Large model intelligent agent unmanned aerial vehicle task planning and instruction generation method and system

Abstract

The application discloses a large model intelligent agent unmanned aerial vehicle task planning and instruction generation method and system, and belongs to the technical field of unmanned aerial vehicle intelligent control and task planning. The method freezes the task state frame, constraint set and instruction template version of the current planning period, and constructs a constraint causal task graph with a constraint responsibility chain; the large model intelligent agent only generates a candidate planning and a planning anchor point table, performs safe projection according to a conflict mask, and generates a typed instruction queue; the task, format and constraint gate are used to identify the mismatch reason, match the corresponding repair subset, perform one-time re-projection, and output a degraded instruction when the re-projection fails; and the execution feedback is only written into the next planning period. Therefore, the planning traceability, instruction safety and processing termination are improved.

Description

Technical Field

[0001] This invention relates to the fields of intelligent control, mission planning, flight control command generation, and artificial intelligence technology for unmanned aerial vehicles (UAVs), and particularly to a method and system for mission planning and command generation for large-scale intelligent agent UAVs. Background Technology

[0002] As the application of drones in scenarios such as inspection, reconnaissance, target search, emergency rescue, low-altitude logistics and collaborative operations continues to expand, the tasks performed by drones are gradually changing from single waypoint flights to complex tasks that include task semantic understanding, task sub-objective decomposition, route selection, time scheduling, platform capability constraints, communication reachability constraints, safety avoidance and execution feedback processing.

[0003] Existing UAV mission planning technologies typically generate mission plans using preset rules, path search algorithms, optimization algorithms, or manual configuration. When using these methods, the task requirements in natural language form need to be converted into structured task parameters beforehand. When the task description is complex, has many types of constraints, or the environmental situation changes rapidly, problems such as low efficiency in task semantic parsing, difficulty in uniformly expressing the coupling relationships between different constraints, and difficulty in tracing the sources of planning conflicts can easily arise.

[0004] Large-scale intelligent agents possess natural language understanding, task decomposition, and tool invocation capabilities, enabling them to generate task plans based on task text, environmental information, and UAV status. However, the output of large-scale intelligent agents is probabilistic, potentially leading to issues such as inconsistencies between the generated content and the original task objective, lack of data sources for planning elements, omission of airspace or platform capability constraints, and output formats that do not meet flight control interface requirements. Directly converting the planning results generated by large-scale intelligent agents into flight control commands may result in semantic shifts, format errors, route violations, altitude exceeding limits, communication unavailability, or platform capability exceeding limitations directly impacting the UAV execution chain.

[0005] To improve the security of large-scale model planning results, some existing solutions incorporate rule validation, path planning, security checks, or feedback correction processes after the large model output. However, these existing solutions still have the following shortcomings:

[0006] First, there is a lack of stable correspondence between task fields, planning elements, graph nodes, graph edges, and constraint sources, making it difficult to identify fictitious planning elements that cannot be traced back to the original task or constraint data, and also making it difficult to determine the constraint sources and affected variables corresponding to conflicting planning elements.

[0007] Second, some solutions prompt the large model agent to generate a new plan after discovering a planning error, or repeatedly modify the current plan based on feedback. Due to the lack of clear limits on the number of repairs and termination conditions, this easily leads to multiple generation, multiple verifications, or repeated replanning, increasing instruction generation latency and potentially creating empty or infinite loops.

[0008] Third, existing verification methods typically treat task semantic errors, instruction format errors, and security constraint violations as the same type of failure, failing to select corresponding repair data based on the cause of the failure. This makes it easy to use incorrect repair methods to handle mismatched failure types. For example, format errors are not suitable for rerouting, and task semantic offsets cannot be eliminated simply by correcting field formats.

[0009] Fourth, if environmental updates or execution feedback modify data already involved in the planning, executable plans, or the current instruction queue within the same planning cycle, it may cause inconsistencies in data versions within the same planning cycle, resulting in pre- and post-processing steps being based on different versions of task states or constraints, thereby causing planning oscillations and instruction write-backs.

[0010] Therefore, a new method and system for UAV mission planning and command generation by a large model intelligent agent is needed. While retaining the semantic understanding capability of the large model intelligent agent, the system implements data anchoring, conflict localization, deterministic projection, cause matching and repair, one-time reprojection, and cross-planning cycle feedback isolation on its output to ensure that the UAV mission planning and flight control command generation process is safe, traceable, verifiable, and can be terminated in a deterministic manner. Summary of the Invention

[0011] The purpose of this invention is to provide a method and system for large-scale intelligent agent unmanned aerial vehicle (UAV) mission planning and instruction generation, in order to solve the following problems existing in the prior art:

[0012] Planning elements generated by large model agents are difficult to trace back to task fields, constraint sources, and graph relationships; planning elements without semantic sources may enter the flight control command conversion process; different types of planning conflicts are difficult to locate to specific constraint sources and affected variables; there is a lack of matching repair paths for task semantic mismatch, command format mismatch, and constraint violations; repeated calls to large models or repeated repairs after planning failures can easily increase processing latency and generate loops; execution feedback writing back to the current command queue within the current planning cycle can easily lead to data version drift, processing loops, and command oscillations.

[0013] To achieve the above objectives, this invention freezes the timestamps and version numbers of the task status frame, constraint set, planning output format constraints, and mapping table at the beginning of the planning cycle; constructs a constraint causal task graph with constraint sources, affected variables, and constraint responsibility chains; establishes the correspondence between candidate planning elements and graph nodes, graph edges, and task fields through a planning anchor table; writes semantically unanchored elements and conflict planning elements into a conflict mask; and sets the constraint set... Divide into hard constraint subsets that must be satisfied and a subset of residual risk used to evaluate residual risk In the hard constraint subset and conflict mask Within the limited feasible planning space, based on the aforementioned subset of remaining risks Candidate Programs Projection as an executable plan Generate or call constraint conflict repair subsets, format repair subsets, and semantic anchoring repair subsets according to different gating mismatch reasons; perform reprojection only once without calling the large model agent again and without adding new task objectives; and write the execution feedback only to the next planning cycle to form a task planning and instruction generation chain with consistent data versions, matching failure reasons and repair methods, limited processing times, and no feedback write-back in the current cycle.

[0014] To achieve the aforementioned objective, this invention provides a method for large-scale intelligent agent unmanned aerial vehicle (UAV) mission planning and instruction generation, comprising:

[0015] During the planning period At the start, freeze the task status frame with the source identifier. Constraint Sets Planning output format constraints And a mapping table from planning primitives to flight control command templates timestamp and version number, and constrain the set Divided into hard-constrained subsets and residual risk subset Based on the task status frame and constraint set Constructing a constrained causal task graph and image embedding And write the constraint source, affected variables, and constraint responsibility chain list into the conflict edges; embed the graph. Constraint Sets and planning output format constraints Input a large model agent and generate only candidate plans. And planning anchor point table From the planning anchor point table Associate candidate planning elements with the constraint causal task graph Nodes, edges, and task status frames Task fields; based on the planning anchor table Reading conflict edges to form a conflict mask And write semantically unanchored elements that cannot be associated into the conflict mask. In hard-constrained subsets and conflict mask Within the limited feasible planning space, based on the remaining risk subset Candidate Programs Projection as an executable plan and according to the conflict mask Generate constraint conflict repair subset Based on the planned output format constraints and mapping table Executable plan Convert to a typed instruction queue It performs checks on task consistency, format compliance, and constraint compliance; if a check fails, it generates a gating mismatch mask. According to the planning output format constraints Mapping table and instruction template fingerprint generation format repair subset Based on the task status frame Candidate planning And planning anchor point table Generate semantic anchoring repair subset According to the gate mismatch mask Repairing subsets from constraint conflicts Format Repair Subset and semantic anchoring repair subset Select the repair items to form a first reprojection set. Without invoking a large model agent or adding new task objectives, perform only one reprojection; if the reprojection still fails the judgment, perform another reprojection. If the feasible planning space is empty or empty, output a degradation instruction and terminate the current planning cycle; execute feedback. Only write to the input buffer for the next planning period, do not write back to the typed instruction queue for the current planning period. .

