assigned to a joint processing domain composed of terminal execution agents and edge coordination agents for in-situ scheduling. The multi-level resource allocation mechanism extracts the time slack from the time constraint parameters of each atomic task unit and obtains the bidirectional link communication latency between the cloud master control agent and the production status map. A three-dimensional joint decision is made: if the time slack is less than or equal to the bidirectional link communication latency, and the corresponding atomic task unit does not involve workstation interaction across physical boundaries, and the peak floating-point operation volume and memory throughput requirements in the load feature vector do not exceed the sum of the real-time computing power margins of the terminal execution agent and the edge coordination agent in the current scenario, it is confirmed as a local task that meets the immediate response requirements. The remaining atomic task units that do not meet the local task determination conditions are confirmed as global collaborative subtasks. The global collaborative sub-tasks are assigned to the cloud-based master control agent to generate a collaborative scheduling strategy. The cloud-based master control agent decomposes the collaborative scheduling strategy into cross-scenario execution instructions and sends them to the edge coordination agent and terminal execution agent in the corresponding scenario to perform collaborative work. To ensure absolute time coordination between agents in different scenarios at the physical boundary, the timestamps at the cross-scenario physical handover nodes are extracted to generate time synchronization tokens.

2. The ubiquitous intelligent AI agent task scheduling method for multi-scenario collaborative applications according to claim 1, characterized in that, The multi-scenario environmental data includes warehousing scenario data, final assembly scenario data, and cross-scenario collaborative data; the warehousing scenario data includes material inventory balance, material storage location, and real-time pose of handling equipment; the final assembly scenario data includes the occupancy status of each workstation, the operating parameters of the assembly robot, and the real-time spatial pose of the drone components and the geometric accuracy of the connection points. The cross-scenario collaborative data includes the occupancy status of logistics channels between scenarios and the communication bandwidth and latency parameters between intelligent agent nodes at all levels, and a production status map is constructed by combining multi-scenario environmental data. The ubiquitous intelligent AI agent includes a cloud-based master control agent, an edge coordination agent, and a terminal execution agent.

3. The ubiquitous intelligent AI agent task scheduling method for multi-scenario collaborative applications according to claim 1, characterized in that, The specific process of decoupling orders into subtask flows with temporal dependencies involves parsing the product structure tree and process path in the UAV production order and mapping them to a preset set of atomic work units; establishing temporal connections between work units based on the physical assembly logic and process sequence between each atomic work unit, defining the in-degree constraints and out-degree triggering conditions for each node; retrieving the corresponding material storage location, handling equipment pose, and workstation occupancy status from the production status map, converting physical space distance and equipment availability into the pre-execution cost of each work unit, dynamically adjusting the weights of logical connection edges in the directed acyclic graph, and generating subtask flows with temporal dependencies.

4. The ubiquitous intelligent AI agent task scheduling method for multi-scenario collaborative applications according to claim 3, characterized in that, The process of extracting the computational load characteristics and time constraint parameters of each subtask flow involves traversing the subtask flow, quantifying the peak floating-point operation volume and memory throughput requirements of each atomic task unit based on each atomic task unit, and generating a load feature vector. The topology of the subtask flow is analyzed using the critical path method. Combined with the communication latency between agents at all levels in the production status map and the order delivery deadline, the latest start time and time slack of each atomic work unit are calculated to generate time constraint parameters.

5. The ubiquitous intelligent AI agent task scheduling method for multi-scenario collaborative applications according to claim 1, characterized in that, The in-situ scheduling involves the terminal executing the intelligent agent calling its own hardware computing power to parse the local task and comparing the real-time monitored instantaneous available computing power with the peak floating-point operation volume of the local task; when the instantaneous available computing power is less than the peak floating-point operation volume, the computing load of the local task is divided into locally retained load and unloaded load according to the difference in computing power between the two. The terminal execution agent processes the locally retained payload and transmits the payload to be unloaded to the edge coordination agent in the same scene via the local area network; the terminal execution agent and the edge coordination agent respectively parse the carried payload and issue control commands to the corresponding physical entity nodes.

6. The ubiquitous intelligent AI agent task scheduling method for multi-scenario collaborative applications according to claim 1, characterized in that, The collaborative scheduling strategy involves the cloud-based master control agent mapping the global collaborative sub-tasks to the production status map, extracting the physical connection relationships and logical link transmission weights across warehousing and final assembly scenarios; sorting the global collaborative sub-tasks by spatiotemporal nodes based on the extracted physical connection relationships and logical link transmission weights, generating a collaborative scheduling strategy that includes cross-domain transfer paths and trigger timestamps for each node; segmenting the collaborative scheduling strategy along the scene physical boundaries of the cross-domain transfer paths to generate local execution sequences belonging to the corresponding scenarios; extracting timestamps at the physical handover nodes across scenarios to generate timing synchronization tokens, and encapsulating and combining the local execution sequences and the timing synchronization tokens into the cross-scenario execution instructions, which are then sent to the edge coordination agent of the corresponding scenario.

Images