Method, device, equipment and system for dynamic permission handover of agent with human-machine symbiotic architecture

Abstract

The application provides an agent dynamic permission transfer method with a man-machine symbiotic architecture and related equipment. The method acquires a task real-time running feature stream and extracts an environment interaction node sequence to perform intention deviation detection. When a deviation metric reaches a baseline, the agent permission is suspended, and an on-site stack snapshot carrying source information is recorded. A man-machine interaction probe is generated based on this and sent to a hardware console to present a panoramic restoration view. An intervention intervention map is received and translated into a state correction chain. The stack snapshot is reconstructed based on the chain, and the running permission is released to continue the task. The application effectively improves the continuity of agent control and the smoothness of scene intervention.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and intelligent agent collaborative control technology, and more specifically, to an intelligent agent dynamic permission handover method, device, electronic device, computer-readable storage medium, intelligent agent and system with a human-machine symbiotic architecture. Background Technology

[0002] With the rapid development of smart gardens and intelligent management of public spaces, management agents are increasingly being used in scenarios such as security monitoring, anomaly identification, and automated handling. Management agents need to continuously monitor the environment through multi-dimensional sensing devices and independently execute various management tasks such as anomaly order processing and security patrols in complex dynamic scenarios. This places higher demands on the agent's adaptability in the event of emergencies and its human-machine intervention and collaboration capabilities.

[0003] Existing intelligent agent anomaly management solutions typically employ a rule-triggered, independent command flow-based blocking architecture. This approach first utilizes visual algorithms and edge computing nodes to monitor abnormal behavior within the area, immediately terminating the agent's current automated task upon detecting a sudden situation. Subsequently, the system extracts isolated video clips of the moment the anomaly occurs, directly generating an anomaly log or dispatching an intervention request to the control terminal. Finally, console personnel reassess the situation based on the static anomaly log and issue a new reset command to the agent to initiate the next phase of the task.

[0004] However, this blocking intervention approach has significant technical flaws. By directly terminating the task and stripping away the runtime context, the system cannot retain the underlying execution chain and interaction semantics when the agent deviates, resulting in the complete loss of the perception state and historical logic in dynamic scenarios. At the same time, relying on isolated image fragments to reset instructions lacks immersive reconstruction of the multi-dimensional stack state of the scene. The control end cannot intuitively inject precise intervention logic, and it cannot smoothly convert external intervention actions into repair instructions of the original task sequence. This results in a severe sense of disconnect in the handover of agent authority, making it impossible to achieve seamless dynamic intervention and task continuation based on the continuation of the underlying stack. Summary of the Invention

[0005] This application provides a method, apparatus, electronic device, computer-readable storage medium, intelligent agent, and system for dynamic authorization transfer of intelligent agents with a human-machine symbiotic architecture, so as to at least alleviate the above-mentioned technical problems.

[0006] The technical advantages of the technical solution provided in this application are: This application presents a dynamic permission handover method for intelligent agents with a human-machine symbiotic architecture. Addressing the technical shortcomings of existing blocking intervention schemes that directly terminate tasks, leading to loss of perception state and fragmented permission handover, this method extracts the sequence of environmental interaction nodes from the real-time runtime feature stream for intent deviation detection. When the deviation scalar reaches a preset trigger baseline, a permission suspension command is generated and the system switches to a permission suspension state. Simultaneously, a snapshot of the on-site stack carrying task state source information is recorded. This solves the problem of lost underlying execution links caused by directly severing the runtime context in traditional schemes. Compared to the traditional, abrupt task termination logic, this application flexibly suspends permissions while preserving the complete execution state of the intelligent agent. This significantly improves the preservation of the intelligent agent's runtime state and the continuity of the logical context, providing a highly coherent data foundation for subsequent precise intervention.

[0007] Based on the recorded on-site stack snapshots, this application generates a human-computer interaction probe and sends it to the hardware console to present a panoramic view, solving the problem that traditional solutions rely on isolated screen fragments, making it impossible for the control end to intuitively perceive the underlying state. Compared to the shallow interaction mode of traditional methods that only push single slices and static logs, this application uses probes to construct a panoramic view on the console that reflects deep execution logic, greatly improving the control end's understanding of complex scene environments and the intuitiveness of formulating intervention actions.

[0008] Finally, this application receives the intervention map issued by the hardware console, translates its execution instructions to obtain a state correction chain, and then releases the agent's running authority to continue advancing the target task after reconstructing the on-site stack snapshot based on this state correction chain. This solves the problem that traditional reset instructions cannot smoothly continue with the original task logic. Compared with the traditional method of forcibly issuing new instructions to start a new task, this application directly translates the external intervention intention and organically integrates it into the reconstruction process of the underlying stack snapshot, realizing in-situ closed-loop repair of the logical link. This greatly enhances the smoothness of human-machine authority handover and the coherence of the final execution of the target task. Attached Figure Description

[0009] Figure 1 This application provides an embodiment of a dynamic permission handover scenario for an intelligent agent with a human-machine symbiotic architecture. Figure 2 This application provides an embodiment of a method for dynamic permission handover of an intelligent agent with a human-machine symbiotic architecture. Figure 3 This application provides an embodiment of a dynamic permission handover device for an intelligent agent with a human-machine symbiotic architecture. Figure 4 This is an electronic device according to an embodiment of the present application.

[0010] Figure 5This is a computer-readable storage medium according to an embodiment of the present application.

[0011] Figure 6 This is an example of an intelligent agent in this application.

[0012] Figure 7 This application provides an embodiment of a dynamic permission handover system for intelligent agents with a human-machine symbiotic architecture. Detailed Implementation

[0013] like Figure 1 As shown, this is a scenario of dynamic permission handover for an intelligent agent with a human-machine symbiotic architecture according to an embodiment of this application; Figure 2 As shown, this application provides an embodiment of a dynamic permission handover method for an intelligent agent with a human-machine symbiotic architecture. The method includes the following steps: Obtain the real-time running feature stream of the target task, and extract the environmental interaction node sequence from the real-time running feature stream; obtain the preset trigger baseline corresponding to the target task, perform intent deviation detection on the environmental interaction node sequence, and generate a deviation scalar; When the deviation scalar reaches the preset trigger baseline, a permission suspension instruction is generated, and based on the permission suspension instruction, the running permission of the agent in the current execution cycle is switched to a permission suspension state. In the permission-suspended state, a snapshot of the agent's context stack is recorded in the current execution cycle, and the context stack snapshot carries the task state source information corresponding to the current execution cycle; A human-computer interaction probe is generated based on the on-site stack snapshot, and the human-computer interaction probe is sent to the hardware console so that the hardware console can present a panoramic restoration view based on the human-computer interaction probe; Receive the intervention atlas issued by the hardware console for the panoramic restoration view; perform instruction translation operation on the intervention atlas to obtain the state correction chain; The state correction chain is used to reconstruct the field stack snapshot, forming a reconstructed field stack snapshot; the execution permission of the agent is released based on the reconstructed field stack snapshot, and the agent is guided to continue to advance the target task according to the reconstructed field stack snapshot.

[0014] Optionally, extracting the sequence of environmental interaction nodes from the real-time running feature stream includes: The real-time running feature stream is divided into at least two temporal feature slices; A spatiotemporal correlation comparison operation is performed on each of the time-series feature slices to identify mutation feature parameters in each of the time-series feature slices through the spatiotemporal correlation comparison operation; The temporal feature slice containing the mutation feature parameter is marked as an environmental interaction candidate node; Extract at least two consecutively occurring environmental interaction candidate nodes, and concatenate the extracted at least two environmental interaction candidate nodes in chronological order to form the environmental interaction node sequence.

[0015] Preferably, the specific implementation process of dividing the real-time running feature stream into at least two temporal feature slices is as follows: The real-time running feature stream is not a single video frame sequence, but a data stream formed by multiple running feature units continuously formed during the execution of the target task, arranged according to the sampling time label. Each running feature unit includes a source field, a scene location field, a sampling time label, a task stage field, a feature content field, and a device feedback field; wherein, the source field is used to distinguish whether the running feature unit originates from camera frame extraction, environmental perception device reporting, smart security identification results, smart broadcast feedback status, smart lighting feedback status, smart irrigation feedback status, or work order flow status; the scene location field is used to express the park area, camera drawing boundary data, and spatial location in the geographic information management platform corresponding to the running feature unit, the camera drawing boundary data comes from the boundary data formed after regional drawing of the camera monitoring screen, and the spatial location in the geographic information management platform is used to map the park area and the camera drawing boundary data to the same spatial reference; the task stage field is used to express which type of execution stage the target task is currently in, such as inspection, identification, alarm, linkage, or handling feedback. With the above field configuration, the time-series feature slices obtained by subsequent segmentation can simultaneously retain the time source, spatial source, device source, and task source, instead of only retaining isolated image fragments, thereby enabling the time-series feature slices to serve as the data basis for subsequent identification of mutation feature parameters.

[0016] Preferably, when the real-time running feature stream is segmented, the sampling time tag, the task stage field, the acquisition source field, and the device feedback field in each running feature unit are read first, and the running sequence is established according to the sampling time tag; then, according to the camera frame-dropping trigger position, the environmental perception device reporting position, the target task stage switching position, and the device feedback state change position, a slice boundary marker is written into the running sequence; then, the running feature units between two adjacent slice boundary markers are used as the carrying content to form a time-series feature slice. The slice boundary marker is not simply mechanically divided according to a fixed time interval, but is jointly determined by the frame-dropping frequency, behavior recognition group, area monitoring boundary, environmental perception device reporting rhythm, and target task stage change; for example, when the camera corresponding to the security sentry is configured as a behavior recognition group for water danger, ground garbage, no fishing, or no climbing, the slice boundary marker will preferentially fall between the frame-dropping trigger position and the recognition result feedback position of the behavior recognition group, so that the formed time-series feature slice can carry the environmental interaction state changes of the behavior recognition group within a continuous observation interval.

[0017] Preferably, the technical reason for dividing the real-time running feature stream into at least two temporal feature slices is that the mutation feature parameter needs to be formed by comparing the data states of adjacent data. A single temporal feature slice can only express the local environmental interaction state within a time interval, and cannot express whether the local environmental interaction state has changed relative to the previous time interval. In specific implementation, the currently formed temporal feature slice is configured as the current comparison slice, and the temporal feature slice formed immediately before the current comparison slice is configured as the preceding reference slice. The current comparison slice and the preceding reference slice jointly participate in the spatiotemporal correlation comparison operation to determine whether the image content within the boundary data drawn by the same camera, the environmental perception data within the same park area, and the device feedback state within the same target task have changed in relation to the execution stage corresponding to the task stage field. Through the connection relationship between at least two temporal feature slices, the formation process of the environmental interaction node sequence can reflect the evolution of the environmental interaction state, rather than directly identifying a certain frame image, a certain sensor report, or a certain alarm record as an environmental interaction node.

[0018] Preferably, the specific implementation process of performing spatiotemporal correlation comparison operation on each of the time-series feature slices is as follows: First, the scene location field in the time-series feature slice is read, and the scene location field is mapped to the spatial location in the geographic information management platform to determine the park area, monitoring device location, and camera drawing boundary data corresponding to the time-series feature slice; then, the sampling time tag in the time-series feature slice is read, and the sampling time tag is matched with the current execution cycle of the target task to determine the time position of the time-series feature slice in the target task; next, the spatial location, the camera drawing boundary data, and the time position are used together as a correlation index, and the running feature units in the current comparison slice and the preceding reference slice that have the same source or have a spatial adjacency relationship are matched according to the correlation index. The above-mentioned corresponding processing ensures that the identification results of behaviors such as water hazards, ground litter, fishing ban, climbing ban, cycling on sidewalks, road litter, and illegal occupation of roads are within the same comparable range as their corresponding camera boundary data, camera frame-taking trigger locations, environmental perception data, and device feedback status. This avoids incorrectly comparing the operational feature units corresponding to fields from different park areas, different time locations, or different task stages.

[0019] Preferably, after completing the spatial and temporal location correspondence, the spatiotemporal correlation comparison operation performs format unification processing on the feature content fields in the current comparison slice to form a current state description item; simultaneously, the feature content fields in the preceding reference slice are subjected to the same format unification processing to form a preceding state description item. The current state description item includes the behavior recognition category obtained by camera frame extraction, the camera drawing boundary data corresponding to the behavior recognition category, and the temperature status, humidity status, light intensity status, wind speed status, smart broadcast feedback status, smart lighting feedback status, smart irrigation feedback status, and work order flow status reported by the environmental sensing device; the preceding state description item includes data content with the same field scope as the current state description item. Subsequently, the current state description item is compared with the previous state description item by item to identify data changes under the same spatial location, the same camera drawing boundary data and the same task stage field. This comparison does not directly mix data with different dimensions, but first completes the category change judgment, state change judgment, numerical range change judgment and task stage change judgment according to the meaning of the field, and then encapsulates the above judgment results into the mutation feature parameter.

[0020] Preferably, the technical essence of the mutation feature parameter is a structured state parameter used to describe the changes in environmental interaction of the current comparison slice relative to the preceding reference slice, rather than a single anomaly score. The mutation feature parameter includes a mutation source field, a mutation location field, a mutation time field, a mutation category field, a mutation magnitude field, and a task association field. Specifically, the mutation source field marks the mutation feature parameter as originating from camera frame extraction, environmental perception device reporting, smart security identification results, smart broadcast feedback status, smart lighting feedback status, smart irrigation feedback status, or work order flow status; the mutation location field indicates the spatial location in the camera's boundary data and the geographic information management platform; the mutation category field distinguishes whether the mutation feature parameter belongs to a change in behavior recognition category, a change in environmental perception data status, a change in device feedback status, or a change in task stage status; and the task association field expresses the correspondence between the mutation feature parameter and the current execution stage of the target task expressed by the task stage field. Through these fields, the mutation feature parameter can retain its source, location, time, category, and task relationship when subsequently marking candidate nodes for environmental interaction, preventing it from becoming an isolated alarm detached from the target task.

[0021] Preferably, when identifying the mutation feature parameters, if the current state description item shows a change from a state of no abnormal behavior to one of the following behavior recognition states within the same camera drawing boundary data: water danger, no fishing, no climbing, sidewalk cycling, road litter, or illegal occupation of the road, and the preceding state description item is recorded as the state of no abnormal behavior within the same camera drawing boundary data, then the change in the behavior recognition state is written into the mutation category field; if the current state description item shows a range change in the temperature, humidity, light intensity, or wind speed reported by the environmental sensing device relative to the preceding state description item, and this range change corresponds to the current execution stage of the target task expressed by the task stage field, then this range change is written into the mutation amplitude field; if the current state description item shows a start / stop state change in the smart broadcast feedback state, smart lighting feedback state, or smart irrigation feedback state relative to the preceding state description item, then this start / stop state change is written into the mutation source field corresponding to the device feedback field. The above identification method correlates camera frame extraction, environmental perception data, and device feedback status under the same spatiotemporal caliber, enabling the mutation feature parameter to express the technical relationship between environmental changes and the target task operation status in the park management scenario.

[0022] Preferably, the specific implementation process of marking the temporal feature slice containing the mutation feature parameter as an environmental interaction candidate node is as follows: When the mutation feature parameter exists in the current comparison slice, firstly, the mutation source field, mutation location field, mutation time field, mutation category field, and task association field in the mutation feature parameter are read; then, the above fields are matched with the acquisition source field, scene location field, sampling time label, and task stage field in the current comparison slice to confirm that the mutation feature parameter does indeed originate from the running feature unit carried by the current comparison slice; then, the current comparison slice that has completed the source matching process is marked as the environmental interaction candidate node. The technical essence of the environmental interaction candidate node is a temporal event anchor point with task context. It is not a single image frame, nor a single sensor value, but candidate node data that binds the mutation feature parameter, spatial location in the geographic information management platform, sampling time label, and task stage field. Therefore, it can continue to express its source and function when subsequently spliced ​​into the environmental interaction node sequence.

[0023] Preferably, before marking the candidate nodes for environmental interaction, interference elimination processing is performed on the mutation feature parameters. The interference elimination processing first reads the frame extraction integrity status, camera drawing boundary data, environmental perception device reporting status, and device feedback status in the current comparison slice; wherein, the frame extraction integrity status is determined by whether the camera frame extraction forms readable image data at the camera frame extraction trigger position, and the environmental perception device reporting status is determined by whether the environmental perception device forms readable environmental perception data at the environmental perception device reporting position; subsequently, the frame extraction integrity status, the camera drawing boundary data, the environmental perception device reporting status, and the device feedback status are respectively mapped to the mutation feature parameters. When the mutation feature parameter originates only from the abnormal frame extraction state, the mutation location field being located outside the camera drawing boundary data, the environmental perception device reporting an abnormal state, or the device feedback state only experiencing a state refresh unrelated to the target task, the corresponding current comparison slice is not marked as the environmental interaction candidate node. When the mutation feature parameter can point back to a specific spatial location within the camera drawing boundary data and can correspond to the current execution cycle of the target task through the sampling time label, the corresponding current comparison slice is then marked as the environmental interaction candidate node. Through the above interference elimination processing, the environmental interaction candidate node can be distinguished from ordinary alarm records and ordinary frame extraction records, and its node attributes originate from environmental interaction changes confirmed through spatiotemporal correlation comparison operations.

[0024] Preferably, the specific implementation process for extracting at least two consecutively occurring environmental interaction candidate nodes is as follows: After obtaining multiple environmental interaction candidate nodes, they are first sorted according to the sampling time label corresponding to each environmental interaction candidate node, and it is determined whether they belong to the same park area, adjacent park areas, or consecutive spatial locations on the same target task path based on the mutation position field corresponding to each environmental interaction candidate node; subsequently, it is determined whether each environmental interaction candidate node corresponds to the same execution cycle of the target task based on the task association field corresponding to each environmental interaction candidate node; when at least two adjacent environmental interaction candidate nodes simultaneously satisfy the time continuity relationship, spatial continuity relationship, and task continuity relationship, these at least two environmental interaction candidate nodes are extracted as candidate node fragments. The candidate node fragments can express the continuous evolution of the environmental interaction state during the execution of the target task, rather than expressing an isolated trigger, and are therefore more suitable for participating in subsequent intent deviation detection.