[0016] To achieve the aforementioned objective, this invention also provides a large-scale intelligent agent UAV mission planning and instruction generation system, including a data graph construction module, used during the planning cycle. Freeze the task status frame with source identifier Constraint Sets Planning output format constraints and mapping table The timestamp and version number will constrain the set. Divided into hard-constrained subsets and residual risk subset Construct a constrained causal task graph and image embedding The system writes the constraint source, affected variables, and constraint responsibility chain list into the conflict edges; a planning projection module is used to embed the graph. Constraint Sets and planning output format constraints Input a large model agent and generate only candidate plans. And planning anchor point table From the planning anchor point table Associative candidate planning elements and constraint causal task graph Nodes, edges, and task status frames The task field reads conflict edges and combines them with semantically anchorless elements to form a conflict mask. In hard-constrained subsets and conflict mask Within the limited feasible planning space, based on the aforementioned subset of remaining risks The candidate plans Projection as an executable plan Command gating module, used to control the conflict mask. Generate constraint conflict repair subset The executable plan Convert to a typed instruction queue Generate a format repair subset and semantic anchoring repair subset It performs checks on task consistency, format compliance, and constraint compliance; if any check fails, it generates a gating mismatch mask. Based on this, a reprojection set is formed. Without invoking the large model agent or adding new task objectives, only one reprojection is performed. If the reprojection still fails the judgment, a reprojection set is obtained. When the feasible planning space is empty or empty, a degradation instruction is output and the current planning cycle is terminated; the feedback isolation module is used to isolate execution feedback. Only write to the input buffer for the next planning period, do not write back to the typed instruction queue for the current planning period. .

[0017] Compared with the prior art, the present invention has at least the following beneficial effects.

[0018] First, the present invention freezes the task status frame at the start of the current planning cycle. Constraint Sets Planning output format constraints and mapping table The source identifier, timestamp, and version number ensure that the constraint causal task graph, candidate plan, executable plan, repair subset, and typed instruction queue generated within the current planning cycle are all based on the same version of data, avoiding inconsistencies in data within the same planning cycle due to environment updates or execution feedback.

[0019] Second, this invention constrains causal task graphs. The conflict edges are written with constraint sources, affected variables, and constraint responsibility chains, so that security conflicts, capacity overruns, communication unreachability, and timing conflicts can be traced back to the corresponding source data, constraint items, and affected planning variables, thereby improving the interpretability and auditability of the planning conflict location and remediation process.

[0020] Third, this invention uses a planned anchor point table. Establish a correspondence between candidate planning elements and nodes, edges, and task fields in the constraint causal task graph and task state frame. Define planning elements that cannot establish a correspondence as semantically unanchored elements. This prevents content generated by the large model agent but not traceable to task or constraint data from entering the typed instruction conversion process, reducing the risk of large model illusions directly entering the UAV flight control link.

[0021] Fourth, this invention limits the large model agent to generating only candidate plans. It is not allowed to directly generate flight control commands that can be issued; candidate plans need to be included in the hard constraint subset. and conflict mask Within the limited feasible planning space, based on the remaining risk subset Only after completing deterministic projections can an executable plan be formed. This ensures that the executable plan meets the hard constraints, while reducing the residual risk in the feasible plan that meets the hard constraints, thereby establishing a safe isolation between the probabilistic large model output and the deterministic flight control execution.

[0022] Fifth, this invention utilizes a collision mask. The conflict relationships in the constraint causal task graph are converted into conflict markers at the planning element level. Constraint conflict repair subsets are generated only for planning elements hit by the conflict mask. This can reduce unnecessary modifications to non-conflict planning elements, allowing the repaired plan to maintain the original task structure and task intent.

[0023] Sixth, this invention generates or invokes constraint conflict repair subsets for constraint violations, format mismatches, and semantic mismatches, respectively. Format Repair Subset and semantic anchoring repair subset And based on the gating mismatch mask Select the corresponding repair item to match the repair method with the cause of failure, and avoid using route repair to handle field errors or using format repair to handle task semantic offsets.

[0024] Seventh, when the format compliance judgment fails, this invention outputs the format according to the planned constraints. Mapping table The command template fingerprint can repair missing fields, incorrect units, incorrect field types, incorrect enumeration values, incorrect field order, or missing rollback conditions without changing the mission objective, route intent, and execution object of the executable plan, thereby improving the reliability of the flight control command format while maintaining the semantics of the plan.

[0025] Eighth, in the event of semantic mismatch, this invention relies on the task state frame. Candidate planning And planning anchor point table Restoring the correspondence between task sub-goals, execution objects, planning primitives or action sequences and task fields eliminates the need to re-invoke the large model agent to regenerate candidate plans, thus reducing task goal drift caused by repeated generation.

[0026] Ninth, when the gating judgment fails, the present invention only allows one reprojection under the condition that the large model agent is not called and no new task target is added; if the judgment still fails after reprojection, the reprojection set is empty or the feasible space is empty, the degradation instruction is directly output and the current planning cycle is terminated, so that the planning and repair process has a clear limit on the number of times and the termination path, avoiding infinite replanning, empty loops and dead loops.

[0027] Tenth, this invention will perform feedback. Only write to the input buffer for the next planning period, do not write back to the typed instruction queue for the current planning period. In the next planning cycle At the beginning, the execution feedback will be... Together with the new task input for the next planning cycle, they form a task state frame. and constraint set Based on this, a constrained causal task graph is constructed. This ensures that the data processing chain in the current planning cycle continues to advance in a unidirectional manner, avoiding data loops, version drift, and instruction oscillations caused by execution feedback modifying the current instruction queue in the opposite direction. Attached Figure Description

[0028] Figure 1 This is an overall flowchart of a large-scale intelligent agent unmanned aerial vehicle (UAV) mission planning and instruction generation method provided in an embodiment of the present invention.

[0029] Figure 2 This is a block diagram of a large-scale intelligent agent unmanned aerial vehicle (UAV) mission planning and instruction generation system provided in an embodiment of the present invention.

[0030] Figure 3 This is a schematic diagram illustrating the construction relationship between the task state frame, constraint set, constraint causal task graph, constraint responsibility chain list, and graph embedding in an embodiment of the present invention.

[0031] Figure 4 This is a schematic diagram illustrating the processing relationships between candidate plans, planning anchor point tables, conflict masks, deterministic safe projections, constraint conflict repair subsets, and constraint projection credentials in an embodiment of the present invention.

[0032] Figure 5 This is a schematic diagram illustrating the processing relationships between the typed instruction queue, shadow instruction queue, task consistency judgment, format compliance judgment, constraint compliance judgment, gating mismatch mask, cause matching repair, first reprojection, downgrade instructions, and feedback cross-cycle isolation in this embodiment of the invention. Detailed Implementation

[0033] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. The following examples are used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0034] This invention uses the planning cycle as the basic data processing unit, confining the large model agent to the candidate planning generation stage. Through source freezing, constraint responsibility chain, planning anchoring, conflict localization, deterministic projection, cause matching and repair, one-time reprojection, and feedback cross-cycle isolation, a unidirectional processing link is formed from task input to UAV command output.

[0035] In this invention, the planning period This represents a complete UAV mission planning and command generation cycle; mission status frame. Indicates the planning period The data consists of structured state data formed by mission text, airspace restrictions, UAV status, and environmental situation; constraint set. This represents constraint data used to define airspace, altitude, speed, time window, platform capabilities, communication reachability, and safety boundaries; planning output format constraints. This indicates the field structure, field type, field unit, field order, enumeration range, and interface format requirements that candidate plans and typed instructions must meet; mapping table. This indicates the mapping relationship between planning primitives and flight control command templates.

[0036] To avoid confusing mandatory constraints with optimizable risk metrics, the constraint set is... Divided into hard-constrained subsets and residual risk subset : , in, This represents a subset of hard constraints that an executable plan must satisfy, including no-fly zone constraints, legal altitude restrictions, maximum speed constraints for drones, minimum return-to-home battery constraints, flight dynamics boundary constraints, prohibited communication loss areas, and preset emergency safety constraints. This represents the subset of remaining risks used to evaluate security, platform capabilities, communication, and timing risks after hard constraints are met. It includes the degree of proximity to the risk zone boundary, remaining power margin, communication quality margin, time margin, and the degree of susceptibility to environmental disturbances. When a constraint has both mandatory restriction and risk assessment attributes, its mandatory restriction portion is included in the hard constraint subset. The margin assessment portion that falls within the mandatory limit is written into the residual risk subset. .

[0037] Constrained Causal Task Graph Indicates based on task status frame and constraint set Constructed graph structure; graph embedding Represents a constrained causal task graph The embedded representation formed by propagating the relationships between nodes, edges, constraint sources, and affected variables; candidate programming This represents the planning results generated by a large model agent that cannot be directly deployed for execution; planning anchor point table. Represents the causal task graph of candidate planning elements and constraints. Nodes, edges, and task status frames in the data. The correspondence between the task fields in the document.

[0038] Conflict Mask Indicates the use of labeling candidate programs Data sets of conflict-prone planning elements and semantically unanchored elements; executable planning Indicate candidate plans Planning results obtained after deterministic projection; constraint conflict repair subset This represents the set of fixes generated for constraint conflicts; format fix subset. Represents the set of fixes generated for typed instruction format errors; semantically anchored fix subset. This represents the set of anchor repair items generated in response to semantic mismatch in the task.

[0039] Typed instruction queue Indicates an executable plan The resulting structured command set that can be recognized by the flight control system; gating mismatch mask A data structure that classifies and labels semantic mismatches, format mismatches, and constraint violations; a set of first-order reprojections. Indicates based on gating mismatch mask A set of repair items selected or combined from the three repair subsets; Execution feedback. This indicates the execution status, actual flight path, platform status, anomaly information, and environmental change data returned by the UAV after executing a typed instruction queue or a downgraded instruction.

[0040] First Embodiment

[0041] like Figure 1 As shown, this embodiment provides a method for large-scale intelligent agent unmanned aerial vehicle (UAV) mission planning and command generation. The method is executed in the following order: data freezing, constraint graph construction, candidate plan generation, conflict localization, deterministic projection, command conversion, gating repair, command output, and next cycle feedback update, including steps S1 to S12.

[0042] S1. Generate and freeze the task status frame, constraint set, planning output format constraints, and mapping table.

[0043] During the planning period At the start, acquire the mission text, airspace restrictions, drone status, and environmental situation.

[0044] The task text may include task descriptions such as inspection, search, tracking, close observation, delivery, return, or collaborative operation, and may also include task object, target area, task priority, task deadline, action requirements, and task completion conditions.

[0045] The airspace restrictions may include no-fly zones, altitude-restricted zones, temporary control zones, safety buffer zones, permitted flight corridors, take-off and landing areas, and alternate landing areas. The UAV status may include current location, current altitude, current speed, heading, remaining battery power, payload status, communication link status, flight control mode, health status, and available execution capabilities. The environmental situation may include wind speed, wind direction, rainfall, visibility, obstacles, moving objects at risk, terrain changes, and communication coverage status.

[0046] The above input data is structured and mapped to generate a task status frame with source identifier. and constraint set Task status frame It should include at least the following fields: task objective field, task priority field, target area field, action intent field, platform status field, environment status field, and communication status field; constraint set. It includes at least airspace constraints, altitude constraints, speed constraints, time window constraints, platform capability constraints, communication reachability constraints, and security risk constraints.

[0047] For example, the phrase "complete the target area patrol within fifteen minutes" in the task text is mapped to the task target field, target area field, and task deadline field, respectively, and further forms a time window constraint; the remaining battery power of the drone is written into the platform status field, and at the same time participates in forming the minimum return battery power hard constraint and the remaining battery power margin risk; the communication coverage area is written into the communication status field, and participates in forming the communication loss area hard constraint and the communication quality margin risk.

[0048] The constraint set is categorized according to whether the constraints can be traded off in the executable plan. Divided into hard-constrained subsets and residual risk subset Violation of hard constraint subset The plan must not be determined as an executable plan; residual risk subset It is not used to relax hard constraints, but rather to select the plan with lower residual risk among multiple plans that all satisfy hard constraints.

[0049] Generate planning output format constraints based on UAV model, flight control interface protocol, payload type, and mission mode. and mapping table Planning output format constraints This is used to specify the fields, field types, units, order, enumeration range, validation rules, and required conditions that candidate plans and typed instructions should include; mapping table Used to define the conversion relationship between planning primitives and flight control command templates.

[0050] For example, the "fly to waypoint" planning primitive can be mapped to a flight control command template that includes target coordinates, target altitude, desired speed, arrival criteria, and backoff conditions; the "hover" planning primitive can be mapped to a flight control command template that includes hover position, hover duration, communication maintenance conditions, and abnormal backoff conditions.

[0051] Each input data item is accompanied by a source identifier. The source identifier records whether the corresponding data originated from task text parsing, airspace database, UAV telemetry, environmental sensors, communication detection, manual input, or feedback from the previous planning cycle.

[0052] After completing the above data mapping, a source freeze window is established. The source freeze window records task status frames. Constraint Sets Hard constraint subset Residual risk subset Planning output format constraints and mapping table Source identifier, timestamp, and version number.

[0053] Before the end of the current planning period, newly arriving routine environment updates or execution feedback are temporarily stored in the input cache of the next planning period, without modifying the data in the source freeze window. Therefore, the constraint-causal task graph generated within the current planning period... Image embedding Candidate planning Planning Anchor Point Table Conflict mask Executable plan Three repair subsets, gated mismatch mask Once the set is reprojected and typed instruction queues All are based on the same version of data.

[0054] When an emergency safety event is received during the current planning cycle, such as a collision warning, a serious flight control malfunction, a battery level below the emergency threshold, or entry into a preset danger boundary, a safety control channel independent of the large model agent and the first reprojection process will directly output preset safety commands for hovering, returning to home, exiting the risk area, or terminating the mission. The emergency safety event is simultaneously written into the input cache of the next planning cycle, but is not used to trigger the regeneration of candidate plans or a second reprojection within the current planning cycle.

[0055] S2. Construct a constrained causal task graph

[0056] like Figure 3 As shown, based on the task status frame and constraint set Constructing a constrained causal task graph .

[0057] Constrained Causal Task Graph The nodes include task sub-target nodes, route unit nodes, platform capability nodes, communication reachable nodes, risk object nodes, and command primitive nodes.

[0058] According to the task status frame The system generates task sub-target nodes based on the task objective field and action intent field; generates route unit nodes based on the origin, target area, destination, alternate landing point, and permitted route corridor; generates platform capability nodes based on remaining battery power, maximum speed, maximum altitude, climb capability, payload capability, and flight control mode; generates communication reachability nodes based on communication link quality, communication coverage area, and communication relay location; generates risk object nodes based on no-fly zones, obstacles, moving risk objects, severe weather areas, and terrain risks; and generates command primitive nodes based on the pre-set planning primitive set.

[0059] Planning primitives represent objects that can be mapped to tables. The minimum planned actions converted into flight control command templates include takeoff, flying to waypoints, hovering, circling, adjusting altitude, adjusting speed, maintaining communication, executing payload actions, returning to home, and awaiting manual takeover. Planning primitives are intermediate expressions between candidate plans and typed commands, not direct commands issued to the flight control system.

[0060] Constrained Causal Task Graph The edges include temporally dependent edges, resource-consuming edges, dynamically feasible edges, communicatively reachable edges, and conflict edges. Conflict edges are further classified according to conflict type into safe conflict edges, capacity-over-limit edges, communicatively unreachable edges, and temporally conflict edges.

[0061] Establish a temporal dependency edge when one task sub-objective must be executed after another task sub-objective is completed; establish a resource occupancy edge when different task sub-objectives occupy the same power, payload, sensors, communication links, or flight time windows; establish a dynamically feasible edge when the route unit meets the UAV's speed, climb rate, turning radius, and remaining power requirements; establish a communication reachable edge when the route unit or task action is within the communication coverage area; and establish a corresponding conflict edge when the route, altitude, speed, action, or time window violates hard constraints or exceeds a preset risk trigger threshold.