[0025] Preferably, at least two consecutive environmental interaction candidate nodes are extracted because camera frame extraction, environmental perception device reporting, and device feedback status in a park management scenario all constitute instantaneous operating states. A single environmental interaction candidate node is insufficient to express whether the target task execution direction has deviated. Through the temporal continuity between at least two environmental interaction candidate nodes, it can be determined whether the mutation feature parameter continues within a continuous time period; through the spatial continuity between at least two environmental interaction candidate nodes, it can be determined whether the mutation feature parameter extends along the same camera drawing boundary data, the same park area, or the target task path; through the task continuity between at least two environmental interaction candidate nodes, it can be determined whether the mutation feature parameter continuously affects the execution phase of the target task. This continuity determination allows the candidate node fragments to filter out short-term refresh states unrelated to the target task and retain environmental interaction processes that contribute to the subsequent generation of the deviation scalar.

[0026] Preferably, the specific implementation process of concatenating at least two extracted environmental interaction candidate nodes into the environmental interaction node sequence in chronological order is as follows: First, read the sampling time label, mutation location field, mutation category field, and task association field corresponding to each environmental interaction candidate node in the candidate node segment, and determine the node arrangement order according to the sampling time label; then, write a node connection field between two adjacent environmental interaction candidate nodes, the node connection field including the collection source field of the previous environmental interaction candidate node, the collection source field of the next environmental interaction candidate node, the time interval relationship between the two, the spatial connection relationship between the two, and the task stage progression relationship between the two; then, connect at least two environmental interaction candidate nodes into the environmental interaction node sequence according to the node connection field. Through the node connection field, the environmental interaction node sequence can express the direction of change from the previous environmental interaction candidate node to the next environmental interaction candidate node, so that when performing intention deviation detection on the environmental interaction node sequence later, the evolution path of the environmental interaction state can be read along the node arrangement order.

[0027] Preferably, after the environmental interaction node sequence is formed, the mutation feature parameter of each environmental interaction candidate node is retained, and the node connection field between two adjacent environmental interaction candidate nodes is also retained. The mutation feature parameter expresses what kind of environmental interaction change has occurred within a single environmental interaction candidate node, and the node connection field expresses how multiple environmental interaction candidate nodes are continuous in time, space, and task stages. Taking a park security patrol task as an example, when the first environmental interaction candidate node records a "no climbing" behavior recognition state within the boundary data drawn by a certain camera, the second environmental interaction candidate node records an activation change in the smart broadcast feedback state, and the third environmental interaction candidate node records a generation change in the work order flow state, these environmental interaction candidate nodes are sequentially concatenated into the environmental interaction node sequence through the node connection field. This environmental interaction node sequence expresses both the occurrence of environmental events and the operational trajectory of the target task between recognition, linkage, and handling feedback, providing continuous data basis for subsequent comparison of the initial intent map and generation of deviation scalar metrics.

[0028] Optionally, the step of performing intent deviation detection on the environmental interaction node sequence and generating a deviation scalar includes: Obtain the pre-configured initial intent map corresponding to the target task; A semantic node mapping operation is performed on the environmental interaction node sequence to extract the actual execution graph of the environmental interaction node sequence through the semantic node mapping operation; Compare the topological differences between the actual execution map and the initial intention map; Node distance parameters are extracted from the topology differences, and the deviation scalar is generated based on the node distance parameters.

[0029] Preferably, the specific implementation process of obtaining the pre-configured initial intent map corresponding to the target task is as follows: Before the target task enters the execution state, the target task configuration record corresponding to the target task is read. The target task configuration record includes a task type field, a task path field, a device call field, a behavior recognition grouping field, camera drawing boundary data, spatial location in the geographic information management platform, and a task stage field. The task type field distinguishes which type of task the target task belongs to: security patrol, anomaly monitoring, broadcast linkage, lighting linkage, intelligent irrigation linkage, or work order flow. The task path field expresses the execution route of the target task in the spatial location of the geographic information management platform. The device call field expresses the camera frame extraction, environmental perception device reporting, intelligent security recognition results, intelligent broadcast feedback status, intelligent lighting feedback status, intelligent irrigation feedback status, or work order flow status that the target task needs to call. Through the above target task configuration record, the task stage fields that the target task should go through and the running feature units that should be received in the normal execution state can be determined first, providing a data source for the subsequent configuration of the initial intent map.

[0030] Preferably, the technical essence of the initial intent map is a data map used to express the expected execution structure of the target task when no intent deviation occurs, rather than a simple task list. The initial intent map includes intent nodes and intent-related edges; the intent nodes are jointly configured by the task stage field, device call field, behavior recognition grouping field, camera drawing boundary data, and spatial location in the geographic information management platform in the target task configuration record. The intent nodes include node semantic fields, node time fields, node spatial fields, and node stage fields. The node semantic field is used to express the expected semantic category corresponding to the intent node. The node time field is used to express the expected time position of the intent node in the target task. The node spatial field is used to express the spatial location in the camera drawing boundary data and geographic information management platform corresponding to the intent node. The node stage field is used to express the task stage field corresponding to the intent node. The intent-related edges are jointly configured by the temporal succession relationship, spatial succession relationship, and task stage progression relationship between two adjacent intent nodes. Through the intent nodes, the initial intent graph can express what kind of runtime feature units the target task should read in a certain execution stage; through the intent association edges, the initial intent graph can express the expected execution node path of the target task from one execution stage to the next execution stage, so that the initial intent graph can serve as a reference object for subsequent topology difference comparison.

[0031] Preferably, when pre-configuring the initial intent map, firstly, a corresponding task stage field combination is selected based on the task type field. This task stage field combination is formed by arranging multiple task stage fields according to the expected execution order of the target task. Then, based on the task path field, the task stage field combination is mapped to a spatial location in the geographic information management platform to form an intent node with spatial constraints. Subsequently, based on the behavior recognition grouping field, behavior recognition categories such as water hazard, ground litter, no fishing, no climbing, sidewalk cycling, road litter, or illegal occupation of roads are written into the corresponding intent node, enabling the intent node to express the behavior recognition categories that the target task should recognize within the boundary data drawn by the corresponding camera. Then, based on the device invocation field, the smart broadcast feedback status, smart lighting feedback status, smart irrigation feedback status, or work order flow status are written into the intent node corresponding to the execution stage. These intent nodes are further connected through intent association edges, enabling the initial intent map to express the expected execution structure of the target task from environmental perception, behavior recognition, device linkage to work order flow.

[0032] Preferably, in a scenario, if the target task is a security patrol task, the task stage field in the target task configuration record can be configured according to the execution order of patrol, identification, alarm, linkage, and handling feedback to form a combination of task stage fields corresponding to the security patrol task. When configuring the initial intention graph, first, the spatial location of the park area corresponding to the patrol and the geographic information management platform is written into the node spatial field of the first intention node, and then the camera drawing boundary data and behavior recognition grouping field are written into the intention node corresponding to identification. When the behavior recognition grouping field includes prohibiting climbing, the smart security identification result corresponding to prohibiting climbing is written into the node semantic field of the intention node corresponding to identification. Subsequently, the smart broadcast feedback status is written into the intention node corresponding to linkage, and the work order flow status is written into the intention node corresponding to handling feedback. The above multiple intention nodes are connected by the intention association edges according to the expected execution order of the task stage fields, and the initial intention graph formed thereby can express the expected execution node path of the security patrol task under normal execution state.

[0033] Preferably, the specific implementation process of performing semantic node mapping operation on the environmental interaction node sequence is as follows: First, read the mutation feature parameters, sampling time label, task stage field, camera drawing boundary data, spatial location in the geographic information management platform, and node acceptance field carried by each environmental interaction candidate node in the environmental interaction node sequence; then, determine the environmental interaction candidate node originating from camera frame extraction, environmental perception device reporting, smart security recognition results, smart broadcast feedback status, smart lighting feedback status, smart irrigation feedback status, or work order flow status based on the mutation source field in the mutation feature parameters; then, determine whether the environmental interaction candidate node corresponds to a change in behavior recognition category, change in environmental perception data status, change in device feedback status, or change in task stage status based on the mutation category field in the mutation feature parameters. Through the above reading and judgment, the semantic node mapping operation can map each environmental interaction candidate node to an execution node with the same semantic scope as the intent node in the initially set intent graph.

[0034] Preferably, when forming the execution node, the semantic node mapping operation does not directly copy the original fields of the environmental interaction candidate nodes into the graph content. Instead, it first performs semantic merging processing on the mutation feature parameters in the environmental interaction candidate nodes. This semantic merging processing categorizes changes in behavior recognition categories into behavior recognition semantics, changes in environmental perception data states into environmental perception recognition semantics, changes in device feedback states into device linkage semantics, and changes in task stage states into task flow semantics. Subsequently, the behavior recognition semantics, the environmental perception recognition semantics, the device linkage semantics, or the task flow semantics are written into the node semantic field of the execution node, the sampling time label is written into the node time field of the execution node, the camera drawing boundary data and the spatial location in the geographic information management platform are written into the node spatial field of the execution node, and the task stage field is written into the node stage field of the execution node. The resulting execution node has the same field caliber as the intent node, enabling subsequent topology difference comparisons to focus on the same type of node attributes, rather than comparing the original video footage, sensor values, and device feedback states together.

[0035] Preferably, the specific implementation process of extracting the actual execution graph of the environmental interaction node sequence through the semantic node mapping operation is as follows: After multiple execution nodes are formed, the node connection field between two adjacent environmental interaction candidate nodes in the environmental interaction node sequence is read first. The node connection field includes the acquisition source field of the previous environmental interaction candidate node, the acquisition source field of the next environmental interaction candidate node, the time interval relationship, the spatial connection relationship, and the task stage progression relationship. Subsequently, according to the node connection field, execution association edges are written between the corresponding two adjacent execution nodes. The execution association edges are used to express the actual connection relationship between two execution nodes in time, space, and task stage. The actual connection relationship is jointly defined by the time interval relationship, the spatial connection relationship, and the task stage progression relationship. The execution association edges and the execution nodes together form the actual execution graph. The technical essence of the actual execution graph is the actual execution structure reconstructed from the environmental interaction node sequence. It expresses the actual execution node path formed by the target task under the influence of environmental interaction changes, rather than statically recording a single environmental interaction candidate node.

[0036] Preferably, after the actual execution graph is formed, the node source correspondence between the execution node and the original environment interaction candidate node is retained, and the edge source correspondence between the execution-related edge and the node-accepting field is also retained. The node source correspondence is used to backtrack the corresponding mutation feature parameters, camera-drawn boundary data, spatial location in the geographic information management platform, and task stage fields when topological differences are subsequently discovered. The edge source correspondence is used to backtrack the time interval relationship, spatial acceptance relationship, and task stage progression relationship between two adjacent environment interaction candidate nodes when extracting node distance parameters. Through the node source correspondence and the edge source correspondence, the actual execution graph will not lose its source relationship with the environment interaction node sequence when participating in intent deviation detection, enabling subsequent deviation scalars to point back to specific running feature units and specific task stage fields.

[0037] Preferably, the specific implementation process of comparing the topological differences between the actual execution graph and the initial intention graph is as follows: First, based on the node semantic field, node spatial field, and node stage field of the execution node, the intention node corresponding to the field definition is searched in the initial intention graph to form a node correspondence; then, based on the time interval relationship, spatial continuity relationship, and task stage progression relationship expressed by the execution association edge, the intention association edge corresponding to the field definition is searched in the initial intention graph to form an edge correspondence; then, based on the node correspondence and the edge correspondence, it is determined whether the actual execution graph has missing nodes, added nodes, misaligned node order, spatial path offset, or task stage jump relative to the initial intention graph. The aforementioned missing nodes, added nodes, misaligned node order, spatial path offset, and task stage jump together constitute the topological difference, enabling the topological difference to express the deviation position and deviation type of the actual execution structure relative to the expected execution structure.

[0038] Preferably, the topological difference is not simply determined by whether the two graphs are identical, but rather by comparing the execution association edges of the actual execution graph segment by segment along the intent association edges of the initial intent graph. Specifically, starting from the initial intent node in the initial intent graph, the intent association edges subsequently connected to that initial intent node are read, and execution nodes with a node correspondence to that initial intent node are found in the actual execution graph. Then, the execution association edges subsequently connected to that execution node are compared with the corresponding intent association edges using the same criteria to determine whether the execution association edges maintain the expected temporal continuity, spatial continuity, and task stage progression. If the execution node pointed to by the execution association edge does not have a node correspondence with the intent node pointed to by the intent association edge, this is recorded as a node sequence misalignment or a task stage jump. If the node spatial field corresponding to the execution association edge deviates from the spatial continuity corresponding to the intent association edge, this is recorded as a spatial path offset. Through the above segment-by-segment comparison, the topological difference can reflect the specific location where the target task deviates from the expected execution node path during actual execution.

[0039] Preferably, the specific implementation process of extracting node distance parameters from the topology differences is as follows: First, read the node missing, node added, node sequence misalignment, spatial path offset, and task stage jump in the topology differences, and determine the difference nodes corresponding to each type of topology difference; then, compare the position of the difference node in the actual execution graph with the position of the corresponding intention node in the initial intention graph to form node distance parameters. The technical essence of the node distance parameters is a structured distance parameter used to describe the deviation of the difference node from the expected execution node path. The node distance parameters include node semantic distance, node temporal distance, node spatial distance, and node stage distance; wherein, the node semantic distance is used to express the degree of difference between the node semantic field of the execution node and the node semantic field of the intention node, the node temporal distance is used to express the degree of difference between the node temporal field of the execution node and the node temporal field of the intention node, the node spatial distance is used to express the degree of difference between the node spatial field of the execution node and the node spatial field of the intention node, and the node stage distance is used to express the degree of difference between the node stage field of the execution node and the node stage field of the intention node.

[0040] Preferably, when extracting the node distance parameter, the distance content of different field dimensions is processed separately, and node semantic fields, node time fields, node spatial fields, and node stage fields are not directly mixed in the calculation. For the node semantic distance, the node semantic field of the executing node and the node semantic field of the intent node are first compared to see if they belong to the same semantic category, and then the node semantic distance is formed according to the task association relationship between the semantic categories; for the node time distance, the node time field of the executing node and the node time field of the intent node are first read, and then the node time distance is formed according to their sequential offset relationship under the same time dimension; for the node spatial distance, the node spatial field of the executing node and the node spatial field of the intent node are first unified to the spatial location in the geographic information management platform, and then the node spatial distance is formed according to their offset relationship under the same spatial reference; for the node stage distance, the node stage field of the executing node and the node stage field of the intent node are first read, and then the node stage distance is formed according to their progressive relationship in the task stage field. Through the above field-specific processing method, the node distance parameter can avoid the direct comparison of data with different dimensions.

[0041] Preferably, the specific implementation process of generating the deviation scalar based on the node distance parameters is as follows: First, read the node distance parameters corresponding to each differential node, and determine the participation order of the node semantic distance, node temporal distance, node spatial distance, and node stage distance in the target task according to the task type field, behavior recognition grouping field, device call field, and task stage field in the target task configuration record; then, perform a standardization process on the node distance parameters according to the participation order to obtain the node deviation description item. The standardization process is used to organize the node semantic distance, node temporal distance, node spatial distance, and node stage distance into the same deviation description standard, but the standardization process does not change the field source of the node semantic distance, node temporal distance, node spatial distance, and node stage distance respectively; then, form the deviation scalar based on the occurrence position, duration, and number of associated execution nodes of the node deviation description item in the actual execution graph. The technical essence of the deviation scalar is to converge the structured deviation degree of multiple difference nodes into a numerical parameter that can be compared with a preset trigger baseline, so that subsequent determination of whether to generate a permission suspension instruction can be made based on the deviation scalar.

[0042] Preferably, the deviation scalar, during its generation, reflects not only the degree of deviation of a single divergent node but also the continuity of multiple divergent nodes within the actual execution graph. Specifically, if the node distance parameter corresponds to only one execution node, and that execution node does not continuously influence subsequent execution nodes through execution-related edges, then the node distance parameter only forms a local node deviation description. If multiple execution nodes correspond to the node distance parameter, and these nodes exhibit continuous temporal, spatial, or task-stage progression relationships through execution-related edges, then the multiple node distance parameters are merged to form a continuous node deviation description. Subsequently, the deviation scalar is formed based on the local node deviation description and the continuous node deviation description. In this way, the deviation scalar can distinguish between an isolated environmental interaction change and an environmental interaction change that continuously affects the path of the actual execution node, thereby reducing the likelihood of a permission suspension command being triggered solely due to a single frame-drop anomaly or a single environmental awareness device reporting anomaly.

[0043] Preferably, in a security patrol task, if the expected execution node path in the initial intent map is a sequential connection of patrol intent node, identification intent node, alarm intent node, linkage intent node, and handling feedback intent node, and the actual execution map shows that after a prohibited climbing execution node appears within the boundary data drawn by a certain camera, no execution node corresponding to the smart broadcast feedback state is formed, but an execution node corresponding to the work order flow state is directly formed, then the topology difference includes node missing and task stage jump; further, when extracting node distance parameters from the node missing and task stage jump, the missing linkage intent node is mapped to node semantic distance and node stage distance, and the execution node directly entering the work order flow state is mapped to node sequence misalignment and node stage distance; subsequently, the deviation scalar is formed based on the above node distance parameters. In this example, the deviation scalar is not directly formed by the single behavior identification category of prohibited climbing, but by the topology difference between the actual execution map and the initial intent map, thus expressing the degree of deviation of the actual execution node path.

[0044] Preferably, in the intelligent irrigation linkage task, if the initial intent map requires the environmental sensing device to first report and form intent nodes corresponding to environmental sensing and identification semantics, and then the intelligent irrigation feedback state to form intent nodes corresponding to device linkage semantics, but the actual execution map shows that execution nodes corresponding to changes in environmental sensing data state appear continuously, but no execution nodes corresponding to the intelligent irrigation feedback state are formed, then the topology difference includes node missingness and spatial path offset; when extracting the node distance parameter, firstly, the node spatial distance is formed based on the difference in node spatial fields between the execution nodes corresponding to changes in environmental sensing data state and the intent nodes corresponding to environmental sensing and identification semantics, and then the node semantic distance and node stage distance are formed based on the missing intent nodes corresponding to device linkage semantics; subsequently, the deviation scalar is formed based on the above node distance parameters. Thus, the deviation scalar can express the deviation state where the environmental sensing data has changed but the actual execution node path has not entered the corresponding device linkage stage.