[0062] Each conflicting edge is written to the constraint source, affected variables, and constraint responsibility chain. The constraint responsibility chain records at least the source data identifier, constraint item identifier, the subset to which the constraint belongs, the affected variables, the associated nodes, the conflict type, and the risk level.

[0063] For example, when a candidate route crosses a no-fly zone, the constraint responsibility chain for the corresponding conflict edge records the source of the no-fly zone data, the airspace hard constraint identifier, the affected variables "route" and "altitude", the associated route unit node, and the type of safety conflict; when a candidate mission causes the expected remaining battery power after returning to base to be lower than the minimum return battery power, the constraint responsibility chain records the source of the UAV telemetry, the minimum return battery power hard constraint, the affected variables "speed", "time window" and "action", the platform capability node, and the type of capability overrun.

[0064] When a plan meets the minimum return-to-base power hard constraint, but the expected remaining power margin is small, it is not directly judged as infeasible. Instead, it is considered within the remaining risk subset based on the remaining power margin. Remaining risks to the computing platform's capabilities.

[0065] S3, Generative Graph Embedding

[0066] Constrained causal task graph Input graph embedding model. Graph embedding model can be implemented using relational graph neural networks or relational graph attention networks, used to perform relation-weighted propagation of node information according to different edge types.

[0067] Relationship scores are determined based on node feature similarity, edge type importance, constraint source credibility, and the risk level of affected variables. ; The first is determined based on the relationship score. Layer relation weights: ; in, Indicates the node index. Indicates the adjacent node index. Indicates the edge type index. This represents the layer index of a graph embedding network. Indicates the first The first in the layer The node and the first Between the nth adjacent nodes Relationship scoring for edge classes Represents the cosine similarity function. Indicates the first The importance of edge type in edge class. This indicates the credibility of the source of the corresponding constraint. This indicates the risk level of the corresponding affected variable. This represents the preset non-negative weights corresponding to the node feature similarity. This represents the preset non-negative weight corresponding to the importance of edge types. This represents the preset non-negative weight corresponding to the credibility of the constraint source. This represents the preset non-negative weight corresponding to the risk level of the affected variable, and satisfies the following conditions: , This represents the normalized relation weights. This represents the edge type index in the normalized computation. This represents the adjacent node index in the normalization calculation. Represents a set of edge types. Indicates the first The node of the first Class adjacency node set This represents the natural exponential function.

[0068] The adjacent node embedding, constraint source code, and affected variable code are concatenated, and the node embedding is updated according to the following formula: , in, Indicates the first The node at the th Layer node embedding, Indicates the first The node at the th Layer node embedding, Indicates the first The neighboring nodes at the th ... Layer node embedding, Indicates the first The node and the first Between the nth adjacent nodes The source encoding of constraints for edge classes. This indicates the corresponding affected variable code. Represents the self-node mapping matrix, Indicates the first The relational mapping matrix of edge classes, The brackets represent non-linear activation functions, and square brackets represent vector concatenation.

[0069] After updating the nodes of the preset number of layers, embed all the nodes of the last layer into a graph embedding according to the node identifier order. Simultaneously, an edge-level risk summary is generated, which includes the number of different conflict edge types, risk level, constraint source, and distribution of affected variables.

[0070] The graph embedding model is trained offline using historical task state frames, historical constraint sets, historical constraint causal task graphs, historical candidate plans, historical conflict annotations, constraint responsibility chain annotations, and historical execution results. After training, the parameters of the graph embedding model remain fixed within the actual planning cycle. In the current planning cycle, only forward inference is performed, and the model parameters are not updated in real time using the execution feedback of the current planning cycle.

[0071] S4. Generate candidate plans and a list of planning anchor points.

[0072] Embedding the graph Constraint Sets and planning output format constraints Input a large model agent.

[0073] The model parameters used are Large model intelligent agent decoding function Candidate plans are generated according to the following formula. : , in, The model parameters represent the parameters of a large-scale intelligent agent. Indicates the use of model parameters The constrained decoding function.

[0074] Large model agents are only allowed to draw from constrained causal task graphs. Candidate planning elements are selected from the set of node identifiers, edge identifiers, and pre-set planning primitives. Outputting commands that can be directly sent to the flight control system is not allowed.

[0075] Candidate Planning Each planning element includes at least a node identifier, planning primitive, task sub-objective, target route unit, time window, desired altitude, desired speed, and action intention.

[0076] In generating candidate programs Simultaneously generate planning anchor point table Planning Anchor Point Table Record the correspondence between each candidate planning element and the task sub-objective node, route unit node, platform capability node, communication reachable node, risk object node, related graph edges, task status frame field, and preset planning primitives.

[0077] When a candidate planning element cannot establish any necessary correspondence, it is marked as a semantically unanchored element. For example, a large model agent generates task state frames. If there is no additional delivery task in the plan and the task cannot be associated with a task sub-target node or task field, the planning element is determined as a semantically unanchored element. Even if the planning element does not violate the spatial domain or velocity constraints, it is not allowed to enter the typed instruction conversion process.

[0078] The large model agent executes candidate planning inference only once per planning cycle. If subsequent gating or reprojection fails, the large model agent is not returned to the current step and re-invoked.

[0079] S5, Generate a collision mask

[0080] For candidate programs The first in Planning elements According to the planning anchor point table Determine its relationship with the first Planning elements In constrained causal task graph The associated nodes or edges are read, and the first node or edge between the nodes or edges is read. Values ​​of conflicting edges in a conflicting edge class .

[0081] Will be used to indicate the Whether a planning element is a semantically anchorless element is marked by the anchorless tag. When the first When a planning element cannot establish the necessary anchoring relationship, Otherwise .

[0082] Based on the total number of candidate planning elements Conflict edge type set and indicator functions Generate conflict markers and conflict mask : , in, and Both represent planning element indices. Indicates the index of the conflict edge type. Indicates the first The planning element and the first There is a first planning element. Conflict-like edges, This indicates that the stated first... Conflict-like edges, This represents the total number of candidate planning elements. This represents the set of conflict edge types, which includes security conflict edges, capacity overrun edges, communication unreachable edges, and temporal conflict edges. This indicates an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. Indicates the first Conflict markers for planning elements, This represents a conflict mask consisting of all conflict markers.

[0083] For planning elements whose remaining risk has not yet exceeded the conflict triggering threshold, their conflict flag is not directly set to 1, but is evaluated in step S6 by the cost of remaining risk; for planning elements that violate hard constraints, whose remaining risk exceeds the conflict triggering threshold, or that lack necessary semantic anchors, their conflict flag is set to 1.

[0084] S6. Project the candidate program into an executable program.

[0085] First, determine the subset with hard constraints. and conflict mask Limited feasible planning space:

[0086]

[0087] in, This indicates a hard-constrained index. Indicates the index of the candidate planning element. This represents the basic planning space consisting of predefined planning primitives, available route units, and permitted action sequences. Represents hard-constrained subsets The index is Hard constraints Indicate planning Compared to hard constraints constraint functions, Indicate planning Satisfy the corresponding hard constraints Indicate planning Should the first one be adopted? One candidate planning element, This indicates that the candidate planning element will not be adopted. Indicates the first One candidate planning element was hit by the conflict mask.

[0088] For any feasible program Calculate its relative to the subset of residual risk Residual risk cost: , in, This represents the cost of remaining safety risks. This indicates the remaining risk and cost of the platform's capabilities. Indicates the residual risk cost of communication. This represents the cost of residual risk in the time series. , , and These represent the corresponding non-negative residual risk cost weights.

[0089] The costs of various residual risks are determined according to their corresponding risk edges: , in, Indicates the index of remaining risk categories. Indicates the first Residual risk cost Indicates belonging to the first The set of risk edges of class residual risk, Indicates the risk edge. Indicates risk edge In the Non-negative risk weights in residual risk class Indicate planning At risk Normalized risk value; when When represents the remaining safety risk, This indicates the remaining risk of platform capabilities. When the remaining risk of communication is indicated, The time indicates the remaining risk of the time series.