[0045] Preferably, after the deviation scalar is formed, the parameter source relationship between the deviation scalar and the node distance parameter is retained, as well as the difference source relationship between the node distance parameter and the topology difference. The parameter source relationship is used to trace back which type of distance content—node semantic distance, node temporal distance, node spatial distance, or node stage distance—the deviation scalar originates from when subsequently determining whether the deviation scalar has reached the preset trigger baseline. The difference source relationship is used to trace back the execution node, execution associated edge, intent node, and intent associated edge corresponding to the topology difference after generating the permission suspension instruction. Through the parameter source relationship and the difference source relationship, when recording the on-site stack snapshot, the task deviation source that triggers the permission suspension state can be written into the task state source information, enabling the on-site stack snapshot to inherit the processing result of the intent deviation detection.

[0046] Optionally, recording a snapshot of the agent's stack trace in the current execution cycle includes: In the permission suspension state corresponding to the permission suspension instruction, the internal bus data stream of the intelligent agent is intercepted; Extract the global and local variable parameters from the internal bus data stream; Capture the external interface call records generated by the intelligent agent during the current execution cycle; The global variable parameters, the local variable parameters, the external interface call records, and the timestamp tag corresponding to the current execution cycle are encapsulated into the on-site stack snapshot.

[0047] Preferably, the specific implementation process of intercepting the internal bus data stream of the agent in the permission suspension state corresponding to the permission suspension instruction is as follows: After the deviation scalar reaches the preset trigger baseline and the permission suspension instruction is generated, the target task identifier, current execution cycle identifier, permission suspension trigger time position, deviation scalar that triggered the permission suspension instruction, and task deviation source corresponding to the deviation scalar are read according to the permission suspension instruction; then, the running permission of the agent in the current execution cycle is switched to the permission suspension state, so that the agent suspends the ability to issue drive instructions to the target task, and retains the running state formed by the task scheduling center, main working memory area, and drive instruction issuance port before the permission suspension trigger time position. The internal bus data stream refers to a data stream formed in chronological order of status data, task data, device call data, and feedback data transmitted between different execution units within the agent. It includes target task scheduling messages, real-time running feature stream transmission messages, environmental interaction node sequence transmission messages, deviation metric scalar transmission messages, device call status transmission messages, and work order flow status transmission messages. The deviation metric scalar transmission messages carry the deviation metric scalar formed by intent deviation detection, enabling the deviation metric scalar to enter the subsequent snapshot recording process through the internal bus data stream. By intercepting the internal bus data stream in a permission-suspended state, the task execution context of the agent within the current execution cycle can be obtained without restarting the target task.

[0048] Preferably, the interception of the internal bus data stream does not involve severing the execution link of the target task, but rather establishing a read-only transfer relationship for the internal bus data stream under the permission suspension state. Specifically, the target task data segment corresponding to the target task in the internal bus data stream is first determined based on the current execution cycle identifier. Then, the truncation boundary in the target task data segment is determined based on the permission suspension trigger time position. Subsequently, the target task scheduling message, real-time running feature stream transmission message, environment interaction node sequence transmission message, deviation metric scalar transmission message, device call status transmission message, and work order flow status transmission message located before the truncation boundary and not yet covered by subsequent execution stages are transferred to the snapshot carrier record. The snapshot carrier record is a record object used to carry the intermediate data required for the formation of the field stack snapshot. The snapshot carrier record is not equivalent to the field stack snapshot; subsequent field alignment processing, source index establishment processing, and structure encapsulation processing need to be performed based on the snapshot carrier record to form the field stack snapshot. Through the snapshot carrier record, the task status in the internal bus data stream is not saved as an isolated log, but is retained as intermediate data that can continue to participate in the field stack snapshot encapsulation.

[0049] Preferably, the technical essence of the current execution cycle is the continuous execution interval formed by the agent around the same target task from the triggering of a task scheduling to the change of permission status. The current execution cycle is not equivalent to a fixed-length time slice, but is jointly defined by the target task scheduling message, the real-time running feature stream, the environmental interaction node sequence, the deviation metric scalar, and the permission suspension instruction. Specifically, the target task scheduling message defines the task source of the current execution cycle; the real-time running feature stream defines the environmental perception input source within the current execution cycle; the environmental interaction node sequence defines the source of environmental interaction changes within the current execution cycle; the deviation metric scalar defines the task deviation source that triggers the permission suspension state within the current execution cycle; and the permission suspension instruction defines the time boundary for the transition from continued execution to permission suspension state within the current execution cycle. Through the above definition, the subsequent extraction of global variable parameters, local variable parameters, and external interface call records are all within the same current execution cycle, preventing data from other target tasks or other execution cycles from being mixed into the on-site stack snapshot.

[0050] Preferably, the specific implementation process for extracting global and local variable parameters from the internal bus data stream is as follows: First, read the target task scheduling message, the real-time running feature stream transmission message, the environment interaction node sequence transmission message, the deviation metric scalar transmission message, and the device call status transmission message from the snapshot carrying record. Then, divide the parameters in the above messages into global variable parameters and local variable parameters according to the parameter scope. The technical essence of the global variable parameters is a state parameter that can span multiple execution stages within the current execution cycle and continuously affect the running direction of the target task. It includes target task identifier, task type field, task path field, device call field, behavior recognition grouping field, spatial location in the geographic information management platform, camera drawing boundary data, permission status field, deviation metric scalar, preset trigger baseline, and task status source information. Among them, the task status source information is used to receive the task deviation source and write the task deviation source into the source field of the global variable parameters. The global variable parameters are used to express the overall task running context of the target task when encapsulating the field stack snapshot, so that the field stack snapshot can retain the task-level state before the permission suspension state occurs.

[0051] Preferably, the technical essence of the local variable parameters is that they are stage-specific parameters that only take effect during a certain execution phase, a certain device invocation process, or a certain environmental interaction change within the current execution cycle. These include time-series feature slices, mutation feature parameters, environmental interaction candidate nodes, node receiving fields, execution nodes, execution association edges, node distance parameters, node deviation descriptions, smart security identification results, smart broadcast feedback status, smart lighting feedback status, smart irrigation feedback status, and work order flow status. Specifically, during extraction, the running feature units in the snapshot record are first divided into execution phases corresponding to inspection, identification, alarm, linkage, or handling feedback based on the task phase field. Then, the local variable parameters are read from each execution phase. Subsequently, a phase source correspondence is established between the read local variable parameters and the corresponding task phase field. Through this phase source correspondence, the local variable parameters can return to their respective execution phase when encapsulating the field stack snapshot, avoiding confusion with the global variable parameters and preventing them from becoming isolated parameters outside the current execution cycle.

[0052] Preferably, in the security patrol task of a smart garden, the global variable parameters may include the task path field corresponding to the security patrol task, camera drawing boundary data, behavior recognition grouping field, preset trigger baseline, deviation metric scalar, and permission status field; the local variable parameters may include a temporal feature slice formed by frame extraction from a certain camera, a mutation feature parameter in the temporal feature slice, an environmental interaction candidate node marked by the mutation feature parameter, and an environmental interaction node sequence formed by connecting multiple environmental interaction candidate nodes. If the environmental interaction node sequence forms a deviation metric scalar after intent deviation detection, and the deviation metric scalar triggers the permission suspension instruction, then the deviation metric scalar is entered into the snapshot carrying record as a global variable parameter, and the mutation feature parameter and the environmental interaction candidate node are entered into the snapshot carrying record as local variable parameters. The above parameter division enables the on-site stack snapshot to express both the overall task execution context of the security patrol task and the specific environmental interaction changes that lead to the permission suspension state.

[0053] Preferably, the specific implementation process of capturing the external interface call records formed by the intelligent agent within the current execution cycle is as follows: In the permission suspension state, firstly, interface call messages that interact with external interface objects are read from the internal bus data stream. The external interface objects include camera frame extraction interface, environmental perception device reporting interface, smart broadcast feedback interface, smart lighting feedback interface, smart irrigation feedback interface, work order flow interface, and geographic information management platform spatial location interface; then, interface call messages belonging to the current execution cycle are filtered according to the current execution cycle identifier; subsequently, the interface source field, interface target field, interface input parameter field, interface return field, call time field, call status field, and call result field are extracted from the filtered interface call messages to form the external interface call record. The technical essence of the external interface call record is to express the external interaction trajectory formed when the agent calls the external interface object within the current execution cycle. Specifically, the interface source field and the interface target field express the call direction of the external interface call record; the interface input parameter field and the interface return field express the data transmission content of the external interface call record; the call time field expresses the call time position of the external interface call record within the current execution cycle; and the call status field and the call result field express the execution status of the external interface call record. When subsequently encapsulating the context stack snapshot, the external interface call record is used to indicate which parameters among the global and local variable parameters originate from the external interface call and which parameters originate from the agent's internal task execution.

[0054] Preferably, the process of capturing external interface call records further includes sorting the interface call order. Specifically, multiple external interface call records are first sorted according to the call time field, and then the call direction corresponding to each external interface call record is determined according to the interface source field and the interface target field. Subsequently, the call direction is mapped to the task stage field to form an interface stage correspondence. This interface stage correspondence expresses which execution stage (inspection, identification, alarm, linkage, or handling feedback) a particular external interface call record occurs in; for example, the external interface call record corresponding to the camera frame extraction interface typically corresponds to the identification stage, the external interface call record corresponding to the smart broadcast feedback interface typically corresponds to the linkage stage, and the external interface call record corresponding to the work order flow interface typically corresponds to the handling feedback stage. Through this interface stage correspondence, the external interface call record, together with the local variable parameters, can describe the external interaction situation of each execution stage within the current execution cycle.

[0055] Preferably, the timestamp tag is uniformly generated based on the current execution cycle before encapsulating the on-site stack snapshot. Specifically, the start time position of the current execution cycle, the permission suspension trigger time position, the sampling timestamp of each running feature unit, and the call time field of each external interface call record are first read. Then, the above time information is arranged according to the same time base to form a timestamp tag. The timestamp tag is not a single moment value, but rather time indexing data used to mark the time position of the global variable parameters, the local variable parameters, and the external interface call records within the current execution cycle. When encapsulating the on-site stack snapshot, the timestamp tag establishes a time correspondence with the global variable parameters, the local variable parameters, and the external interface call records, enabling the subsequent generation of a human-computer interaction probe based on the on-site stack snapshot to reconstruct the execution process of the target task before entering the permission suspension state in chronological order.

[0056] Preferably, the specific implementation process of encapsulating the global variable parameters, the local variable parameters, the external interface call records, and the timestamp tag corresponding to the current execution cycle into the on-site stack snapshot is as follows: First, establish a task-level field area, a stage-level field area, an interface-level field area, and a timestamp field area in the snapshot carrier record; then, write the global variable parameters into the task-level field area, write the local variable parameters into the stage-level field area according to the stage source correspondence, write the external interface call records into the interface-level field area according to the interface stage correspondence, and write the timestamp tag into the timestamp field area; then, establish a time correspondence relationship between the data in the task-level field area, the stage-level field area, and the interface-level field area according to the timestamp tag, and establish a source index according to the task status source information. Based on the snapshot carrier record with the completion time correspondence relationship and the source index, perform structure encapsulation processing to form the on-site stack snapshot.

[0057] Preferably, the structure encapsulation process specifically includes field validation, field arrangement, and field indexing. The field validation first checks whether the target task identifier, task type field, task path field, and permission status field in the task-level field area correspond to the current execution cycle identifier. Then, it checks whether the local variable parameters in the stage-level field area can be referenced back to the corresponding task stage field through the stage source correspondence, and whether the external interface call records in the interface-level field area can be referenced back to the corresponding task stage field through the interface stage correspondence. After the field validation passes, the data in the task-level field area, stage-level field area, and interface-level field area are arranged according to the timestamp label to form a snapshot field sequence. Subsequently, a source index and a time index are established for the snapshot field sequence, enabling the snapshot field sequence to reference the internal bus data stream, the global variable parameters, the local variable parameters, the external interface call records, and the timestamp label. After completing the source index and time index, the snapshot field sequence is encapsulated as the on-site stack snapshot.

[0058] Preferably, the technical essence of the on-site stack snapshot is a structured snapshot data used to describe the task execution context, parameter state context, and external interface call context of the agent when the permission suspension state occurs. The on-site stack snapshot includes a task-level field area, a stage-level field area, an interface-level field area, a timestamp field area, a source index, and a time index; wherein, the task-level field area carries the global variable parameters, the stage-level field area carries the local variable parameters, the interface-level field area carries the external interface call records, the timestamp field area carries the timestamp tag, the source index indicates which type of data in the on-site stack snapshot originates from—the internal bus data stream, the global variable parameters, the local variable parameters, or the external interface call records—and the time index indicates the time position of the data in the on-site stack snapshot within the current execution cycle. Through this structure, the on-site stack snapshot is not an ordinary log file, but a state carrier capable of subsequently generating human-computer interaction probes, presenting a panoramic restoration view, and reconstructing the on-site stack snapshot.

[0059] Preferably, in the example corresponding to the security patrol task, when the permission suspension instruction is triggered by the deviation scalar corresponding to the prohibition of climbing, the target task data segment in the internal bus data stream corresponding to the security patrol task is transferred to the snapshot carrier record; subsequently, the task type field, task path field, camera drawing boundary data, preset trigger baseline, deviation scalar, and permission status field are extracted from the snapshot carrier record to form the global variable parameters; then, the time-series feature slice, mutation feature parameter, environmental interaction candidate node, environmental interaction node sequence, execution node, node distance parameter, and node deviation description item are extracted to form the local variable parameters; next, the external interface call records corresponding to the camera frame extraction interface, smart security recognition result, smart broadcast feedback interface, and work order flow interface are captured; finally, the global variable parameters, the local variable parameters, the external interface call records, and the timestamp tag are structurally encapsulated through the snapshot carrier record to form the on-site stack snapshot. This on-site stack snapshot can express the environmental interaction changes that the agent has identified before the permission suspension state occurs, the intention deviation detection results that have been formed, and the external interface objects that have been called, providing a continuous data foundation for the subsequent generation of human-computer interaction probes based on the on-site stack snapshot.

[0060] Optionally, generating the human-computer interaction probe based on the on-site stack snapshot includes: Obtain the preset human-computer interaction probe pruning rules, and perform data pruning on the field stack snapshot according to the human-computer interaction probe pruning rules to obtain a simplified state feature string; Obtain a preset visual rendering pipeline configuration, and transmit the simplified state feature string to the visual rendering pipeline according to the preset visual rendering pipeline configuration to generate three-dimensional spatial scene data for the hardware console to present the panoramic restored view; Identify the interactive anchor points in the three-dimensional spatial scene data; The human-computer interaction probe is generated based on the simplified state feature string, the three-dimensional spatial scene data, and the interactive anchor points.

[0061] Preferably, the specific implementation process of obtaining the preset human-computer interaction probe pruning rules is as follows: Before the target task enters the execution state, the target task configuration record, hardware console presentation configuration record, and on-site stack snapshot field configuration record corresponding to the target task are read first. The target task configuration record provides fields for task type, task path, device call, behavior recognition grouping, camera drawing boundary data, spatial location in the geographic information management platform, and task stage. The hardware console presentation configuration record provides the view hierarchy, scene object type, and interactive control type that the hardware console can present. The on-site stack snapshot field configuration record provides the field scope for task-level, stage-level, interface-level, time index, source index, and time index. Subsequently, the human-computer interaction probe pruning rules are configured based on the target task configuration record, the hardware console presentation configuration record, and the on-site stack snapshot field configuration record. The technical essence of the human-computer interaction probe pruning rules is to filter field retention rules from the on-site stack snapshot that can restore the target task's running state, environmental interaction changes, and external interface call state, rather than simply compressing the on-site stack snapshot.

[0062] Preferably, the human-computer interaction probe pruning rules include task-level field retention guidelines, stage-level field retention guidelines, interface-level field retention guidelines, timestamp field retention guidelines, source index retention guidelines, and time index retention guidelines; wherein, the task-level field retention guidelines are used to retain the target task identifier, task type field, task path field, permission status field, deviation metric scalar, preset trigger baseline, and task status source information from the task-level field area; the stage-level field retention guidelines are used to retain temporal feature slices, mutation feature parameters, and environmental interaction candidates from the stage-level field area. The data pruning process includes nodes, environmental interaction node sequences, execution nodes, node distance parameters, and node deviation descriptions. The interface-level field retention criteria are used to retain external interface call records corresponding to camera frame extraction interfaces, environmental sensing device reporting interfaces, smart broadcast feedback interfaces, smart lighting feedback interfaces, smart irrigation feedback interfaces, work order flow interfaces, and geographic information management platform spatial location interfaces. The timestamp field retention criteria are used to retain the timestamp labels. The source index retention criteria are used to retain the source index. The time index retention criteria are used to retain the time index. These retention criteria collectively define the retention scope of the data pruning process, enabling the simplified state feature string to encompass the state content related to the panoramic restoration view in the on-site stack snapshot.

[0063] Preferably, the specific implementation process of pruning the data of the field stack snapshot according to the human-computer interaction probe pruning rules is as follows: First, read the task-level field area, stage-level field area, interface-level field area, timestamp field area, source index, and time index from the field stack snapshot, and perform field matching on the above field areas and indices according to the human-computer interaction probe pruning rules; then, write global variable parameters that conform to the retention criteria of the task-level fields into the task state pruning fragment, write local variable parameters that conform to the retention criteria of the stage-level fields into the stage state pruning fragment, write external interface call records that conform to the retention criteria of the interface-level fields into the interface state pruning fragment, write timestamp tags that conform to the retention criteria of the timestamp fields and time indexes that conform to the retention criteria of the time indexes into the time state pruning fragment, and write source indexes that conform to the retention criteria of the source indexes into the source state pruning fragment. The aforementioned task state pruning fragments, stage state pruning fragments, interface state pruning fragments, time state pruning fragments, and source state pruning fragments are not independent results of each other. Instead, they are arranged by the time index in the time state pruning fragments and maintain the source correspondence with the on-site stack snapshot by the source index in the source state pruning fragments.

[0064] Preferably, the data pruning process further includes redundant field exclusion processing. Specifically, firstly, based on the time index in the time state pruning segment, it is determined whether there are duplicate timestamp tags in the on-site stack snapshot. Then, based on the source index in the source state pruning segment, it is determined whether the duplicate timestamp tags point to the same running feature unit, the same external interface call record, or the same local variable parameters. When the data content corresponding to the duplicate timestamp tags is the same and the task stage field has not changed, the duplicate data content is excluded as a redundant field. When the data content corresponding to the duplicate timestamp tags comes from the same external interface object, but its call status field or call result field changes, the data content is retained in the interface state pruning segment. Through the above redundant field exclusion processing, the data pruning process can reduce duplicate fields unrelated to the panoramic restoration view, while retaining field content that reflects changes in the target task's running status.