[0090] Within the feasible planning space, select an executable plan based on the remaining risk cost, disturbance distance, and disturbance penalty weight: , in, Indicates an executable plan. Represents the feasible planning variables. Indicates feasible planning Compared to candidate programs The disturbance distance can be determined based on a normalized weighted sum of the number of replaced planning elements, the distance of route change, the amount of time change, the amount of altitude change, the amount of speed change, and the number of action changes. This represents the non-negative perturbation penalty weight.

[0091] Through the above processing, the hard-constrained subset Determining whether a plan can enter the feasible planning space, the remaining risk subset It is only used to prioritize risks among multiple plans that meet hard constraints, avoiding the logical contradiction that "a plan has met all constraints but the cost of violating the constraints is still not zero".

[0092] After projection is complete, a constraint projection certificate is generated. The constraint projection certificate records the constraint set version, hard constraint verification results, conflict mask items that have been repaired, the remaining risk cost before and after projection, the perturbation distance, and the identifier of the final selected plan.

[0093] When feasible planning space If the value is empty, the large model agent is not called again; instead, the process proceeds directly to step S11 to output the downgrade instruction.

[0094] S7. Generate a subset for constraint conflict repair.

[0095] like Figure 4 As shown, for conflict markers The The planning element, with the first Replace the planning element with a replacement item to obtain the counterfactual program. .

[0096] Based on the difference in residual risk cost and perturbation distance between the candidate program and the counterfactual program, calculate the repair benefit density of the replacement term: , in, Indicates the index of the replacement item. Indicates the first The replacement item replaces the first one. Counterfactual planning is formed after planning elements. Indicate candidate plans The remaining risk cost, Representing counterfactual planning The remaining risk cost, This represents the distance of the counterfactual program from the candidate program. This indicates a positive number used to prevent the denominator from being zero. Indicates the first The planning element of the first Repair benefit density of each replacement item.

[0097] First, remove subsets that do not meet the hard constraints. The replacement option is then determined according to the repair benefit density. Select the remaining replacement items from high to low to form a constraint conflict repair subset. .

[0098] When multiple replacement items have the same repair benefit density, the unique first replacement item is selected in the following order: larger hard constraint safety margin, higher constraint source credibility, smaller perturbation distance, and smaller replacement item index, to avoid uncertain repair order caused by the same input.

[0099] The replacement options may include adjusting the flight path, adjusting the time window, adjusting the altitude, adjusting the speed, replacing actions, or adding communication maintenance actions. A repair responsibility label is written for each selected replacement option, which includes the corresponding conflict edge type, constraint source, affected variables, and repair action.

[0100] When conflict marker When this happens, constraint conflict replacement items are not generated for the corresponding planning elements to avoid unnecessary modifications to non-conflicting planning elements.

[0101] S8. Generate a typed instruction queue, a shadow instruction queue, and a format repair subset.

[0102] Executable plan Constrained by the planned output format and mapping table Convert to a typed instruction queue .

[0103] For executable plans Middle and the first Typed instructions Corresponding planning segment Generate the first according to the following formula Typed instructions are grouped into a typed instruction queue: , in, Indicates the index of the typed instruction. Indicates the relationship with the first The planning segment corresponding to each typed instruction, This represents the conversion function from planning primitives to flight control command templates. Indicates the first A typed instruction, Indicates the planning period The total number of typed instructions within, This represents a queue of typed instructions consisting of all typed instructions in the order of execution.

[0104] Each typed instruction must include at least the instruction type, the target of execution, spatial parameters, time parameters, speed parameters, communication parameters, and backoff conditions.

[0105] In generating typed instruction queues Simultaneously, a shadow instruction queue is generated. The shadow instruction queue is a typed instruction queue. The simulated copy is not sent to the actual flight control system, but is instead pre-run in the simulation environment or rule executor. Shadow verification is used to determine whether route violations, altitude exceeding limits, speed exceeding limits, communication failures, incorrect action sequences, or missing backoff conditions occur after command execution.

[0106] Simultaneously, an instruction template fingerprint is generated. The instruction template fingerprint records the planning primitive, flight control instruction template version, mapping table version, field integrity verification result, and field unit verification result corresponding to each typed instruction.

[0107] When the format compliance check fails, the output format constraints of the plan will apply. Mapping table and instruction template fingerprint generation format repair subset Formatting Repair Subset Used to complete missing fields, standardize field units, correct field types, correct field enumeration values, adjust field order, add rollback conditions, or replace with the correct template corresponding to the same planning primitive, without changing the executable plan. The mission objectives, flight route intentions, and targets of execution.

[0108] S9. Perform three types of gating judgments.

[0109] like Figure 5 As shown, in the typed instruction queue Before issuing the document, perform a task consistency check, a format compliance check, and a constraint compliance check.

[0110] Embedded based on task state frame consistency Consistent embedding with typed instruction queues Based on task consistency similarity, task consistency threshold, format error rate, format error threshold, constraint violation rate, and constraint violation threshold, a gating mismatch mask is generated. , , , , in, Indicates a gate mismatch mask. This represents the task state frame consistency embedding function. This represents a typed instruction queue consistency embedding function. This represents a task consistency similarity function. This represents the task consistency threshold. Functions indicating formatting errors Indicates the threshold for formatting errors. Represents the constraint violation function. Indicates the threshold for constraint violation. Indicates semantic mismatch markers, Indicates a format mismatch flag. Indicates a constraint violation flag.

[0111] Constraint violation degree for hard constraint subset Perform calculations; residual risk subset The risk ranking among feasible plans has been used in deterministic projections, and a residual risk cost greater than zero is not directly equated to a violation of hard constraints.

[0112] When the shadow command queue experiences route violations, altitude overruns, speed overruns, or communication failures, a violation flag will be set. Set to 1; when a field is missing, a unit error, a field type error, an enumeration value error, or a fallback condition is missing, a format mismatch flag will be displayed. Set to 1; when there is a missing task sub-goal, an incorrect execution object, an error in the planning primitive, or an action sequence that is inconsistent with the task intent, a semantic mismatch will be marked. Set to 1. When the same failure belongs to multiple types, multiple mismatch flags can be set to 1 simultaneously.

[0113] When all three mismatch flags are 0 and the shadow instruction queue verification passes, the typed instruction queue is issued. .

[0114] When semantic mismatch tags At that time, based on the task status frame Candidate planning And planning anchor point table Generate semantic anchoring repair subset Semantic anchoring repair subset Used to restore the correspondence between task sub-goals, execution objects, planning primitives, or action sequences and task fields; does not allow the generation of task status frames. New mission objectives that do not exist in the game.

[0115] The consistency verification model is trained offline using task state frames, candidate plans, planning anchor point tables, constraint sets, planning output format constraints, mapping tables, and typed instruction queues. During the actual planning cycle, the consistency verification model only performs inference and does not utilize execution feedback from the current planning cycle to update model parameters.

[0116] S10. Form a reprojection set and perform a reprojection.

[0117] When gating mismatch mask When at least one mismatch flag is 1, the repair item is used. For each set element, construct a reprojected set based on the three mismatch markers: , in, This represents the set of reprojections. Indicates the repair item. This represents a subset for constraint conflict repair. This indicates a subset of format repairs. This indicates a semantically anchored repair subset.

[0118] When multiple mismatch flags are simultaneously set to 1, the repair process is performed in the following fixed order:

[0119] First, apply the corrective measures corresponding to the constraint violations to the executable plan. and in the hard-constrained subset First, the plan is redefined within the limited feasible planning space. Second, the repair items corresponding to semantic mismatch are applied to the planning anchoring relationship to restore the correspondence between task sub-goals, execution objects, planning primitives, action order and task fields. Third, the typed instruction queue is regenerated based on the repaired executable plan and anchoring relationship. Finally, the repair items corresponding to format mismatch are applied to the regenerated typed instruction queue to correct fields, units, types, enumeration values, order and fallback conditions.