[0065] Preferably, the specific implementation process of obtaining the simplified state feature string is as follows: After forming the task state pruning fragment, the stage state pruning fragment, the interface state pruning fragment, the time state pruning fragment, and the source state pruning fragment, the task state pruning fragment, the stage state pruning fragment, and the interface state pruning fragment are first arranged in order according to the timestamp label and time index in the time state pruning fragment. Then, the corresponding source index is configured for each arranged field content according to the source index in the source state pruning fragment. Subsequently, the field content that has completed the sequential arrangement and source index configuration is encapsulated into the simplified state feature string. The technical essence of the simplified state feature string is a renderable state data string obtained by pruning the on-site stack snapshot, which includes task state pruning fragment, stage state pruning fragment, interface state pruning fragment, time state pruning fragment, and source state pruning fragment. Therefore, the simplified state feature string retains both the overall running state of the target task and the environmental interaction changes and external interface call states that cause the permission suspension state.

[0066] Preferably, in the scenario corresponding to the security patrol task, if the on-site stack snapshot records the mutation feature parameter corresponding to the prohibition of climbing, the environmental interaction candidate node corresponding to the mutation feature parameter, the camera drawing boundary data to which the environmental interaction candidate node belongs, the deviation scalar corresponding to the environmental interaction node sequence, the external interface call record corresponding to the smart broadcast feedback interface, and the external interface call record corresponding to the work order flow interface, then the data pruning process first writes the deviation scalar, the preset trigger baseline, and the permission status field into the task status pruning fragment, then writes the mutation feature parameter, the environmental interaction candidate node, and the environmental interaction node sequence into the stage status pruning fragment, and writes the external interface call record corresponding to the smart broadcast feedback interface and the external interface call record corresponding to the work order flow interface into the interface status pruning fragment. Subsequently, the time status pruning fragment records the time position of the above field content within the current execution cycle, and the source status pruning fragment records the source index of the above field content in the on-site stack snapshot, so that the simplified status feature string can express the state evolution of the security patrol task from anomaly identification to the permission suspension state before it occurs.

[0067] Preferably, the specific implementation process of obtaining the preset visual rendering pipeline configuration is as follows: the spatial coordinate mapping standard, scene object mapping standard, view hierarchy arrangement standard, state style mapping standard and interactive anchor point generation standard are pre-configured in the configuration record of the hardware console to form the preset visual rendering pipeline configuration. The spatial coordinate mapping method is used to convert spatial location, camera drawing boundary data, and task path fields in the geographic information management platform into scene coordinate data in the 3D spatial scene data. The scene object mapping method is used to map the camera frame extraction interface, environmental perception device reporting interface, smart broadcast feedback interface, smart lighting feedback interface, smart irrigation feedback interface, work order flow interface, and geographic information management platform spatial location interface into scene object data in the 3D spatial scene data. The view hierarchy arrangement method is used to arrange the task stage field, timestamp label, and environmental interaction node sequence into view hierarchy data in the 3D spatial scene data. The state style mapping method is used to convert the deviation scalar, permission state field, call state field, and call result field into state presentation data in the 3D spatial scene data. The interactive anchor point generation method is used to provide field sources for subsequently determining the interactive anchor points based on the interactive control type, the source index, and the time index. The technical essence of the preset visual rendering pipeline configuration is to convert the simplified state feature string into configuration rules for a panoramic restored view readable data structure.

[0068] Preferably, the technical essence of the visual rendering pipeline is a data processing flow that performs spatial transformation, object generation, hierarchical arrangement, and state binding on the simplified state feature string according to the preset visual rendering pipeline configuration. Specifically, firstly, based on the spatial coordinate mapping standard, spatial location, camera drawing boundary data, and task path fields from the geographic information management platform are read from the simplified state feature string to form scene coordinate data; then, based on the scene object mapping standard, external interface call records are read from the interface state pruning fragment, and the external interface objects corresponding to the external interface call records are converted into scene object data; next, based on the view hierarchy arrangement standard, task stage fields, environmental interaction node sequences, timestamp labels, and time indices are read from the stage state pruning fragment and the time state pruning fragment to form view hierarchy data; finally, based on the state style mapping standard, deviation scalar, permission state fields, call state fields, and call result fields are bound to the corresponding scene object data and view hierarchy data. The aforementioned scene coordinate data, scene object data, view hierarchy data, and state presentation data together form the three-dimensional spatial scene data.

[0069] Preferably, the specific implementation process of generating the 3D spatial scene data for the hardware console to present the panoramic restored view is as follows: After obtaining the scene coordinate data, the visual rendering pipeline first arranges the spatial location in the geographic information management platform, the camera drawing boundary data, and the task path field in the scene coordinate data into a spatial scene skeleton; then, the scene object data is attached to the corresponding position in the spatial scene skeleton to form a scene object layout; next, the view hierarchy data is superimposed onto the scene object layout in the order corresponding to the timestamp label and the time index to form a scene temporal layout; then, the state presentation data is bound to the corresponding scene object data in the scene temporal layout to form the 3D spatial scene data. The 3D spatial scene data is not a simple 3D graphics file, but a structured scene data including a spatial scene skeleton, scene object layout, scene temporal layout, and state presentation data. The hardware console can present the panoramic restored view based on the 3D spatial scene data.

[0070] Preferably, the specific implementation process for determining the interactive anchor points in the 3D spatial scene data is as follows: First, read the state presentation data from the 3D spatial scene data, and determine the scene object data in the state presentation data that corresponds to the deviation scalar, permission state field, call state field, call result field, mutation feature parameter, and environmental interaction candidate node; then, according to the interactive anchor point generation caliber and the interactive control type, mark the scene object data that can point back to the task-level field area, stage-level field area, interface-level field area, time index field area, source index, or time index in the on-site stack snapshot as the interactive anchor points. The technical essence of the interactive anchor point is the data attachment position in the 3D spatial scene data that can trigger viewing, positioning, backtracking, or intervention input. It has both a spatial position in the 3D spatial scene data and a source index and time index that point back to the on-site stack snapshot.

[0071] Preferably, when determining the interactive anchor point, source verification and time verification are also performed on the interactive anchor point based on the source state pruning fragment and time state pruning fragment in the simplified state feature string. The source verification is used to determine whether the interactive anchor point can be referenced back to the global variable parameter, local variable parameter, or external interface call record in the scene stack snapshot through the source index in the source state pruning fragment; the time verification is used to determine whether the interactive anchor point can be referenced back to the corresponding time position in the current execution cycle through the time index in the time state pruning fragment. If a scene object data fails the source verification or fails the time verification, then the scene object data is not marked as the interactive anchor point; if a scene object data can pass both the source verification and the time verification, then the scene object data is marked as the interactive anchor point, and the interactive anchor point is written into the interactive anchor point record. The interactive anchor point record is used to carry the source correspondence and time correspondence between the interactive anchor point and the scene stack snapshot.

[0072] Preferably, in the scenario corresponding to the security patrol task, if a scene object data in the three-dimensional spatial scene data corresponds to a mutation feature parameter that prohibits climbing, and this scene object data is located in the spatial scene skeleton corresponding to the camera drawing boundary data, and can be referenced back to the environmental interaction candidate node in the stage-level field area through the source index in the source state pruning fragment, and can be referenced back to the sampling time label before the permission suspension trigger time position through the time index in the time state pruning fragment, then this scene object data is marked as the interactive anchor point. If another scene object data corresponds to the external interface call record of the smart broadcast feedback interface, and this scene object data can be referenced back to the external interface call record in the interface-level field area through the source index in the source state pruning fragment, and can be referenced back to the call time field corresponding to the linkage stage through the time index in the time state pruning fragment, then this scene object data is marked as the interactive anchor point. Thus, the interactive anchor point can associate the scene object data in the panoramic restoration view with the state data in the on-site stack snapshot.

[0073] Preferably, the specific implementation process for generating the human-computer interaction probe based on the simplified state feature string, the three-dimensional spatial scene data, and the interactive anchor points is as follows: First, a probe carrying record is established. This record carries the intermediate data required for the formation of the human-computer interaction probe; the probe carrying record is not equivalent to the human-computer interaction probe itself. Then, the simplified state feature string is written into the state field area of ​​the probe carrying record, the three-dimensional spatial scene data is written into the scene field area of ​​the probe carrying record, and the interactive anchor point record is written into the interaction field area of ​​the probe carrying record. Next, a probe source index is established based on the source correspondence between the state field area, the scene field area, and the interaction field area. A probe time index is established based on the time correspondence between the time state pruned fragment and the three-dimensional spatial scene data. Based on the probe carrying record with the probe source index and the probe time index, probe encapsulation processing is performed to form the human-computer interaction probe.

[0074] Preferably, the probe encapsulation process includes state binding, scene binding, and interaction binding. State binding binds the task state pruning fragments, stage state pruning fragments, interface state pruning fragments, time state pruning fragments, and source state pruning fragments in the simplified state feature string to the corresponding scene object data in the 3D spatial scene data. Scene binding binds the spatial scene skeleton, scene object layout, scene temporal layout, and state presentation data in the 3D spatial scene data to the scene field area of ​​the probe carrying record. Interaction binding establishes a correspondence between the interactive anchor points in the interactive anchor point record and the probe source index and the probe time index. Through state binding, scene binding, and interaction binding, the state content, scene content, and interaction content in the probe carrying record are encapsulated into the human-computer interaction probe.

[0075] Preferably, the technical essence of the human-computer interaction probe is presentable, traceable, and interactive state probe data converted from the on-site stack snapshot. The human-computer interaction probe includes the simplified state feature string, the three-dimensional spatial scene data, the interactive anchor point, the probe source index, and the probe time index; wherein, the simplified state feature string carries the task state content retained after data pruning in the on-site stack snapshot; the three-dimensional spatial scene data is used to present a panoramic view on the hardware console; the interactive anchor point is used to locate the data position in the panoramic view that points back to the on-site stack snapshot; the probe source index points back to the data source of the human-computer interaction probe; and the probe time index points back to the time position of the human-computer interaction probe in the current execution cycle. Through the above structure, the human-computer interaction probe can convert the task state content in the on-site stack snapshot into a data carrier that can be presented and interacted with by the hardware console.

[0076] Preferably, in the example corresponding to the security patrol task, if the on-site stack snapshot includes the mutation feature parameter corresponding to the prohibition of climbing, the environmental interaction candidate node, the deviation scalar, the camera drawing boundary data, the external interface call record corresponding to the smart broadcast feedback interface, the external interface call record corresponding to the work order flow interface, and the permission status field, then the human-computer interaction probe pruning rule first filters the above field content from the on-site stack snapshot to form the simplified state feature string; then, the visual rendering pipeline converts the simplified state feature string into three-dimensional spatial scene data containing the camera drawing boundary data, the scene object data corresponding to the prohibition of climbing, the scene object data corresponding to the smart broadcast feedback interface, and the scene object data corresponding to the work order flow interface according to the preset visual rendering pipeline configuration; next, according to the source index in the source state pruning fragment and the time index in the time state pruning fragment, interactive anchor points are determined in the above scene object data; finally, the simplified state feature string, the three-dimensional spatial scene data, and the interactive anchor points are encapsulated into the human-computer interaction probe through the probe carrying record. This human-computer interaction probe enables the hardware console to present the panoramic restoration view based on the spatial state, temporal state, interface state, and permission state of the same target task.

[0077] Optionally, the step of performing a command translation operation on the intervention map to obtain a state correction chain includes: Extract the human control action sequences from the intervention map; Perform an action semantic decomposition operation on the human-controlled action sequence to output an atomic-level operation instruction sequence through the action semantic decomposition operation; Obtain a preset intelligent agent hardware execution constraint table, which includes hardware execution order constraints, hardware execution port constraints, and hardware execution state constraints corresponding to the intelligent agent; The atomic-level operation instruction sequence is sorted and reorganized according to the intelligent agent hardware execution constraint table; The state correction chain is generated based on the sorted and reorganized atomic-level operation instruction sequence.

[0078] Preferably, the specific implementation process for extracting the human control action sequence from the intervention map is as follows: After receiving the intervention map from the hardware console for the panoramic restoration view, the intervention nodes, intervention-related edges, interactive anchor point identifiers, probe source indexes, and probe time indexes in the intervention map are first read. The intervention nodes originate from the human intervention input executed by the hardware console on the interactive anchor point in the panoramic restoration view. The intervention-related edges express the temporal and object connections between two adjacent intervention nodes. The interactive anchor point identifier points back to the interactive anchor point in the human-computer interaction probe. The probe source index points back to the task-level field area, stage-level field area, or interface-level field area in the on-site stack snapshot. The probe time index points back to the time position in the current execution cycle. Subsequently, based on the probe source index and the probe time index, the intervention nodes are verified for source and time to confirm that the intervention nodes correspond to the state data in the on-site stack snapshot. Then, the human control action sequence is extracted according to the temporal connection of the intervention-related edges. The technical essence of the human-controlled action sequence is an intervention action data sequence issued by the hardware console, with interactive anchor points and the current execution cycle time position, rather than ordinary control commands detached from the scene stack snapshot.

[0079] Preferably, the human control action sequence includes at least one human control action record. Each human control action record includes an action source field, an action target field, an action type field, an action parameter field, an action time field, and an anchor point reference field. The action source field indicates which interactive anchor point in the panoramic view the human control action record originates from. The action target field indicates which type of status data the human control action record corresponds to: permission status field, mutation feature parameter, environmental interaction candidate node, external interface call record, smart broadcast feedback status, smart lighting feedback status, smart irrigation feedback status, or work order flow status. The action type field indicates which type of action the human control action record belongs to: status confirmation, status withdrawal, interface redirection, linkage resending, task phase rollback, or task phase continuation. The action parameter field carries the variable correction content, interface input parameter correction content, task phase correction content, or permission status correction content corresponding to the human control action record. The action time field carries the time position of the human control action record in the current execution cycle. The anchor point reference field points the human control action record back to the corresponding interactive anchor point. Through the above fields, the human-controlled action sequence can retain its technical origin relationship with the intervention map, the human-computer interaction probe, and the on-site stack snapshot.

[0080] Preferably, the specific implementation process of performing action semantic decomposition on the human-controlled action sequence is as follows: First, each human-controlled action record in the human-controlled action sequence is read, and the state data type of the human-controlled action record is determined according to the action target field; then, the human-controlled action record is determined according to the action type field to determine whether it needs to cause variable correction, interface call correction, task stage correction, or permission status correction; then, the target variable location information, variable correction value, target interface object, interface input parameter correction content, interface return expected content, task stage correction content, and permission status correction content are decomposed according to the action parameter field. The above decomposed content continues to establish a correspondence with the action time field and the anchor point reference field to form an action decomposition record. The action decomposition record is used to carry the intermediate data obtained from the action semantic decomposition operation, and subsequent atomic encapsulation based on the action decomposition record is required to form the atomic-level operation instruction sequence.

[0081] Preferably, when processing the action parameter field, the action semantic decomposition operation does not directly convert a human control action record into an execution instruction, but rather decomposes it according to the executable granularity of the state data. If the action target field points to the permission status field, the permission status correction content is extracted from the action parameter field and written into the permission status correction item in the action decomposition record; if the action target field points to an external interface call record, the target interface object, interface input parameter correction content, and expected interface return content are extracted from the action parameter field and written into the interface call correction item in the action decomposition record; if the action target field points to a mutation feature parameter, an environmental interaction candidate node, or a node deviation description item, the target variable location information and variable correction value are extracted from the action parameter field and written into the variable correction item in the action decomposition record; if the action target field points to a task stage field, the task stage correction content is extracted from the action parameter field and written into the task stage correction item in the action decomposition record. Through the above-described decomposition method, the action semantic decomposition operation can convert intervention actions for panoramic view restoration into atomic correction content that can subsequently participate in hardware constraint matching.

[0082] Preferably, the specific implementation process of the step of outputting an atomic-level operation instruction sequence through the action semantic decomposition operation is as follows: After the action decomposition record is formed, the permission status correction item, interface call correction item, variable correction item, and task stage correction item in the action decomposition record are read first, and then atomically encapsulated according to the granularity of changing only one target state object at a time to form multiple atomic-level operation instructions; each atomic-level operation instruction includes an atomic instruction type field, target variable location information, variable correction value, target interface object, interface input parameter correction content, interface return expected content, task stage correction content, permission status correction content, action time field, and anchor point reference field. Subsequently, the multiple atomic-level operation instructions are arranged into the atomic-level operation instruction sequence according to the action time field. The technical essence of the atomic-level operation instruction sequence is a fine-grained instruction sequence formed by decomposing the human control action sequence, which can accept hardware execution order constraints, hardware execution port constraints, and hardware execution state constraint verification item by item, rather than a high-level intervention text directly issued to the intelligent agent for execution.

[0083] Preferably, in the scenario corresponding to the security patrol task, if the human control action record is used to mark the environmental interaction candidate node corresponding to the prohibition of climbing as requiring linkage broadcast, and to reissue the external interface call record corresponding to the smart broadcast feedback interface, then the action semantic decomposition operation first extracts the target variable positioning information and variable correction value corresponding to the environmental interaction candidate node from the human control action record to form an atomic-level operation instruction pointing to the stage-level field area; then it extracts the target interface object, interface input parameter correction content, and interface return expected content corresponding to the smart broadcast feedback interface to form an atomic-level operation instruction pointing to the interface-level field area; subsequently, it reads the action time field and anchor point retracement field corresponding to the human control action record, and arranges the above two atomic-level operation instructions into the atomic-level operation instruction sequence according to the action time field. Thus, the atomic-level operation instruction sequence can decompose an intervention action for panoramic restoration view into two executable instructions with different granularities: variable correction and interface call correction.

[0084] Preferably, the specific implementation process of obtaining the preset intelligent agent hardware execution constraint table is as follows: Before the target task enters the execution state, the target task configuration record, external interface object configuration record, main working memory area status configuration record, driver instruction issuing port status configuration record, and task scheduling center execution stage configuration record are read first; wherein, the target task configuration record is used to provide task type field, task path field, device call field, behavior recognition grouping field, and task stage field; the external interface object configuration record is used to provide the interface call parameters of camera frame extraction interface, environmental perception device reporting interface, smart broadcast feedback interface, smart lighting feedback interface, smart irrigation feedback interface, work order flow interface, and geographic information management platform spatial location interface; the main working memory area status configuration record is used to provide the switching conditions between read-only state and read-write state; the driver instruction issuing port status configuration record is used to provide the port status conditions between port opening, port suspension, and port recovery; and the task scheduling center execution stage configuration record is used to provide the task stage progression relationship between inspection, identification, alarm, linkage, and handling feedback. Subsequently, the intelligent agent hardware execution constraint table is established based on the target task configuration record, the external interface object configuration record, the main working memory area status configuration record, the driver instruction issuance port status configuration record, and the task scheduling center execution phase configuration record.

[0085] Preferably, the technical essence of the agent hardware execution constraint table is a constraint data table used to limit whether the atomic-level operation instruction sequence can enter the subsequent state correction chain according to the agent's hardware execution conditions. The agent hardware execution constraint table includes hardware execution order constraints, hardware execution port constraints, and hardware execution state constraints. Specifically, the hardware execution order constraints limit the sequential relationship between different atomic-level operation instructions, for example, completing variable correction in the main working memory area first, and then resuming the interface call corresponding to the driver instruction issuance port. The hardware execution port constraints limit the external interface objects that a certain atomic-level operation instruction can call and the port state of the external interface objects in the current execution cycle. The hardware execution state constraints limit which state the agent should be in before executing a certain atomic-level operation instruction: permission suspended, read-only, read-write, or port suspended. Through the agent hardware execution constraint table, the atomic-level operation instruction sequence can be arranged into an order that meets the hardware execution conditions before entering the state correction chain.

[0086] Preferably, the intelligent agent hardware execution constraint table, when preset, also establishes constraint source fields for the hardware execution order constraint, the hardware execution port constraint, and the hardware execution state constraint, respectively. The constraint source field of the hardware execution order constraint points back to the execution phase configuration record of the task scheduling center; the constraint source field of the hardware execution port constraint points back to the configuration record of the external interface object; and the constraint source field of the hardware execution state constraint points back to the status configuration record of the main working memory area and the status configuration record of the driver instruction issuing port. The constraint source field is used in the subsequent sorting and reorganization process to explain why a certain atomic operation instruction needs to be advanced, delayed, split, or kept in its original position, so that the sorting and reorganization process of the atomic operation instruction sequence can point back to the specific constraint source, rather than relying on ordinary instruction arrangement experience.

[0087] Preferably, the specific implementation process of sorting and reorganizing the atomic-level operation instruction sequence according to the intelligent agent hardware execution constraint table is as follows: First, each atomic-level operation instruction in the atomic-level operation instruction sequence is read, and it is divided into variable correction instructions, interface call correction instructions, task stage correction instructions, or permission status correction instructions according to the atomic instruction type field; then, the variable correction instructions are matched with the hardware execution status constraints to determine whether the variable correction instructions need to be executed after the main working memory area is in a read / write state; the interface call correction instructions are matched with the hardware execution port constraints to determine whether the target interface object corresponding to the interface call correction instructions is in a port callable state; the task stage correction instructions are matched with the hardware execution order constraints to determine whether the task stage correction instructions conform to the task stage progression relationship between inspection, identification, alarm, linkage, and handling feedback; the permission status correction instructions are matched with the hardware execution status constraints to determine whether the permission status correction instructions can enter the subsequent correction process in a permission suspended state. After the above matching is completed, a constraint matching record is formed. The constraint matching record is used to carry the matching result between each atomic-level operation instruction and the hardware execution order constraint, the hardware execution port constraint, or the hardware execution state constraint.

[0088] Preferably, after the constraint matching record is formed, the atomic-level operation instruction sequence is further sorted and reorganized according to the hardware execution order constraint, the hardware execution port constraint, and the hardware execution state constraint. Specifically, atomic-level operation instructions that do not meet the hardware execution state constraint are first marked as pending state transition instructions. These pending state transition instructions indicate that atomic-level operation instructions related to read / write state switching need to be inserted or executed earlier than the current atomic-level operation instruction. Then, atomic-level operation instructions that do not meet the hardware execution port constraint are marked as pending port recovery instructions. These pending port recovery instructions indicate that atomic-level operation instructions related to port recovery need to be inserted or executed earlier than the current atomic-level operation instruction. Subsequently, according to the hardware execution order constraint, atomic-level operation instructions used to release read-only state or restore port state are arranged before the corresponding variable correction instructions and interface call correction instructions. Then, according to the action time field and task stage field, multiple atomic-level operation instructions under the same task stage field are arranged in chronological order. The above sorting and reorganization process forms a sorting and reorganization record, which carries the original order, reorganized order, constraint matching result, and constraint source field of each atomic-level operation instruction.

[0089] Preferably, during the sorting and reorganization process, if multiple atomic-level operation instructions extracted from a human-controlled action record point to a stage-level field area and an interface-level field area respectively, then the variable correction instruction corresponding to the stage-level field area should be determined before the interface call correction instruction corresponding to the interface-level field area according to the hardware execution order constraints. If the interface call correction instruction corresponding to the intelligent broadcast feedback interface needs to be executed after the driver instruction is issued and the port is restored, then the atomic-level operation instructions corresponding to the port restoration are arranged before the interface call correction instruction corresponding to the intelligent broadcast feedback interface according to the hardware execution port constraints. If a variable correction instruction needs to be written to the main working memory area, then the atomic-level operation instructions corresponding to the read / write state switching are arranged before the variable correction instruction according to the hardware execution state constraints. Through the above sorting and reorganization, the atomic-level operation instruction sequence can be transformed from the input order of the human-controlled action sequence into a reorganized atomic-level operation instruction sequence acceptable to the intelligent agent's hardware execution conditions.

[0090] Preferably, the specific implementation process of generating the state correction chain based on the sorted and recombined atomic-level operation instruction sequence is as follows: First, read the recombined atomic-level operation instruction sequence, constraint matching result, and constraint source fields from the sorted and recombined record, and establish a state correction chain carrier record; then, write each atomic-level operation instruction into the chain node field of the state correction chain carrier record in the recombined order, and write the hardware execution order constraint, hardware execution port constraint, or hardware execution state constraint between two adjacent atomic-level operation instructions into the chain constraint field of the state correction chain carrier record; then, write the target variable positioning information, variable correction value, target interface object, interface input parameter correction content, interface return expected content, task stage correction content, permission state correction content, action time field, and anchor point retracement field corresponding to each atomic-level operation instruction into the chain correction field of the state correction chain carrier record. Based on the state correction chain carrier record with the chain node field, chain constraint field, and chain correction field configured, perform chain encapsulation processing to form the state correction chain.

[0091] Preferably, the chain encapsulation process specifically includes node concatenation, constraint binding, and source backpointing. Node concatenation connects multiple atomic-level operation instructions in the chain node field according to the order of the recombined atomic-level operation instruction sequence, forming a chained instruction structure that can be read by subsequent reconstructed on-site stack snapshots. Constraint binding binds the hardware execution order constraints, hardware execution port constraints, and hardware execution state constraints in the chain constraint field to two adjacent atomic-level operation instructions, enabling the identification of the execution preconditions of each atomic-level operation instruction during subsequent reconstruction of the on-site stack snapshot. Source backpointing writes the anchor point backpointing field, the probe source index, and the probe time index into the state correction chain carrying record, enabling the state correction chain to backpoint to the corresponding interactive anchor point, human-machine interaction probe, on-site stack snapshot, and current execution cycle. After completing node concatenation, constraint binding, and source backpointing, the state correction chain carrying record is encapsulated into the state correction chain.

[0092] Preferably, the technical essence of the state correction chain is to convert the sequence of human control actions in the intervention map into chain-like correction data that can be used to reconstruct the on-site stack snapshot. The state correction chain includes multiple chain node fields, multiple chain constraint fields, multiple chain correction fields, anchor point reference fields, probe source index, and probe time index; wherein, the chain node fields are used to carry the sorted and reorganized atomic-level operation instructions; the chain constraint fields are used to carry the hardware execution order constraints, hardware execution port constraints, and hardware execution state constraints corresponding to the atomic-level operation instructions; the chain correction fields are used to carry the target variable location information, variable correction values, target interface objects, interface input parameter correction content, interface return expected content, task stage correction content, and permission status correction content required for subsequent reconstruction of the on-site stack snapshot; the anchor point reference field is used to reference the interactive anchor point corresponding to the human control action record that formed the chain node field; and the probe source index and the probe time index are used to reference the data source and time position in the current execution cycle of the human-computer interaction probe. Using the aforementioned chained correction data, the target variable location information, variable correction value, and execution timing can be read from the state correction chain to continue the on-site stack snapshot reconstruction process.

[0093] Preferably, in the example corresponding to the security patrol task, if the intervention map indicates that supplementary linkage processing is needed for the environmental interaction candidate node corresponding to the prohibition of climbing, and that the smart broadcast feedback interface re-forms the external interface call record during the linkage phase, then the corresponding human control action sequence is first extracted from the intervention map; subsequently, the human control action sequence is subjected to action semantic decomposition to form variable correction instructions pointing to the environmental interaction candidate node, interface call correction instructions pointing to the smart broadcast feedback interface, and task phase correction instructions pointing to the task phase field; then, according to the hardware execution order constraints, hardware execution port constraints, and hardware execution state constraints in the intelligent agent hardware execution constraint table, the above atomic-level operation instructions are sorted and reorganized so that the task phase correction instructions are placed before the interface call correction instructions corresponding to the smart broadcast feedback interface, and the atomic-level operation instructions corresponding to port restoration are placed before the interface call correction instructions corresponding to the smart broadcast feedback interface; finally, the state correction chain is formed based on the sorted and reorganized atomic-level operation instruction sequence. This state correction chain can express the conversion relationship between the intervention action in the panoramic restoration view and the correction content required for subsequent on-site stack snapshot reconstruction.

[0094] Optionally, reconstructing the field stack snapshot based on the state correction chain includes: Create a mirror copy of the live stack snapshot in the memory control area of ​​the agent, and establish a snapshot source correspondence between the mirror copy and the live stack snapshot; Read the target variable location information, variable correction value, and execution sequence from the state correction chain; According to the execution sequence, the target variable parameters corresponding to the target variable location information are located one by one in the mirror copy, and the target variable parameters are overwritten using the variable correction value; Perform a closed-loop logical verification on the overwritten mirror copy; When the logical closed-loop verification passes, the original on-site stack snapshot is replaced with the overwritten mirror copy to form the reconstructed on-site stack snapshot.

[0095] Preferably, the specific implementation process of creating a mirror copy of the live stack snapshot in the agent's memory control area is as follows: After the state correction chain is formed, the state of the memory control area when the agent is in a permission-suspended state is read first. The memory control area is used to carry temporary data objects during the reconstruction of the live stack snapshot. The temporary data objects include a copy carrier record, a location exception record, an overwrite exception record, and a reconstruction snapshot carrier record, and are used to isolate the live stack snapshot from the correction content in the state correction chain, avoiding direct rewriting of the live stack snapshot before verification is completed. Subsequently, a snapshot reconstruction workspace is established in the memory control area, and a copy carrier record with the same field structure as the live stack snapshot is generated based on the task-level field area, stage-level field area, interface-level field area, timestamp field area, source index, and time index of the live stack snapshot. Then, the task-level field area, stage-level field area, interface-level field area, timestamp field area, source index, and time index in the live stack snapshot are copied to the copy carrier record to form the mirror copy. The technical essence of the mirror copy is that it is a correctable snapshot data with the same field structure, the same source index, and the same time index as the on-site stack snapshot. It is used to carry over the correction content in the state correction chain, rather than directly replacing the on-site stack snapshot.

[0096] Preferably, when creating the mirror copy, it retains the task-level field area, stage-level field area, interface-level field area, timestamp field area, source index, and time index from the live stack snapshot, and adds a copy field status to each field content; the copy field status includes uncorrected state, pending correction state, corrected state, and verification passed state. Specifically, the global variable parameters from the live stack snapshot are first copied to the task-level field area of ​​the mirror copy, then the local variable parameters from the live stack snapshot are copied to the stage-level field area of ​​the mirror copy, and the external interface call records from the live stack snapshot are copied to the interface-level field area of ​​the mirror copy; subsequently, the timestamp tag is copied to the timestamp field area of ​​the mirror copy, and the source index and the time index are copied to the mirror copy. Through the above copying method, the mirror copy can completely retain the task execution context, parameter state context, and external interface call context of the live stack snapshot when the permission suspension state occurs, providing a fallback reference data basis when subsequently overwriting the target variable parameters.

[0097] Preferably, the specific implementation process of establishing a snapshot source correspondence between the mirror copy and the on-site stack snapshot is as follows: After forming the mirror copy, firstly, the source index and time index in the on-site stack snapshot are read, and the source index and time index are written into the copy source field of the mirror copy. The copy source field is used to carry the source indexing content between the mirror copy and the on-site stack snapshot. Subsequently, a field source correspondence is established between the task-level field area in the on-site stack snapshot and the task-level field area in the mirror copy; a field source correspondence is established between the stage-level field area in the on-site stack snapshot and the stage-level field area in the mirror copy; a field source correspondence is established between the interface-level field area in the on-site stack snapshot and the interface-level field area in the mirror copy; and a field source correspondence is established between the time index field area in the on-site stack snapshot and the time index field area in the mirror copy. The above field source correspondences and the copy source field together form the snapshot source correspondence. The technical essence of the snapshot source correspondence is to express the mapping relationship between which field area and which time position in the on-site stack snapshot the content of each field in the mirror copy originates from. When performing overwrite, verification and replacement processes in the future, the original field source is pointed back through the snapshot source correspondence.

[0098] Preferably, the specific implementation process of reading the target variable location information, variable correction value, and execution sequence from the state correction chain is as follows: First, read the chain node field, chain constraint field, chain correction field, anchor point retracement field, probe source index, and probe time index from the state correction chain; then, extract the target variable location information, variable correction value, target interface object, interface input parameter correction content, interface return expected content, task stage correction content, and permission status correction content from the chain correction field; and then determine the execution sequence based on the hardware execution order constraint, hardware execution port constraint, and hardware execution status constraint in the chain constraint field. The technical essence of the target variable location information is the structured address information used to locate the position of the field to be corrected in the mirror copy, which includes field area location information, field name location information, source index location information, and time index location information; the technical essence of the variable correction value is the data content used to replace or supplement the target variable parameter, which can come from variable correction content, interface input parameter correction content, interface return expected content, task stage correction content, or permission status correction content; the technical essence of the execution sequence is the correction order relationship jointly defined by the reorganized order of the chain node fields in the status correction chain and the chain constraint fields.

[0099] Preferably, after the target variable location information is read, it is first matched with the task-level field area, stage-level field area, interface-level field area, and time index field area in the mirror copy. When the field area location information in the target variable location information points to the task-level field area, the target variable parameters corresponding to the permission status field, deviation metric scalar, preset trigger baseline, or task status source information are searched in the task-level field area of ​​the mirror copy. When the field area location information in the target variable location information points to the stage-level field area, the mutation feature parameters and environmental interaction candidate nodes are searched in the stage-level field area of ​​the mirror copy. The target variable parameters are defined as follows: environmental interaction node sequence, execution node, node distance parameter, or node deviation description item. When the target variable positioning information points to the interface-level field area, the external interface call record corresponding to the camera frame extraction interface, environmental sensing device reporting interface, smart broadcast feedback interface, smart lighting feedback interface, smart irrigation feedback interface, work order flow interface, or geographic information management platform spatial location interface is searched within the interface-level field area of ​​the mirror copy. The interface input parameter field, interface return field, call status field, or call result field in the external interface call record are used as the target variable parameters. Through the above field matching, the target variable positioning information can locate the target variable parameters corresponding to the state correction chain in the mirror copy.

[0100] Preferably, the specific implementation process of locating the target variable parameters corresponding to the target variable location information one by one in the mirror copy according to the execution sequence is as follows: First, read the first chain node field in the state correction chain according to the execution sequence, and read the target variable location information from the chain correction field corresponding to the chain node field; then, enter the corresponding field area in the mirror copy according to the field area location information in the target variable location information, and then search for the corresponding target variable parameter according to the field name location information, source index location information, and time index location information in the target variable location information; when the target variable parameter is found, change the replica field status of the target variable parameter from uncorrected to pending correction, and record the chain source correspondence between the target variable parameter and the chain node field; when a target variable location information cannot locate the corresponding target variable parameter, do not perform variable correction value overwrite processing, but write the target variable location information into the location exception record. The location exception record is used to carry the target variable location information that cannot locate the corresponding target variable parameter, and is used to determine whether the mirror copy has the field basis for continued reconstruction during subsequent logic closed-loop verification.

[0101] Preferably, the technical essence of the target variable parameter is field-level state data in the mirror copy that can be corrected by the state correction chain. This data can be permission status fields or task status source information in the task-level field area, mutation feature parameters, environmental interaction candidate nodes, or node deviation description items in the stage-level field area, or external interface call records in the interface-level field area. Taking a security patrol task as an example, if the target variable positioning information in the state correction chain points to the environmental interaction candidate node corresponding to the "prohibited climbing" rule, then the target variable parameter can be the node status field corresponding to that environmental interaction candidate node in the stage-level field area of ​​the mirror copy; if the target variable positioning information points to the external interface call record corresponding to the smart broadcast feedback interface, then the target variable parameter can be the call status field, interface input parameter field, or call result field corresponding to that smart broadcast feedback interface in the interface-level field area of ​​the mirror copy. Through the above positioning method, the target variable parameter can form a one-to-one correspondence with the variable correction values ​​in the state correction chain.

[0102] Preferably, the specific implementation process of overwriting the target variable parameter using the variable correction value is as follows: After the target variable parameter is located and enters the state to be corrected, the variable correction value corresponding to the target variable parameter is read first, and the overwriting method is determined according to the field area to which the target variable parameter belongs; when the target variable parameter belongs to the task-level field area, the variable correction value is written to the corresponding permission status field, task status source information, or deviation metric scalar association field, and the status of the copy field of the field content is changed to the corrected state; when the target variable parameter belongs to the stage-level field area, the variable correction value is written to the corresponding mutation feature parameter, environment interaction candidate node, node deviation description item, or task stage field, and the status of the copy field of the field content is changed to the corrected state; when the target variable parameter belongs to the interface-level field area, the variable correction value is written to the corresponding interface input parameter field, interface return field, call status field, or call result field, and the status of the copy field of the field content is changed to the corrected state. After the variable correction value overwrites the target variable parameter, the overwrite source is also retained according to the chain source correspondence, so that each overwrite can point back to the chain node field in the state correction chain.