[0120] The above fixed order can prevent the format repair completed earlier from being overwritten by the subsequent instruction regeneration process, and can also prevent semantic repair from acting on plans that have not yet eliminated hard constraint violations.

[0121] During a single reprojection, no large model agent is invoked, no new task objectives are added, no new inputs outside the current planning cycle are read, and only the task state frames in the source freeze window are used. Constraint Sets Planning output format constraints Mapping table And planning anchor point table .

[0122] After the reprojection is completed, the task consistency judgment, format compliance judgment, constraint compliance judgment and shadow instruction verification are re-executed according to the same judgment conditions in step S9.

[0123] If the reprojected typed instruction queue passes all the checks, then the reprojected typed instruction queue is issued; if the reprojection set fails... If the space is empty, the feasible planning space corresponding to the reprojection is empty, or the reprojection still fails the judgment, then proceed directly to step S11.

[0124] The current planning cycle sets a reprojection status flag. Before a reprojection begins, the status flag is 0; immediately after the reprojection begins, it is set to 1. When the status flag is 1, re-entering the reprojection process is not allowed. Therefore, the current planning cycle only has three termination paths: initial gating pass, gating pass after a reprojection, or outputting a degradation instruction. There are no data paths that return to the large model agent, involve a second reprojection, result in an empty loop, or lead to an infinite loop.

[0125] S11, Output downgrade command

[0126] When the feasible planning space is empty, a single reprojection set is used. If the result is empty or fails the gating judgment after one reprojection, output a degradation instruction and terminate the current planning cycle.

[0127] Degradation commands can include hovering, returning to base, maintaining communication, exiting a risky area, entering a safe altitude, waiting for manual takeover, or performing pre-defined safety actions.

[0128] The system pre-generates a degraded trajectory cache. This cache stores available safe hovering points, return routes, safe altitudes, communication hold-up routes, and manual takeover waiting areas for the current planning period. The degraded trajectory cache does not participate in candidate planning. The generation does not modify the typed instruction queue. Pre-defined security action data is only provided when a downgrade is triggered.

[0129] After the downgrade command is output, the current planning cycle will no longer perform candidate planning generation, deterministic projection, or reprojection; it will only receive execution feedback returned by the UAV. .

[0130] S12. Receive execution feedback and implement cross-cycle isolation.

[0131] After executing the initial typed instruction queue, the reprojected typed instruction queue, or the downgraded instructions, the UAV receives execution feedback. .

[0132] Execution feedback It can include command execution status, actual flight path, current position, actual altitude, actual speed, remaining battery power, communication link quality, payload status, abnormal alarms, and environmental change data.

[0133] Execution feedback Do not write back to the typed instruction queue of the current planning cycle It also does not modify the task state frame of the current planning cycle. Constraint Sets Planning output format constraints Mapping table Candidate planning Or an executable plan .

[0134] Execution feedback It is written to the input cache for the next planning period and in the next planning period. At the start, together with the new task text, airspace restrictions, UAV status, and environmental situation, a new task state frame and constraint set are formed, which are used to construct the constraint causal task graph for the next planning cycle. .

[0135] Execution feedback can be categorized into completion feedback, deviation feedback, anomaly feedback, and safety feedback. Completion feedback indicates that the command was completed as expected; deviation feedback indicates that the actual flight path deviates from the expected path but has not yet triggered a safety risk; anomaly feedback indicates abnormalities in communication, payload, battery power, or environmental conditions; and safety feedback indicates that the UAV has entered a risk boundary or triggered a degrade condition. The feedback classification results are only used for the next planning cycle and do not modify commands already generated or issued in the current planning cycle.

[0136] The data processing chain for the current planning cycle is as follows: ,

[0137] This data processing chain proceeds in one direction. Execution feedback. There is no way to return the current typed instruction queue. The data path; there is no data path to return to the large model agent after gating failure; there is no data path to enter the reprojection process again after one reprojection is completed.

[0138] Exemplary task processing procedure

[0139] Taking the target area inspection task as an example, the task status frame The mission objective recorded in the data is "to complete the inspection of the target area within fifteen minutes". The drone's current altitude is 90 meters, the remaining battery power is 48%, and the current communication link quality is normal.

[0140] Hard constraint subset This includes not entering no-fly zones, maintaining a flight altitude between 80 and 120 meters, a flight speed not exceeding 18 meters per second, and expecting to have at least 25% battery remaining after returning to base; remaining risk subset. This includes the degree of approach to the no-fly zone boundary, the expected return battery capacity, the communication quality capacity, and the mission completion time capacity.

[0141] Candidate programs generated by large model agents This includes planning elements such as takeoff, flight to the first inspection point, inspection execution, flight to the second inspection point, and return. (Planning Anchor Point Table) Each planning element is associated with a task sub-target node, a route unit node, a task field, and a planning primitive.

[0142] When the first inspection segment crosses a no-fly zone, the corresponding safety conflict edge is set to 1, and the conflict mask sets the conflict marker for that segment to 1. The deterministic projection process excludes this segment. Among the bypass segments that meet the hard constraints of altitude, speed, and minimum return-home charge, an executable plan is selected based on the remaining safety risk, remaining communication risk, remaining timing risk, and the disturbance distance relative to the candidate plans. .

[0143] If the height unit is incorrectly written as feet during type conversion, the format compliance check will set the format mismatch flag to 1, and the format repair subset will be adjusted. This includes a fix to standardize the altitude unit to meters. Each reprojection only corrects the field unit; it does not change the inspection target, inspection route, or execution object. The corrected shadow command queue is then issued after verification.

[0144] If the revised instructions still violate hard constraints, or a reprojection set cannot be formed, then a return or hover instruction is output and the current planning cycle is terminated. The execution result is only written to the input buffer of the next planning cycle.

[0145] Second Embodiment

[0146] like Figure 2 As shown, this embodiment provides a large-scale intelligent agent UAV mission planning and command generation system for executing the method described in the first embodiment. The system includes a data graph construction module, a planning projection module, a command gating module, and a feedback isolation module.

[0147] The data graph construction module includes an input mapping unit, a source freezing unit, a constrained causal task graph construction unit, and a graph embedding reasoning unit.

[0148] The input mapping unit is used to acquire mission text, airspace restrictions, UAV status, and environmental situation, and generate a mission status frame with source identifier. Constraint Sets Planning output format constraints and mapping table and the constraint set Divided into hard-constrained subsets and residual risk subset .

[0149] The source freeze unit is used to record task status frames. Constraint Sets Hard constraint subset Residual risk subset Planning output format constraints and mapping table The source identifier, timestamp, and version number are used to temporarily store new general environment updates or execution feedback arriving in the current planning cycle into the input cache of the next planning cycle.

[0150] The source freezing unit is also connected to an independent safety control channel. When a preset emergency safety event is detected, the independent safety control channel directly outputs a preset safety command without invoking the candidate planning generation unit or the first reprojection unit.

[0151] The constrained causal task graph construction unit is used to generate task sub-objective nodes, route unit nodes, platform capability nodes, communication reachable nodes, risk object nodes, and instruction primitive nodes. It generates temporal dependency edges, resource occupancy edges, dynamic feasible edges, communication reachable edges, and conflict edges, and writes the constraint source, affected variables, and constraint responsibility chain list into the conflict edges.

[0152] The graph embedding inference unit is used to embed the constrained causal task graph. Input the trained graph embedding model and generate graph embeddings. And edge-level risk summary.

[0153] The planning projection module includes a candidate planning generation unit, a planning anchor point generation unit, a conflict mask generation unit, and a safe projection unit.

[0154] Candidate planning generation units are used to embed graphs Constraint Sets and planning output format constraints Input a large model agent and generate only candidate plans. It does not generate flight control commands that can be executed directly.

[0155] The planning anchor point generation unit is used to generate the planning anchor point table. Record candidate planning elements and constraint causal task graph Middle nodes, edges, and task status frames The mapping relationship between task fields is determined, and planning elements that cannot establish the necessary mapping relationship are marked as semantically unanchored elements.