[0103] Preferably, when the variable correction value overwrites the target variable parameter, data with different units or different field definitions are not directly replaced. Specifically, the field type of the target variable parameter is first read based on its field name location information. Then, based on the source content of the variable correction value, it is determined whether the variable correction value belongs to variable correction content, interface input parameter correction content, interface return expected content, task phase correction content, or permission status correction content. Subsequently, the field type and the source content of the variable correction value are validated using field definitions. When the field definition validation passes, the variable correction value is used to overwrite the target variable parameter. When the field definition validation fails, the variable correction value and the target variable parameter are written to an overwrite exception record. The overwrite exception record is used to carry the variable correction value and target variable parameter that failed the field definition validation, and continues to enter the logical closed-loop validation, enabling the logical closed-loop validation to determine whether the mirror copy contains correction content with inconsistent field definitions.

[0104] Preferably, in the scenario corresponding to the security patrol task, if the state correction chain requires supplementary linkage processing for the environmental interaction candidate node corresponding to the prohibited climbing, then firstly, based on the target variable location information, the target variable parameter corresponding to the environmental interaction candidate node is located in the stage-level field area of ​​the mirror copy, and then the variable correction value indicating that the linkage stage needs to be entered is overwritten to the target variable parameter; subsequently, based on the target variable location information in the next chain node field, the external interface call record corresponding to the smart broadcast feedback interface is located in the interface-level field area of ​​the mirror copy, and the interface input parameter correction content is overwritten to the interface input parameter field of the external interface call record, and the expected interface return content is overwritten to the interface return field or call result field of the external interface call record. The above overwriting process enables the mirror copy to express the continuous state change from environmental interaction candidate node correction to smart broadcast feedback interface call correction, rather than just replacing a single field in isolation.

[0105] Preferably, the specific implementation process of performing logical closed-loop verification on the overwritten mirror copy is as follows: After the variable correction value overwrites the target variable parameter, the task-level field area, stage-level field area, interface-level field area, timestamp field area, source index, time index, and copy field status in the overwritten mirror copy are read first, and the location exception record and the overwrite exception record are read; then, field integrity verification, source consistency verification, time sequence verification, task stage progression verification, and interface call status verification are performed on the overwritten mirror copy. The field integrity check is used to determine whether the target variable parameters corresponding to each chain node field in the state correction chain have been located and entered the corrected state; the source consistency check is used to determine whether the target variable parameters whose replica field status is in the corrected state can be pointed back to the state correction chain through the chain source correspondence, and can be pointed back to the field stack snapshot through the snapshot source correspondence; the time sequence check is used to determine whether the time index of each target variable parameter conforms to the execution sequence; the task stage progression check is used to determine whether the corrected task stage fields conform to the task stage progression relationship between inspection, identification, alarm, linkage and handling feedback; the interface call status check is used to determine whether the external interface call records corresponding to the smart broadcast feedback interface, smart lighting feedback interface, smart irrigation feedback interface or work order flow interface have a correspondence between the interface input parameter field, call status field and call result field.

[0106] Preferably, the logical closed-loop verification does not involve abstract judgment on the overwritten mirror copy, but rather reads the target variable parameters whose copy field status is in the corrected state item by item along the execution sequence of the state correction chain, and determines whether the correction result of the previous target variable parameter can serve as the source condition for the correction of the next target variable parameter. Specifically, if the target variable parameter corresponding to the previous chain node field is a task stage field, it determines whether the external interface call record in the interface-level field area corresponding to the next chain node field belongs to the external interface object allowed to be called by the task stage field; if the target variable parameter corresponding to the previous chain node field is an environment interaction candidate node, it determines whether the smart broadcast feedback interface, smart lighting feedback interface, smart irrigation feedback interface, or work order transfer interface corresponding to the next chain node field has a stage source correspondence with the task stage field to which the environment interaction candidate node belongs; if the target variable parameter corresponding to the previous chain node field is a permission status field, it determines whether the main working memory area write status or the driver instruction issuance port recovery status corresponding to the next chain node field meets the hardware execution status constraints. Through the above item-by-item verification method, the logical closed-loop verification can determine whether the overwritten mirror copy has a continuous state structure that can be used to release running permissions in the future.

[0107] Preferably, when the logical closed-loop verification finds that the location anomaly record is empty, the overwrite anomaly record is empty, the field integrity verification passes, the source consistency verification passes, the time sequence verification passes, the task phase progression verification passes, and the interface call status verification passes, the content of the fields in the overwritten mirror copy whose status is in the corrected state is changed to the verification passed state, and the verification passed state is written into the mirror copy's field status. If the location anomaly record or the overwrite anomaly record is not empty, or if any of the field integrity verification, source consistency verification, time sequence verification, task phase progression verification, and interface call status verification fails, then the overwritten mirror copy is not used to replace the on-site stack snapshot, and the on-site stack snapshot is retained as the original state carrier in the permission suspended state. Through the above processing, the on-site stack snapshot will not be overwritten when the mirror copy fails verification.

[0108] Preferably, the specific implementation process of replacing the original on-site stack snapshot with the overwritten mirror copy to form the reconstructed on-site stack snapshot when the logical closed-loop verification passes is as follows: After all the copy field states in the overwritten mirror copy reach the verification passed state, the task-level field area, stage-level field area, interface-level field area, timestamp field area, source index, time index, snapshot source correspondence, and chain source correspondence in the overwritten mirror copy are read first; then, the overwritten mirror copy is encapsulated into a reconstruction snapshot carrier record, which is used to carry the reconstruction data object before replacing the on-site stack snapshot; then, a reconstruction source index and a reconstruction time index are established based on the reconstruction snapshot carrier record. The reconstruction source index is used to refer back to the on-site stack snapshot and the state correction chain, and the reconstruction time index is used to refer back to the time position in the current execution cycle. Based on the reconstruction snapshot carrier record with the completed reconstruction source index and reconstruction time index, a replacement process is performed to form the reconstructed on-site stack snapshot.

[0109] Preferably, the replacement process is not simply deleting the original context stack snapshot, but rather marking the context stack snapshot as the original snapshot version in the memory control area, and marking the reconstructed snapshot carrying record as a continueable snapshot version. Subsequently, the continueable snapshot version is used as the snapshot data loaded into the main working memory area when the agent subsequently releases its running permissions, while the original snapshot version is retained as rollback reference data. The technical essence of the reconstructed context stack snapshot is a task state snapshot corrected by the state correction chain and verified through logical closed-loop checks. It retains the task running context of the context stack snapshot before the permission suspension state occurred, and adds the variable correction content, interface input parameter correction content, interface return expected content, task stage correction content, and permission state correction content corresponding to the state correction chain. Through the reconstructed context stack snapshot, when the agent's running permissions are subsequently released, the target task can continue to advance according to the reconstructed context stack snapshot.

[0110] Preferably, in the security patrol task of a smart garden, when the state correction chain requires the addition of the environmental interaction candidate node corresponding to the prohibited climbing phase to the linkage stage, and requires the smart broadcast feedback interface to form an external interface call record corresponding to the environmental interaction candidate node, a mirror copy of the on-site stack snapshot is first created in the memory control area, and the field source between the mirror copy and the on-site stack snapshot is retained according to the snapshot source correspondence. Subsequently, the target variable positioning information pointing to the environmental interaction candidate node, the variable correction value used to supplement the linkage stage, and the corresponding execution sequence are read from the state correction chain, and the target variable parameter corresponding to the environmental interaction candidate node is located in the stage-level field area of ​​the mirror copy; then, the target variable positioning information pointing to the smart broadcast feedback interface, the interface input parameter correction content, and the interface return expected content are read from the state correction chain, and the external interface call record corresponding to the smart broadcast feedback interface is located in the interface-level field area of ​​the mirror copy. After completing the above positioning, the corresponding target variable parameters are overwritten sequentially according to the execution sequence, and a logical closed-loop verification is performed on the overwritten mirror copy. When the logical closed-loop verification passes, the overwritten mirror copy is used to form the reconstructed on-site stack snapshot. This reconstructed on-site stack snapshot can express the continuous correction relationship between the environmental interaction candidate nodes corresponding to the prohibited climbing, the linkage stage, the smart broadcast feedback interface, and the external interface call records.

[0111] Optionally, before sending the human-computer interaction probe to the hardware console, the method further includes: Obtain a preset console authentication configuration record, and send a handshake request data packet to the hardware console according to the console authentication configuration record; Receive the biometric identifier returned by the hardware console based on the handshake request data packet; The biometric identifier is compared with the locally stored authorized administrator feature database to obtain the console identity comparison result; When the console identity verification result matches, an encrypted transmission tunnel is established, and the human-computer interaction probe is sent to the hardware console through the encrypted transmission tunnel.

[0112] Preferably, the specific implementation process of obtaining the preset console authentication configuration record is as follows: Before the target task enters the execution state, the target task configuration record, hardware console presentation configuration record, authorized administrator configuration record, and console device configuration record corresponding to the target task are read first. The target task configuration record provides fields for task type, task path, device invocation, behavior recognition grouping, camera drawing boundary data, spatial location in the geographic information management platform, and task stage. The hardware console presentation configuration record provides the panoramic restoration view hierarchy, interactive anchor point type, and probe receiving port that the hardware console can present. The authorized administrator configuration record provides the authorized administrator identifier, authorized task scope, and authorized validity status that can receive the human-computer interaction probe. The console device configuration record provides the device identifier, console communication address, probe receiving port, and console certificate digest of the hardware console. Subsequently, the console authentication configuration record is configured based on the target task configuration record, the hardware console presentation configuration record, the authorized administrator configuration record, and the console device configuration record. The technical essence of the console authentication configuration record is an authentication constraint record used to limit the identity verification of the human-computer interaction probe before it is sent, which hardware console, which authorized administrator ID, and which probe receiving port it needs to pass through. It is not a regular login information record.

[0113] Preferably, the console authentication configuration record includes a task authentication field, a console authentication field, an administrator authentication field, a probe transmission field, and an authentication timing field. The task authentication field carries the task identifier, current execution cycle identifier, task stage field, and permission suspension trigger time position of the target task. The console authentication field carries the device identifier, console communication address, probe receiving port, and console certificate digest of the hardware console. The administrator authentication field carries the authorized administrator identifier, authorized task scope, authorized validity status, and authorized administrator feature database index. The probe transmission field carries the probe digest, probe source index, probe time index, and probe data segment identifier of the human-computer interaction probe. The authentication timing field carries the handshake request sending time, handshake response receiving time, identity comparison time, and encrypted transmission tunnel establishment time. Through the above field configuration, the console authentication configuration record can fix the authentication relationship between the target task, hardware console, authorized administrator, and human-computer interaction probe into the same data structure, enabling subsequent handshake request data packets to point back to the on-site stack snapshot and the source of the human-computer interaction probe.

[0114] Preferably, the authorized administrator feature library is established before the console authentication configuration record is preset. Specifically, the authorized administrator identifier, authorized task scope, and authorized validity status are first read from the authorized administrator configuration record. Then, biometric samples corresponding to the authorized administrator are collected through the hardware console, and feature fragment extraction, noise field removal, and irreversible summary transformation are performed on the biometric samples to form an authorized administrator feature summary. Subsequently, the authorized administrator feature summary is bound to the authorized administrator identifier, the authorized task scope, the authorized validity status, and the console device configuration record, and the bound data is written to the locally stored authorized administrator feature library. The technical essence of the authorized administrator feature library is a data set used to store authorized administrator feature summaries and their authorized task scopes. It does not directly store reproducible biometric samples, but rather stores summarized feature data that can be used to compare the biometric identifier. Through the authorized administrator feature library, subsequent console identity comparison results can simultaneously reflect the biometric source, authorized task scope, and hardware console source.

[0115] Preferably, when forming the authorized administrator feature summary, the biometric sample is first split into fields according to the acquisition channel type to form an acquisition channel field, an acquisition time field, a feature quality field, and a feature content field. The acquisition channel field indicates which acquisition channel in the hardware console the biometric sample originates from; the acquisition time field indicates the time and position of the biometric sample entering the hardware console; the feature quality field indicates whether the biometric sample meets the acquisition conditions for subsequent irreversible summary conversion; and the feature content field carries the biometric fragments extracted from the feature fragments. Subsequently, feature content fields that are incompletely acquired or repeatedly sampled are removed based on the feature quality field. Then, irreversible summary conversion is performed on the remaining feature content fields to form the authorized administrator feature summary. The authorized administrator feature summary continues to establish an administrator source correspondence with the authorized administrator identifier and a console source correspondence with the console device configuration record. The administrator source correspondence is used to confirm the authorized administrator identifier corresponding to the biometric identifier code during subsequent comparisons, and the console source correspondence is used to confirm whether the hardware console belongs to the hardware console defined by the console authentication configuration record during the subsequent establishment of the encrypted transmission tunnel.

[0116] Preferably, the specific implementation process of sending a handshake request data packet to the hardware console according to the console authentication configuration record is as follows: After the human-computer interaction probe is generated, the task authentication field, console authentication field, administrator authentication field, probe transmission field, and authentication timing field in the console authentication configuration record are read first, and the console communication address and probe receiving port of the hardware console are determined according to the console authentication field; then, the task identifier, current execution cycle identifier, and permission suspension trigger time position of the target task are written according to the task authentication field, the probe summary, probe source index, and probe time index of the human-computer interaction probe are written according to the probe transmission field, and the authorized administrator identifier and authorized task scope are written according to the administrator authentication field, and the above fields are then encapsulated into the handshake request data packet. The technical essence of the handshake request data packet is a data packet used to trigger hardware console identity authentication before sending the human-computer interaction probe. It carries the target task source, human-computer interaction probe source, authorized task scope, console communication address, and probe receiving port, enabling the hardware console to return the biometric identification code based on the same authentication standard.

[0117] Preferably, the handshake request data packet further includes a request integrity digest and a request expiration field before being sent. Specifically, a request integrity digest is first generated based on the task authentication field, the console authentication field, the administrator authentication field, and the probe transmission field, and then written into the handshake request data packet. Subsequently, a request expiration field is generated based on the permission suspension trigger time and the handshake request sending time, and then written into the handshake request data packet. The request integrity digest is used to determine whether the field content has been modified after the hardware console receives the handshake request data packet, and the request expiration field is used to determine whether the handshake request data packet still belongs to the current execution cycle after the hardware console receives the handshake request data packet. Through the request integrity digest and the request expiration field, the handshake request data packet can establish a correspondence with the permission suspension state within the current execution cycle, preventing the human-computer interaction probe from being sent to a hardware console that does not correspond to the target task.

[0118] Preferably, the specific implementation process of receiving the biometric identifier returned by the hardware console based on the handshake request data packet is as follows: After receiving the handshake request data packet, the hardware console first reads the authorized administrator identifier, authorized task scope, task identifier, current execution cycle identifier, probe digest, and request validity field from the handshake request data packet, and calls the local biometric acquisition component of the hardware console according to the authorized administrator identifier; subsequently, the hardware console collects the biometric response sample of the currently authenticated object through the biometric acquisition component, and performs feature fragment extraction, noise field removal, and irreversible digest conversion on the biometric response sample to form the biometric identifier; then, the biometric identifier is bound to the task identifier, the current execution cycle identifier, the probe digest, and the device identifier of the hardware console, and the bound biometric identifier is returned. The technical essence of the biometric identifier is a temporary identity response digest formed by the hardware console for this handshake request data packet, which is used to express the identity source of the returned biometric identifier, which is simultaneously bound to the device identifier of the hardware console and the corresponding authorized administrator identifier.

[0119] Preferably, the biometric identifier is not a directly reproducible biometric image or raw biometric data, but rather a bound digest data generated based on the biometric response sample, the request integrity digest in the handshake request data packet, the request expiration field, and the device identifier of the hardware console. The biometric response sample provides the source of the biometrics of the currently authenticated object; the request integrity digest binds the biometric identifier to the current handshake request data packet; the request expiration field limits the biometric identifier to the current execution cycle; and the device identifier of the hardware console limits the biometric identifier to the corresponding hardware console. Through this binding method, even if different target tasks involve the same authorized administrator identifier, the biometric identifier will generate different authentication response results depending on the task identifier, the current execution cycle identifier, and the request expiration field.

[0120] Preferably, the specific implementation process of comparing the biometric identifier with the locally stored authorized administrator feature database to obtain the console identity comparison result is as follows: After receiving the biometric identifier, first read the authorized administrator feature database index, authorized administrator identifier, authorized task scope, console certificate digest, task identifier, and current execution cycle identifier from the console authentication configuration record, and read the corresponding authorized administrator feature digest from the locally stored authorized administrator feature database according to the authorized administrator feature database index; then, compare the biometric digest portion of the biometric identifier with the authorized administrator feature digest using the same caliber to form a biometric digest comparison result; then compare the device identifier of the hardware console bound to the biometric identifier with the console device configuration record in the console authentication configuration record to form a console source comparison result; next, compare the task identifier and current execution cycle identifier bound to the biometric identifier with the task authentication field to form a task source comparison result; finally, compare the authorized task scope in the authorized administrator feature database with the task type field, task path field, and task stage field corresponding to the target task to form an authorized task scope comparison result. Based on the biometric summary comparison results, the console source comparison results, the task source comparison results, and the authorized task scope comparison results, the console identity comparison results are generated.

[0121] Preferably, the technical essence of the console identity comparison result is to express whether the hardware console, the authorized administrator, and the current execution cycle simultaneously conform to the data judgment result of the console authentication configuration record, rather than a single biometric similarity result. The console identity comparison result includes a biometric summary comparison result, a console source comparison result, a task source comparison result, an authorized task range comparison result, and a comparison time field. The authorized task range comparison result is used to determine whether the authorized task range in the authorized administrator's feature library covers the task type field, task path field, and task stage field corresponding to the target task. The comparison time field is used to point back the current identity comparison process to the time position in the current execution cycle. The console identity comparison result is considered consistent only when the biometric summary comparison result, the console source comparison result, the task source comparison result, and the authorized task range comparison result are all consistent; if any comparison result is inconsistent, the console identity comparison result is considered inconsistent.