[0156] The conflict mask generation unit is used to generate a table of planned anchor points. Read the conflict edges associated with each candidate planning element and generate a conflict mask by combining semantically anchorless elements. .

[0157] The secure projection unit is used in hard-constrained subsets and conflict mask Within the limited feasible planning space, based on the remaining risk subset Candidate Programs Projection as an executable plan The system generates constraint projection credentials. When the feasible planning space is empty, the safe projection unit sends a degradation trigger signal to the degradation output unit and does not return to the candidate planning generation unit.

[0158] The instruction gating module includes a constraint repair unit, an instruction conversion unit, a shadow verification unit, a format repair unit, a semantic anchoring repair unit, a gating mismatch identification unit, a first reprojection unit, and a degradation output unit.

[0159] Constraint repair units are used for collision masks The matched planning elements generate replacement items, and the subsets that do not meet the hard constraints are removed. The replacement terms are used to calculate the repair benefit density of the remaining replacement terms, forming a subset for constraint conflict repair. And write a repair responsibility label for the repair item.

[0160] The instruction conversion unit is used to constrain the planned output format. and mapping table , will be an executable plan Convert to a typed instruction queue And generate instruction template fingerprints.

[0161] The shadow verification unit is used to generate and type instruction queues. The shadow command queue corresponds to but is not sent to the flight control system, and the typed commands are pre-rendered in the simulation environment or rule executor.

[0162] The format repair unit is used to correct format mismatches according to the planned output format constraints. Mapping table and instruction template fingerprint generation format repair subset .

[0163] The semantic anchoring repair unit is used to, when semantic mismatch occurs, adjust the anchoring based on the task state frame. Candidate planning And planning anchor point table Generate semantic anchoring repair subset .

[0164] The gating mismatch identification unit is used to perform task consistency, format compliance, and constraint compliance judgments, and generates a gating mismatch mask that includes semantic mismatch markers, format mismatch markers, and constraint violation markers. .

[0165] The reprojection unit is used based on the gated mismatch mask. Repairing subsets from constraint conflicts Format Repair Subset and semantic anchoring repair subset Select the repair items to form a first reprojection set. .

[0166] The reprojection unit performs only one reprojection without calling a large model agent, adding new task objectives, or reading new inputs outside the current planning cycle. After completing one reprojection, the gating mismatch identification unit performs the same judgment again.

[0167] If the reprojected instruction passes the judgment, then the reprojected instruction is issued; if the reprojection set fails... If the feasible planning space is empty, or the reprojected instructions still fail the judgment, the degradation output unit outputs a degradation instruction and terminates the current planning cycle.

[0168] The instruction gating module does not set up an interface for the reprojection unit to return to the candidate planning generation unit, nor does it set up an interface for the reprojection unit to return to itself for re-execution, thus blocking recursive calls to large model agents and repeated reprojections from the system structure.

[0169] The feedback isolation module includes a feedback receiving unit, a feedback hierarchical unit, and a next cycle write unit.

[0170] The feedback receiving unit is used to receive execution feedback generated by the UAV executing the initial typed instruction queue, the reprojected instructions, or the downgraded instructions. .

[0171] The feedback hierarchy unit is used to classify execution feedback. It is divided into completion feedback, deviation feedback, anomaly feedback and safety feedback.

[0172] The next cycle write unit is used to write execution feedback. The feedback and hierarchical results are written into the input cache of the next planning cycle, enabling them to participate in the formation of the task state frame and constraint set, as well as the constraint causal task graph of the next planning cycle. The construction.

[0173] The feedback isolation module does not have the ability to write to the current typed instruction queue. Current candidate plans Or the currently executable plan The data interface, and therefore execute feedback. The planning results and instruction generation results of the current planning cycle cannot be modified in reverse.

[0174] Model training and inference deployment

[0175] The system involves models including graph embedding models, large model agents, and consistency verification models.

[0176] The graph embedding model is trained offline based on historical task state frames, historical constraint sets, historical constraint causal task graphs, historical candidate plans, conflict annotations, chain of responsibility annotations, and historical execution results.

[0177] The large model agent performs instruction fine-tuning or parameter adaptation based on historical task text, task state frames, constraint sets, constraint causal task graphs, graph embeddings, planning anchor point tables, and candidate plans confirmed by humans.

[0178] The consistency verification model is trained offline based on the task state frame, candidate plan, planning anchor table, typed instruction queue, constraint set, planning output format constraints and mapping table.

[0179] The model training and actual planning inference processes are separated. During the actual planning cycle, the graph embedding model only performs graph embedding inference, the large model agent performs candidate planning inference only once, and the consistency verification model only performs consistency and gating inference. Execution feedback for the current planning cycle is provided. No model parameters are updated immediately.

[0180] Third Embodiment

[0181] This embodiment provides a computer program product. The computer program product includes computer program instructions, which, when executed on a processor, cause the processor to perform the method described in the first embodiment.

[0182] The computer program product may be stored in a non-transitory computer-readable storage medium. The computer-readable storage medium may include a read-only memory, random access memory, solid-state memory, disk storage, optical disk storage, or other non-transitory storage medium capable of storing computer program instructions.

[0183] The computer program instructions include at least: Used to generate and freeze task status frames Constraint Sets Hard constraint subset Residual risk subset Planning output format constraints and mapping table Program instructions; Used to construct a constrained causal task graph Write the constraint source, affected variables, and constraint responsibility chain list and generate a graph embedding. Program instructions; Used to invoke a large model agent to generate candidate plans. And planning anchor point table Program instructions; Used according to the planning anchor point table and constrained causal task graph Generate conflict mask Program instructions; Used in hard-constrained subsets and conflict mask Within the limited feasible planning space, based on the remaining risk subset Generate executable plan Program instructions; Used to generate a subset of constraint conflict repairs. Format Repair Subset and semantic anchoring repair subset Program instructions; Used to execute plans Convert to a typed instruction queue And generate program instructions that include a shadow instruction queue and instruction template fingerprint; Used to perform task consistency, format compliance, and constraint compliance checks, and generate gating mismatch masks. Program instructions; Used for gating mismatch mask Form a first reprojection set And execute the reprojection program instruction only once without calling a large model agent or adding new task objectives; This is a program instruction used to output a degradation instruction and terminate the current planning cycle when the decision is not passed after a reprojection, the reprojection set is empty, or the feasible planning space is empty. Used to receive execution feedback and will execute feedback Write only to the input buffer for the next planning cycle without writing back to the current typed instruction queue. Program instructions; This is a program instruction used to output a preset security command through an independent security control channel when a preset emergency security event is detected, and to prevent the emergency security event from triggering the regeneration or second reprojection of candidate plans within the current planning cycle.

[0184] In one deployment method, the computer program product runs on the UAV ground control equipment. The ground control equipment receives mission text, UAV telemetry and environmental data, completes candidate planning generation, deterministic projection, gating and typed instruction generation, and sends the gated typed instructions to the UAV.

[0185] In another deployment method, data graph construction, model inference, planning projection, command gating, and feedback isolation processes run on servers, ground stations, and UAV-borne computing devices, respectively. Data is transmitted between devices according to the version number recorded in the source freeze window. Before processing, the receiving end verifies the planning cycle identifier, timestamp, and version number of the input data; if the verification is inconsistent, the current processing step is not continued, but a version inconsistency flag is output and a degradation process is initiated.

[0186] The reprojection procedure is configured to execute at most once within the current planning period. After a reprojection is completed, the program flow only enters the regating decision stage; it cannot return to the candidate planning generation procedure, nor can it re-enter the reprojection procedure again.

[0187] The above implementation methods form a complete technical chain around the safety constraints of large model candidate planning, traceable anchoring of planning elements, separation of hard constraints and residual risks, failure cause matching and repair, at most one reprojection, and cross-cycle isolation of execution feedback. Equivalent substitutions can be made to the data acquisition methods, model structures, graph embedding layers, projection solution methods, flight control command template formats, storage media types, or equipment deployment methods in the above implementation methods, as long as the aforementioned data coupling relationships and unidirectional processing logic are still adopted, all of which fall under the implementation forms of the technical concept of this invention.