[0122] Preferably, when comparing the biometric identifier with the locally stored authorized administrator feature database, data from different sources are not directly mixed for comparison. Specifically, first, the biometric digest portion of the biometric identifier is compared with the authorized administrator feature digest to form the biometric digest comparison result; then, the device identifier of the hardware console and the console certificate digest are compared to form the console source comparison result; subsequently, the task identifier, the current execution cycle identifier, and the permission suspension trigger time position are compared to form the task source comparison result; finally, the authorized task scope and the task stage field are compared to form the authorized task scope comparison result. The above comparison process generates the biometric digest comparison result, the console source comparison result, the task source comparison result, and the authorized task scope comparison result, respectively. All comparison results, provided the field definitions are consistent, jointly generate the console identity comparison result, avoiding the direct comparison of biometric digests, console device sources, and authorized task scopes as the same type of data.

[0123] Preferably, when the console identity verification result is consistent, a session establishment record is first generated based on the handshake request data packet, the biometric identifier, and the console authentication configuration record. The session establishment record includes a task identifier, current execution cycle identifier, hardware console device identifier, authorized administrator identifier, probe digest, session establishment time, and session validity status. Subsequently, session key material is generated based on the session establishment record, and the session key material is written to the probe sending port of the intelligent agent and the probe receiving port of the hardware console, respectively, to establish the encrypted transmission tunnel. The technical essence of the encrypted transmission tunnel is a data transmission channel used to carry the human-computer interaction probe. It binds the human-computer interaction probe to the target task, the current execution cycle, the hardware console, and the authorized administrator identifier, ensuring that the human-computer interaction probe can only be received via the hardware console that has completed the identity verification.

[0124] Preferably, when the encrypted transmission tunnel is established, the human-computer interaction probe is segmented and indexed according to the probe transmission field. Specifically, the simplified state feature string, 3D spatial scene data, interactive anchor points, probe source index, and probe time index from the human-computer interaction probe are first read, and a probe integrity digest is generated based on the probe digest. Then, the simplified state feature string, the 3D spatial scene data, the interactive anchor points, the probe source index, and the probe time index are configured with probe data segmentation identifiers, and the data content corresponding to the probe data segmentation identifiers is sent sequentially through the encrypted transmission tunnel. The probe integrity digest is used to verify whether the human-computer interaction probe is consistent with the probe digest in the handshake request data packet after the hardware console receives the human-computer interaction probe. The probe data segmentation identifier is used to maintain the correspondence between the simplified state feature string, the 3D spatial scene data, and the interactive anchor points when the hardware console reassembles the human-computer interaction probe.

[0125] Preferably, the reason for sending the human-computer interaction probe to the hardware console via the encrypted transmission tunnel is that the human-computer interaction probe carries the task status content, environmental interaction change content, external interface call status, probe source index, and probe time index retained after data pruning of the on-site stack snapshot. If the consistency of the console identity comparison result is not confirmed first, and the human-computer interaction probe is sent directly, the panoramic restoration view may be presented on a hardware console that does not correspond to the target task, thus causing the source of the subsequent intervention map to be unable to point back to the on-site stack snapshot. By establishing the encrypted transmission tunnel before sending the human-computer interaction probe, the human-computer interaction probe can be transmitted along the authenticated hardware console channel, so that when the hardware console presents the panoramic restoration view based on the human-computer interaction probe, it can continue to retain the technical source relationship between the target task, the current execution cycle, the authorized administrator identifier, and the on-site stack snapshot.

[0126] Preferably, in the security patrol task of a smart garden, after the on-site stack snapshot records the environmental interaction candidate nodes corresponding to the prohibited climbing, the deviation scalar, the camera drawing boundary data, the smart broadcast feedback status, and the work order flow status, and generates the human-computer interaction probe based on the on-site stack snapshot, a preset console authentication configuration record is first obtained, and the handshake request data packet is sent to the hardware console according to the console authentication configuration record. After the hardware console returns the biometric identification code based on the handshake request data packet, it compares the biometric identification code with the locally stored authorized administrator feature database to generate the console identity comparison result; when the console identity comparison result is consistent, the encrypted transmission tunnel is established, and the human-computer interaction probe is sent to the hardware console through the encrypted transmission tunnel. Thus, the panoramic restoration view subsequently presented by the hardware console can maintain the same source link as the environmental interaction candidate nodes corresponding to the prohibited climbing, the smart broadcast feedback status, and the work order flow status, which facilitates the subsequent formation of an intervention map that can refer back to the on-site stack snapshot.

[0127] Optionally, releasing the agent's execution permissions based on the reconstructed on-site stack snapshot includes: Send a wake-up trigger signal to the task scheduling center of the intelligent agent; The task scheduling center switches the intelligent agent from the read-only state corresponding to the permission suspension state to the read-write state according to the wake-up trigger signal; In the read / write state, the reconstructed scene stack snapshot is loaded into the main working memory area of ​​the agent; The driver instruction sending port of the intelligent agent is opened, and the driver instruction sending port is used to restore the intelligent agent's ability to send driver instructions for the target task, thereby releasing the running permissions of the intelligent agent.

[0128] Preferably, the specific implementation process of releasing the agent's running permissions based on the reconstructed on-site stack snapshot is as follows: After the reconstructed on-site stack snapshot is formed, the task-level field area, stage-level field area, interface-level field area, timestamp field area, reconstruction source index, and reconstruction time index in the reconstructed on-site stack snapshot are read first, and the permission status field, task stage field, task status source information, and current execution cycle identifier in the reconstructed on-site stack snapshot are read. Subsequently, it is determined whether the agent is still in the permission suspended state according to the permission status field, and whether the reconstructed on-site stack snapshot belongs to the target task to be continued in the current execution cycle according to the current execution cycle identifier. Then, the state correction chain and the original on-site stack snapshot are referenced back according to the reconstruction source index to form a permission release pre-verification record. The permission release pre-verification record is used to carry the verification result that the reconstructed on-site stack snapshot can participate in the running permission release. When the wake-up trigger signal is generated later, the permission release pre-verification record is used as the signal source to avoid directly restoring the agent's drive instruction issuing capability before the reconstructed on-site stack snapshot has established a correspondence with the target task.

[0129] Preferably, the wake-up trigger signal is essentially an internal control signal used to cause the task scheduling center to transition from the waiting scheduling state under the permission suspension state to the snapshot continuation state, rather than a startup instruction to restart the target task. Specifically, the current execution cycle identifier, permission status field, task stage field, reconstruction source index, and reconstruction time index are first read from the permission release pre-verification record. Then, the current execution cycle identifier, permission status field, task stage field, reconstruction source index, and reconstruction time index are written into the wake-up signal bearer record. Subsequently, the snapshot position identifier of the reconstructed context stack snapshot in the memory control area is written into the wake-up signal bearer record, and the scheduling entry identifier corresponding to the task scheduling center is also written into the wake-up signal bearer record, thus forming the wake-up trigger signal based on the wake-up signal bearer record. The wake-up trigger signal carries the snapshot position identifier of the reconstructed context stack snapshot and the current execution cycle identifier, enabling the task scheduling center to read the reconstructed context stack snapshot within the same current execution cycle after receiving the wake-up trigger signal, instead of regenerating a new task running state from the initial task configuration.

[0130] Preferably, the specific implementation process of sending a wake-up trigger signal to the task scheduling center of the intelligent agent is as follows: After the wake-up trigger signal is formed, the scheduling queue identifier, task suspension position identifier, and drive instruction delivery port suspension identifier retained by the task scheduling center in the permission suspension state are read first, and the scheduling queue identifier, the task suspension position identifier, and the drive instruction delivery port suspension identifier are matched with the current execution cycle identifier in the wake-up trigger signal; then, when the scheduling queue identifier, the task suspension position identifier, and the drive instruction delivery port suspension identifier all point to the same current execution cycle, the wake-up trigger signal is sent to the task scheduling center. The scheduling queue identifier is used to define the queuing position of the target task in the task scheduling center. The task suspension position identifier is used to define the task stage field when the target task enters the permission suspension state. The driver instruction issuing port suspension identifier is used to define the port state in which the driver instruction issuing port is suspended in the permission suspension state. The scheduling queue identifier, the task suspension position identifier, and the driver instruction issuing port suspension identifier jointly participate in the process of sending the wake-up trigger signal, so that the wake-up trigger signal can continue the original target task instead of triggering a new target task.

[0131] Preferably, after receiving the wake-up trigger signal, the task scheduling center first reads the current execution cycle identifier, snapshot position identifier, task stage field, and reconstruction time index from the wake-up trigger signal. Then, based on the snapshot position identifier, it searches for the reconstructed state stack snapshot in the memory control area. Subsequently, it compares the task stage field in the reconstructed state stack snapshot with the task stage field corresponding to the task suspension position identifier, and verifies the chronological relationship between the reconstruction time index and the permission suspension trigger time position. When the task stage field in the reconstructed state stack snapshot can accommodate the task stage field corresponding to the task suspension position identifier, and the reconstruction time index is within the time range corresponding to the current execution cycle, the task scheduling center generates a state switching request record. The state transition request record is used to carry the state transition task source field, state transition time source field, and state transition snapshot source field required for the agent to switch from the read-only state corresponding to the permission suspension state to the read-write state; when the state transition is subsequently executed, the state transition request record limits the read-write state to the target task corresponding to the reconstructed on-site stack snapshot through the state transition task source field, the state transition time source field, and the state transition snapshot source field.

[0132] Preferably, when the task scheduling center switches the agent from the read-only state corresponding to the permission suspension state to the read-write state according to the wake-up trigger signal, it first reads the state switch task source field, state switch time source field, and state switch snapshot source field in the state switch request record, and reads the memory usage status, field write lock status, and snapshot loading entry status of the main working memory area; subsequently, when the memory usage status of the main working memory area can accommodate the reconstructed on-site stack snapshot, and the field write lock status allows writing to the task-level field area, the stage-level field area, the interface-level field area, and the timestamp field area, the agent is switched from the read-only state corresponding to the permission suspension state to the read-write state. The read-only state is used to restrict the on-site stack snapshot from being directly rewritten in the permission suspension state, and the read-write state is used to allow the main working memory area to receive snapshot data after the reconstructed on-site stack snapshot passes source, time, and stage verification; through the above state switching method, the formation of the read-write state has a direct source relationship with the reconstructed on-site stack snapshot.

[0133] Preferably, after the read / write state is established, a snapshot loading bearer record is first established. This snapshot loading bearer record is used to carry the field mapping data before the reconstructed on-site stack snapshot enters the main working memory area. The snapshot loading bearer record is not equivalent to the reconstructed on-site stack snapshot. Specifically, the task-level field area, stage-level field area, interface-level field area, timestamp field area, reconstruction source index, and reconstruction time index are read from the reconstructed on-site stack snapshot, and these fields are written into the snapshot loading bearer record. Subsequently, based on the task status storage location, stage status storage location, interface status storage location, timestamp storage location, and source index storage location in the main working memory area, a loading field mapping relationship is established between the snapshot loading bearer record and the main working memory area. The loading field mapping relationship is used to express which type of storage location in the main working memory area the field content in the reconstructed on-site stack snapshot should be written to. When the reconstructed on-site stack snapshot is loaded subsequently, the field writing is performed through the loading field mapping relationship.

[0134] Preferably, the specific implementation process of loading the reconstructed scene stack snapshot into the main working memory area of ​​the agent in the read / write state is as follows: First, according to the loading field mapping relationship, the task-level field area in the reconstructed scene stack snapshot is written to the task status storage location in the main working memory area to restore the task identifier, task type field, task path field, permission status field, and task status source information of the target task; then, the stage-level field area in the reconstructed scene stack snapshot is written to the stage status storage location in the main working memory area to restore the target task in inspection, identification, alarm, and linkage. Alternatively, it handles the local variable parameters in the corresponding execution phase of the feedback; then, it writes the interface-level field area from the reconstructed field stack snapshot into the interface status storage location in the main working memory area to restore the external interface call records corresponding to the camera frame extraction interface, environmental perception device reporting interface, smart broadcast feedback interface, smart lighting feedback interface, smart irrigation feedback interface, work order flow interface, and geographic information management platform spatial location interface; then, it writes the timestamp field area and the reconstructed time index into the timestamp storage location in the main working memory area, and writes the reconstructed source index into the source index storage location in the main working memory area. Through the above loading process, the main working memory area can accommodate the task running context, parameter status context, and external interface call context expressed by the reconstructed field stack snapshot.

[0135] Preferably, after the reconstructed on-site stack snapshot is loaded into the main working memory area, a load consistency check is performed on the main working memory area. The load consistency check first reads the snapshot load record, the load field mapping relationship, and the task status storage location, stage status storage location, interface status storage location, timestamp storage location, and source index storage location already written in the main working memory area. Then, it determines whether the task identifier in the task status storage location corresponds to the current execution cycle identifier, whether the task stage field in the stage status storage location can continue the task suspension position identifier, whether the external interface call record in the interface status storage location can point back to the state correction chain through the reconstructed source index in the source index storage location, and whether the reconstructed time index in the timestamp storage location is within the time range corresponding to the current execution cycle. After the load consistency check passes, a main working memory area continuation record is generated; this record carries the status result that the reconstructed on-site stack snapshot has entered the main working memory area and can participate in subsequent driver instruction issuance.

[0136] Preferably, the specific implementation process of opening the driver command sending port of the intelligent agent is as follows: After generating the main working memory area connection record, first read the port suspension state, port suspension source, port target object, and port allowed command type formed by the driver command sending port in the permission suspension state; wherein, the port suspension state is used to express the state in which the driver command sending port stops sending driver commands in the permission suspension state, the port suspension source is used to indicate the permission suspension command that triggered the suspension of the driver command sending port, the port target object is used to express the smart broadcast feedback interface, smart lighting feedback interface, smart irrigation feedback interface, or work order flow interface that the driver command sending port can send driver commands to, and the port allowed command type is used to express the device linkage command, task stage connection command, or work order flow command that the driver command sending port can send. Subsequently, the port target object and the port allowed command type are correspondingly processed with the main working memory area connection record to form a port opening verification record. The port opening verification record is used to determine whether the driver instruction issuing port can receive the connection status from the connection record in the main working memory area, rather than directly restoring the calling capability of all external interface objects.

[0137] Preferably, when the port open verification record indicates that the port target object corresponding to the driver instruction issuing port is consistent with the external interface object recorded in the interface status storage location in the main working memory area, and the port's allowed instruction type is consistent with the execution stage corresponding to the task stage field, the driver instruction issuing port is switched from a port suspended state to a port open state. Subsequently, a driver instruction continuation queue is generated based on the task stage field, external interface call record, reconstruction source index, and reconstruction time index in the continuation record of the main working memory area. The driver instruction continuation queue is used to queue the driver instructions corresponding to the smart broadcast feedback interface, smart lighting feedback interface, smart irrigation feedback interface, or work order flow interface according to the task stage field recovered from the reconstructed field stack snapshot. Each driver instruction in the driver instruction continuation queue points back to the continuation record in the main working memory area, enabling the driver instruction issuing port to send driver instructions based on the reconstructed field stack snapshot when the port is open.

[0138] Preferably, when restoring the agent's ability to issue drive commands for the target task through the drive command issuance port, the first drive command in the drive command continuation queue is read first, along with the corresponding task stage field, port target object, external interface call record, and reconstruction time index. Then, the task stage field corresponding to the first drive command is mapped to the task stage field of the target task in the task scheduling center; the port target object corresponding to the first drive command is mapped to the port target object of the drive command issuance port; and the external interface call record corresponding to the first drive command is mapped to the interface status storage location in the main working memory area. After the above mapping process is successful, the first drive command is sent through the drive command issuance port, and a drive command issuance status is written to the main working memory area. The drive command issuance status indicates that the first drive command has been issued to the corresponding external interface object based on the reconstructed on-site stack snapshot. Subsequently, the task scheduling center continues to advance the target task based on the drive command issuance status.

[0139] Preferably, the release of the agent's running permission is not simply a change from the permission suspension state to a state where drive instructions can be issued, but is formed jointly by the reconstructed context stack snapshot, the read / write state, the main working memory area continuation record, the port open state of the drive instruction issuance port, and the drive instruction issuance state. Specifically, first, based on the read / write state, it is confirmed that the reconstructed context stack snapshot can be written to the main working memory area; then, based on the main working memory area continuation record, it is confirmed that the task running context of the target task has entered the main working memory area; subsequently, based on the port open state of the drive instruction issuance port, it is confirmed that the drive instruction issuance port can receive drive instructions in the drive instruction continuation queue; finally, based on the drive instruction issuance state, it is confirmed that the agent has restored its drive instruction issuance capability for the target task. The read / write status, the main working memory area continuation record, the port open status of the driver instruction issuing port, and the driver instruction issuing status are all written into the running permission release record. The running permission release record is used to carry the source of the permission status, the source of the reconstruction snapshot, the source of the main working memory area, and the source of the driver instruction issuing port when the agent completes the running permission release, so that the agent can continue to advance the target task according to the reconstructed on-site stack snapshot.

[0140] Preferably, in the security patrol task of a smart garden, after the reconstructed site stack snapshot records the environmental interaction candidate node corresponding to the prohibited climbing, the task stage field corresponding to the linkage stage, the external interface call record corresponding to the smart broadcast feedback interface, and the external interface call record corresponding to the work order flow status, the task scheduling center first receives the wake-up trigger signal and switches the intelligent agent from the read-only state corresponding to the permission suspension state to the read-write state according to the wake-up trigger signal; then, in the read-write state, the task-level field area, stage-level field area, interface-level field area, and timestamp field area in the reconstructed site stack snapshot are loaded into the main working memory area; then, according to the continuation record in the main working memory area, the driver instruction issuing port is opened, so that the driver instruction issuing port first restores the driver instruction issuing capability corresponding to the smart broadcast feedback interface, and then restores the driver instruction issuing capability corresponding to the work order flow status. Through this continuation method, the intelligent agent does not re-execute the security patrol task, but continues to advance the target task along the environmental interaction candidate node corresponding to the prohibited climbing, the linkage stage, the smart broadcast feedback interface, and the work order flow status.

[0141] Preferably, in the task of linking environmental perception and intelligent irrigation, after the reconstructed field stack snapshot records the external interface call records corresponding to the environmental perception device reporting interface, the local variable parameters corresponding to the environmental perception data state changes, the external interface call records corresponding to the intelligent irrigation feedback interface, and the task stage fields corresponding to the linkage stage, the task scheduling center enters the snapshot continuation state according to the wake-up trigger signal, and loads the reconstructed field stack snapshot into the main working memory area in the read-write state; subsequently, the main working memory area continuation records establish a stage source correspondence relationship between the environmental perception data state changes and the external interface call records corresponding to the intelligent irrigation feedback interface, and the driver instruction issuing port restores the driver instruction issuing capability corresponding to the intelligent irrigation feedback interface according to the stage source correspondence relationship. In this example, the release of running permissions can preserve the continuation relationship between the environmental perception device reporting, environmental perception data state changes, intelligent irrigation feedback interface, and linkage stage, allowing the target task to continue from the state recorded in the reconstructed field stack snapshot.

[0142] like Figure 3 As shown, this embodiment of the present application discloses a dynamic permission handover device for an intelligent agent with a human-machine symbiotic architecture, which includes: The node sequence extraction module is used to obtain the real-time running feature stream of the target task and extract the environmental interaction node sequence in the real-time running feature stream. The intent deviation detection module is used to obtain the preset trigger baseline corresponding to the target task, perform intent deviation detection on the environmental interaction node sequence, and generate a deviation scalar. The permission suspension control module is used to generate a permission suspension instruction when the deviation scalar reaches the preset trigger baseline, and to switch the running permission of the agent in the current execution cycle to a permission suspension state based on the permission suspension instruction; The on-site stack snapshot recording module is used to record the on-site stack snapshot of the agent in the current execution cycle under the permission suspension state. The on-site stack snapshot carries the task status source information corresponding to the current execution cycle. The human-computer interaction probe generation module is used to generate a human-computer interaction probe based on the on-site stack snapshot, and send the human-computer interaction probe to the hardware console so that the hardware console can present a panoramic restoration view based on the human-computer interaction probe. The intervention map receiving module is used to receive the intervention map issued by the hardware console for the panoramic restoration view; The instruction translation module is used to perform instruction translation operations on the intervention map to obtain the state correction chain; The on-site stack snapshot reconstruction module is used to reconstruct the on-site stack snapshot based on the state correction chain to form a reconstructed on-site stack snapshot. The execution permission release module is used to release the execution permission of the agent based on the reconstructed on-site stack snapshot, and to guide the agent to continue to advance the target task according to the reconstructed on-site stack snapshot.

[0143] like Figure 4 As shown, an electronic device according to an embodiment of this application is characterized by including a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements the intelligent agent dynamic permission handover method with human-machine symbiotic architecture as described in any one of the claims of this application.

[0144] like Figure 5 As shown, a computer-readable storage medium is provided in an embodiment of this application. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the dynamic permission handover method for an intelligent agent with a human-machine symbiotic architecture as described in any one of the claims of this application.

[0145] like Figure 6 As shown, this is an embodiment of an intelligent agent according to this application, which includes a task scheduling center, a main working memory area, a driver instruction issuing port, and a dynamic permission handover component; The task scheduling center is used to execute target tasks and generate a real-time running feature stream during the execution of the target tasks; The dynamic permission handover component is used to acquire the real-time running feature stream of the target task, extract the environmental interaction node sequence in the real-time running feature stream, acquire the preset trigger baseline corresponding to the target task, perform intent deviation detection on the environmental interaction node sequence, generate a deviation scalar, and when the deviation scalar reaches the preset trigger baseline, generate a permission suspension instruction, and switch the running permission of the agent in the current execution cycle to a permission suspension state based on the permission suspension instruction. The dynamic permission handover component is also used to record a snapshot of the agent's field stack in the current execution cycle when the permission is suspended. The snapshot carries the task status source information corresponding to the current execution cycle. A human-computer interaction probe is generated based on the snapshot and sent to the hardware console so that the hardware console can present a panoramic restoration view based on the human-computer interaction probe. The dynamic permission handover component is also used to receive the intervention map issued by the hardware console for the panoramic restoration view, perform instruction translation operation on the intervention map to obtain the state correction chain, and reconstruct the field stack snapshot based on the state correction chain to form the reconstructed field stack snapshot. The main working memory area is used to load the reconstructed on-site stack snapshot; The drive instruction issuing port is used to restore the agent's ability to issue drive instructions for the target task after the dynamic permission handover component releases the agent's running permission based on the reconstructed field stack snapshot, so that the agent can continue to advance the target task according to the reconstructed field stack snapshot.

[0146] like Figure 7 As shown, this application embodiment provides a dynamic permission handover system for an intelligent agent with a human-machine symbiotic architecture, which includes an intelligent agent, a hardware control console, and a dynamic permission handover device. The intelligent agent is used to execute the target task and provides a real-time running feature stream during the execution of the target task; The dynamic permission handover device is used to acquire the real-time running feature stream of the target task, extract the environmental interaction node sequence in the real-time running feature stream, acquire the preset trigger baseline corresponding to the target task, perform intent deviation detection on the environmental interaction node sequence, generate a deviation scalar, and when the deviation scalar reaches the preset trigger baseline, generate a permission suspension instruction, and switch the running permission of the agent in the current execution cycle to a permission suspension state based on the permission suspension instruction. The dynamic permission handover device is also used to record a snapshot of the agent's field stack in the current execution cycle when the permission is suspended. The snapshot carries the task status source information corresponding to the current execution cycle. A human-computer interaction probe is generated based on the snapshot and sent to the hardware console. The hardware console is used to present a panoramic restoration view based on the human-computer interaction probe, receive human intervention inputs formed in response to the panoramic restoration view, and generate an intervention atlas based on the human intervention inputs. The dynamic permission handover device is also used to receive the intervention map issued by the hardware console for the panoramic restoration view, perform instruction translation operation on the intervention map to obtain a state correction chain, reconstruct the field stack snapshot based on the state correction chain to form a reconstructed field stack snapshot, release the running permission of the agent based on the reconstructed field stack snapshot, and guide the agent to continue to advance the target task according to the reconstructed field stack snapshot.

[0147] The above Figures 3-7 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. A method for dynamic permission handover of an intelligent agent with a human-machine symbiotic architecture, characterized in that, include: Obtain the real-time running feature stream of the target task, and extract the environmental interaction node sequence from the real-time running feature stream; Obtain the preset trigger baseline corresponding to the target task, perform intent deviation detection on the environmental interaction node sequence, and generate a deviation scalar. When the deviation scalar reaches the preset trigger baseline, a permission suspension instruction is generated, and based on the permission suspension instruction, the running permission of the agent in the current execution cycle is switched to a permission suspension state. In the permission-suspended state, a snapshot of the agent's context stack is recorded in the current execution cycle, and the context stack snapshot carries the task state source information corresponding to the current execution cycle; A human-computer interaction probe is generated based on the on-site stack snapshot, and the human-computer interaction probe is sent to the hardware console so that the hardware console can present a panoramic restoration view based on the human-computer interaction probe; Receive the intervention map issued by the hardware console for the panoramic restoration view; The intervention map is translated to obtain a state correction chain. The field stack snapshot is reconstructed based on the state correction chain to form a reconstructed field stack snapshot. Based on the reconstructed on-site stack snapshot, the agent's running permission is released, and the agent is guided to continue advancing the target task according to the reconstructed on-site stack snapshot; Optionally, the step of performing a command translation operation on the intervention map to obtain a state correction chain includes: Extract the human control action sequences from the intervention map; Perform an action semantic decomposition operation on the human-controlled action sequence to output an atomic-level operation instruction sequence through the action semantic decomposition operation; Obtain a preset intelligent agent hardware execution constraint table, which includes hardware execution order constraints, hardware execution port constraints, and hardware execution state constraints corresponding to the intelligent agent; The atomic-level operation instruction sequence is sorted and reorganized according to the intelligent agent hardware execution constraint table; The state correction chain is generated based on the sorted and reorganized atomic-level operation instruction sequence; The human-controlled action sequence includes at least one human-controlled action record, and each human-controlled action record includes an action source field, an action target field, an action type field, an action parameter field, an action time field, and an anchor point retracement field. The recording of the agent's current execution stack snapshot includes: In the permission suspension state corresponding to the permission suspension instruction, the internal bus data stream of the intelligent agent is intercepted; Extract the global and local variable parameters from the internal bus data stream; Capture the external interface call records generated by the intelligent agent during the current execution cycle; The global variable parameters, the local variable parameters, the external interface call records, and the timestamp tag corresponding to the current execution cycle are encapsulated into the on-site stack snapshot.

2. The method according to claim 1, characterized in that, The extraction of the environmental interaction node sequence from the real-time running feature stream includes: The real-time running feature stream is divided into at least two temporal feature slices; A spatiotemporal correlation comparison operation is performed on each of the time-series feature slices to identify mutation feature parameters in each of the time-series feature slices through the spatiotemporal correlation comparison operation; The temporal feature slice containing the mutation feature parameter is marked as an environmental interaction candidate node; Extract at least two consecutively occurring environmental interaction candidate nodes, and concatenate the extracted at least two environmental interaction candidate nodes in chronological order to form the environmental interaction node sequence.

3. The method according to claim 1, characterized in that, The step of performing intent deviation detection on the sequence of environmental interaction nodes and generating a deviation scalar includes: Obtain the pre-configured initial intent map corresponding to the target task; A semantic node mapping operation is performed on the environmental interaction node sequence to extract the actual execution graph of the environmental interaction node sequence through the semantic node mapping operation; Compare the topological differences between the actual execution map and the initial intention map; Node distance parameters are extracted from the topology differences, and the deviation scalar is generated based on the node distance parameters.

4. The method according to claim 1, characterized in that, The step of generating the human-computer interaction probe based on the on-site stack snapshot includes: Obtain the preset human-computer interaction probe pruning rules, and perform data pruning on the field stack snapshot according to the human-computer interaction probe pruning rules to obtain a simplified state feature string; Obtain a preset visual rendering pipeline configuration, and transmit the simplified state feature string to the visual rendering pipeline according to the preset visual rendering pipeline configuration to generate three-dimensional spatial scene data for the hardware console to present the panoramic restored view; Identify the interactive anchor points in the three-dimensional spatial scene data; The human-computer interaction probe is generated based on the simplified state feature string, the three-dimensional spatial scene data, and the interactive anchor points.

5. A dynamic permission handover device for an intelligent agent with a human-machine symbiotic architecture, characterized in that, include: The node sequence extraction module is used to obtain the real-time running feature stream of the target task and extract the environmental interaction node sequence in the real-time running feature stream. The intent deviation detection module is used to obtain the preset trigger baseline corresponding to the target task, perform intent deviation detection on the environmental interaction node sequence, and generate a deviation scalar. The permission suspension control module is used to generate a permission suspension instruction when the deviation scalar reaches the preset trigger baseline, and to switch the running permission of the agent in the current execution cycle to a permission suspension state based on the permission suspension instruction; The on-site stack snapshot recording module is used to record the on-site stack snapshot of the agent in the current execution cycle under the permission suspension state. The on-site stack snapshot carries the task status source information corresponding to the current execution cycle. The human-computer interaction probe generation module is used to generate a human-computer interaction probe based on the on-site stack snapshot, and send the human-computer interaction probe to the hardware console so that the hardware console can present a panoramic restoration view based on the human-computer interaction probe. The intervention map receiving module is used to receive the intervention map issued by the hardware console for the panoramic restoration view; The instruction translation module is used to perform instruction translation operations on the intervention map to obtain the state correction chain; The on-site stack snapshot reconstruction module is used to reconstruct the on-site stack snapshot based on the state correction chain to form a reconstructed on-site stack snapshot. The execution permission release module is used to release the execution permission of the agent based on the reconstructed scene stack snapshot, and to guide the agent to continue to advance the target task according to the reconstructed scene stack snapshot. The step of performing instruction translation operations on the intervention map to obtain a state correction chain includes: Extract the human control action sequences from the intervention map; Perform an action semantic decomposition operation on the human-controlled action sequence to output an atomic-level operation instruction sequence through the action semantic decomposition operation; Obtain a preset intelligent agent hardware execution constraint table, which includes hardware execution order constraints, hardware execution port constraints, and hardware execution state constraints corresponding to the intelligent agent; The atomic-level operation instruction sequence is sorted and reorganized according to the intelligent agent hardware execution constraint table; The state correction chain is generated based on the sorted and reorganized atomic-level operation instruction sequence; The human-controlled action sequence includes at least one human-controlled action record, and each human-controlled action record includes an action source field, an action target field, an action type field, an action parameter field, an action time field, and an anchor point retracement field. The recording of the agent's current execution stack snapshot includes: In the permission suspension state corresponding to the permission suspension instruction, the internal bus data stream of the intelligent agent is intercepted; Extract the global and local variable parameters from the internal bus data stream; Capture the external interface call records generated by the intelligent agent during the current execution cycle; The global variable parameters, the local variable parameters, the external interface call records, and the timestamp tag corresponding to the current execution cycle are encapsulated into the on-site stack snapshot.

6. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein when the processor executes the computer program, it implements the intelligent agent dynamic permission handover method with a human-machine symbiotic architecture as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the dynamic permission handover method for an intelligent agent with a human-machine symbiotic architecture as described in any one of claims 1 to 4.

8. An intelligent agent, characterized in that, This includes a task scheduling center, main working memory area, driver instruction delivery port, and dynamic permission handover component; The task scheduling center is used to execute the target task and to generate a real-time running feature stream during the execution of the target task; The dynamic permission handover component is used to acquire the real-time running feature stream of the target task, extract the environmental interaction node sequence in the real-time running feature stream, acquire the preset trigger baseline corresponding to the target task, perform intent deviation detection on the environmental interaction node sequence, generate a deviation scalar, and when the deviation scalar reaches the preset trigger baseline, generate a permission suspension instruction, and switch the running permission of the agent in the current execution cycle to a permission suspension state based on the permission suspension instruction. The dynamic permission handover component is also used to record a snapshot of the agent's field stack in the current execution cycle when the permission is suspended. The snapshot carries the task status source information corresponding to the current execution cycle. A human-computer interaction probe is generated based on the snapshot and sent to the hardware console so that the hardware console can present a panoramic restoration view based on the human-computer interaction probe. The dynamic permission handover component is also used to receive the intervention map issued by the hardware console for the panoramic restoration view, perform instruction translation operation on the intervention map to obtain the state correction chain, and reconstruct the field stack snapshot based on the state correction chain to form the reconstructed field stack snapshot. The main working memory area is used to load the reconstructed on-site stack snapshot; The drive instruction issuing port is used to restore the intelligent agent's ability to issue drive instructions for the target task after the dynamic permission handover component releases the intelligent agent's running permission based on the reconstructed field stack snapshot, so that the intelligent agent can continue to advance the target task according to the reconstructed field stack snapshot; The step of performing instruction translation operations on the intervention map to obtain a state correction chain includes: Extract the human control action sequences from the intervention map; Perform an action semantic decomposition operation on the human-controlled action sequence to output an atomic-level operation instruction sequence through the action semantic decomposition operation; Obtain a preset intelligent agent hardware execution constraint table, which includes hardware execution order constraints, hardware execution port constraints, and hardware execution state constraints corresponding to the intelligent agent; The atomic-level operation instruction sequence is sorted and reorganized according to the intelligent agent hardware execution constraint table; The state correction chain is generated based on the sorted and reorganized atomic-level operation instruction sequence; The human-controlled action sequence includes at least one human-controlled action record, and each human-controlled action record includes an action source field, an action target field, an action type field, an action parameter field, an action time field, and an anchor point retracement field. The recording of the agent's current execution stack snapshot includes: In the permission suspension state corresponding to the permission suspension instruction, the internal bus data stream of the intelligent agent is intercepted; Extract the global and local variable parameters from the internal bus data stream; Capture the external interface call records generated by the intelligent agent during the current execution cycle; The global variable parameters, the local variable parameters, the external interface call records, and the timestamp tag corresponding to the current execution cycle are encapsulated into the on-site stack snapshot.

9. A dynamic permission handover system for intelligent agents with a human-machine symbiotic architecture, characterized in that, This includes intelligent agents, hardware control consoles, and dynamic permission handover devices; The intelligent agent is used to execute the target task and provides a real-time running feature stream during the execution of the target task; The dynamic permission handover device is used to acquire the real-time running feature stream of the target task, extract the environmental interaction node sequence in the real-time running feature stream, acquire the preset trigger baseline corresponding to the target task, perform intent deviation detection on the environmental interaction node sequence, generate a deviation scalar, and when the deviation scalar reaches the preset trigger baseline, generate a permission suspension instruction, and switch the running permission of the agent in the current execution cycle to a permission suspension state based on the permission suspension instruction. The dynamic permission handover device is also used to record a snapshot of the agent's field stack in the current execution cycle when the permission is suspended. The snapshot carries the task status source information corresponding to the current execution cycle. A human-computer interaction probe is generated based on the snapshot and sent to the hardware console. The hardware console is used to present a panoramic restoration view based on the human-computer interaction probe, receive human intervention inputs formed in response to the panoramic restoration view, and generate an intervention atlas based on the human intervention inputs. The dynamic permission handover device is also used to receive the intervention map issued by the hardware console for the panoramic restoration view, perform instruction translation operation on the intervention map to obtain the state correction chain, reconstruct the field stack snapshot based on the state correction chain to form a reconstructed field stack snapshot, release the running permission of the agent based on the reconstructed field stack snapshot, and guide the agent to continue to advance the target task according to the reconstructed field stack snapshot. The step of performing instruction translation operations on the intervention map to obtain a state correction chain includes: Extract the human control action sequences from the intervention map; Perform an action semantic decomposition operation on the human-controlled action sequence to output an atomic-level operation instruction sequence through the action semantic decomposition operation; Obtain a preset intelligent agent hardware execution constraint table, which includes hardware execution order constraints, hardware execution port constraints, and hardware execution state constraints corresponding to the intelligent agent; The atomic-level operation instruction sequence is sorted and reorganized according to the intelligent agent hardware execution constraint table; The state correction chain is generated based on the sorted and reorganized atomic-level operation instruction sequence; The human-controlled action sequence includes at least one human-controlled action record, and each human-controlled action record includes an action source field, an action target field, an action type field, an action parameter field, an action time field, and an anchor point retracement field. The recording of the agent's current execution stack snapshot includes: In the permission suspension state corresponding to the permission suspension instruction, the internal bus data stream of the intelligent agent is intercepted; Extract the global and local variable parameters from the internal bus data stream; Capture the external interface call records generated by the intelligent agent during the current execution cycle; The global variable parameters, the local variable parameters, the external interface call records, and the timestamp tag corresponding to the current execution cycle are encapsulated into the on-site stack snapshot.

Images