Claims

1. A method for mission planning and instruction generation for large-scale intelligent agent unmanned aerial vehicles (UAVs), characterized in that, include: During the planning period At the start, freeze the task status frame with the source identifier. Constraint Sets Planning output format constraints And a mapping table from planning primitives to flight control command templates timestamp and version number, and constrain the set Divided into hard-constrained subsets and residual risk subset ; Based on the task status frame and constraint set Constructing a constrained causal task graph and image embedding And write the constraint source, affected variables, and constraint responsibility chain list into the conflict edges; embed the graph. Constraint Sets and planning output format constraints Input a large model agent and generate only candidate plans. And planning anchor point table From the planning anchor point table Associate candidate planning elements with the constraint causal task graph Nodes, edges, and task status frames Task fields; based on the planning anchor table Reading conflict edges to form a conflict mask And write semantically unanchored elements that cannot be associated into the conflict mask. In hard-constrained subsets and conflict mask Within the limited feasible planning space, based on the remaining risk subset Candidate Programs Projection as an executable plan and according to the conflict mask Generate constraint conflict repair subset Based on the constraints of the planned output format and mapping table Executable plan Convert to a typed instruction queue It performs checks on task consistency, format compliance, and constraint compliance; if a check fails, it generates a gating mismatch mask. According to the planning output format constraints Mapping table and instruction template fingerprint generation format repair subset Based on the task status frame Candidate planning And planning anchor point table Generate semantic anchoring repair subset According to the gate mismatch mask Repairing subsets from constraint conflicts Format Repair Subset and semantic anchoring repair subset Select the repair items to form a first reprojection set. Without invoking a large model agent or adding new task objectives, perform only one reprojection; if the reprojection still fails the judgment, perform another reprojection. If the feasible planning space is empty or empty, output a degradation instruction and terminate the current planning cycle; execute feedback. Only write to the input buffer for the next planning period, do not write back to the typed instruction queue for the current planning period. .

2. The method for large-scale intelligent agent UAV mission planning and instruction generation according to claim 1, characterized in that, The frozen task status frame with source identifier Constraint Sets Planning output format constraints and mapping table The timestamp and version number include: establishing a source freeze window at the start of the current planning cycle and recording the task status frame. Constraint Sets Planning output format constraints and mapping table The source identifier, timestamp, and version number; environmental updates or execution feedback arriving before the end of the current planning period are temporarily stored for the next planning period, without using the environmental updates or execution feedback to rewrite the source freeze window, thus ensuring the constraint causal task graph within the current planning period is preserved. Image embedding Candidate planning Planning Anchor Point Table Conflict mask Executable plan Constraint Conflict Repair Subset Format Repair Subset Semantic anchoring repair subset Gating mismatch mask Once the set is projected again and typed instruction queues All use the same version of data corresponding to the source freeze window; when a preset emergency safety event is received in the current planning cycle, the preset safety command for hovering, returning or terminating the task is directly output by the safety control channel independent of the large model agent and the first reprojection process, and the preset emergency safety event is recorded as the input for the next planning cycle. The preset emergency safety event is not used to trigger the regeneration of candidate plans or the second reprojection in the current planning cycle.

3. The method for large-scale intelligent agent UAV mission planning and instruction generation according to claim 1, characterized in that, Based on the planning anchor point table Reading conflict edges to form a conflict mask This includes: for the candidate plans The first in Planning elements According to the aforementioned planning anchor point table Determine its relationship with the first Planning elements In the constrained causal task graph The associated nodes or edges in the middle, read the first node or edge between the nodes or edges. Values ​​of conflicting edges in a conflicting edge class ; will be used to indicate the first Planning elements Whether an element is a semantically anchorless element is marked as an anchorless tag. When the planning anchor point table Unable to provide the first Planning elements Establishing the mapping between nodes, edges, and task fields is required. Otherwise According to the candidate plan Total number of planning elements Conflict edge type set and indicator functions Generate the first according to the following formula Planning elements Conflict markers and the conflict mask : ， in, and All are planning element indexes. Index for conflict edge type, Indicates the first The planning element and the first There is a first planning element. Conflict-like edges, This indicates that the stated first... Conflicting edges, the set of conflict edge types Including security conflict edges, capacity overrun edges, communication unreachable edges, and timing conflict edges, the indicator function The value is 1 if the condition inside the parentheses is true, and 0 otherwise.

4. The method for large-scale intelligent agent UAV mission planning and instruction generation according to claim 1, characterized in that, The generation constraint causal task graph Graph embedding , including: for the first The node at the th Layer node embedding In the set of edge types Inside, targeting the first The node of the first Adjacency Node Set The first in The nth adjacent node will The neighboring nodes at the th ... Layer node embedding , No. The node and the first Between the nth adjacent nodes Constraint source encoding of edge class and the encoding of affected variables Perform splicing and update according to the following formula to obtain the first... The node at the th Layer node embedding : ， in, Indicates the node index. Indicates the adjacent node index. This represents the layer index of a graph embedding network. Indicates the edge type index. Represents a set of edge types. Indicates the first The node of the first Class adjacency node set Indicates the first The first in the layer The node and the first Between the nth adjacent nodes Relationship weights of edge classes Represents the self-node mapping matrix, Indicates the first The relational mapping matrix of edge classes, The square brackets represent vector concatenation, indicating a nonlinear activation function. The relation weights are determined based on node feature similarity, edge type importance, constraint source reliability, and the risk level of affected variables. The graph embedding... It consists of all the embedded nodes in the last layer.

5. The method for large-scale intelligent agent UAV mission planning and instruction generation according to claim 1, characterized in that, The only candidate plan generation And planning anchor point table This includes: using model parameters as Large model intelligent agent decoding function Embedded according to the following formula based on the diagram Constraint Sets and planning output format constraints Generate the candidate plan : ， in, These are the model parameters of the large model agent. To use the model parameters The decoding function; the large model agent is only allowed to decode from the constrained causal task graph. Candidate planning elements are selected from the node identifiers, edge identifiers, and pre-set planning primitives, and no directly executable flight control commands are output; the planning anchor point table Record the node identifier, edge identifier, and task status frame corresponding to each candidate planning element respectively. In the task field, candidate planning elements that cannot establish any of the aforementioned correspondences are identified as semantically unanchored elements, and are then processed through the conflict mask. Prevent it from entering the typed instruction conversion process.

6. The method for large-scale intelligent agent UAV mission planning and instruction generation according to claim 3, characterized in that, The hard-constrained subset and conflict mask Within the limited feasible planning space, based on the remaining risk subset Candidate Programs Projection as an executable plan This includes: satisfying the subset of hard constraints. And without using the aforementioned collision mask The set of planning elements with a conflict marker of 1 in the middle is determined as the feasible planning space. For any feasible program According to the feasible plan Relative to the constrained causal task graph and residual risk subset Residual risk cost The feasible planning Relative to the candidate plan Distance of disturbance and perturbation penalty weights Select the executable plan according to the following formula. : ， Based on the cost of residual safety risk Residual risks and costs of platform capabilities Residual risk cost of communication Time-series residual risk cost and their corresponding non-negative weights , , and The remaining risk cost is determined according to the following formula. : ， in, For feasible planning variables, For feasible planning space, For only the remaining risk subset Calculated residual risk cost, The disturbance distance, The perturbation penalty weight is non-negative. , , and These represent the remaining risk costs in terms of security, platform capabilities, communication, and timing, respectively. , , and These represent the corresponding residual risk cost weights. This indicates selecting the plan that minimizes the objective function within the feasible planning space; generating constraint projection credentials that record the constraint set version, hard constraint verification results, repaired conflict mask terms, remaining risk cost changes, and perturbation distances; when the feasible planning space When empty, the large model agent is not re-invoked, and a degradation instruction is directly output.

7. The method for large-scale intelligent agent UAV mission planning and instruction generation according to claim 6, characterized in that, The collision mask Generate constraint conflict repair subset This includes: for conflict markers The The planning element, with the first The replacement item replaces the first... Each planning element yields a counterfactual program. According to the candidate plan Relative to the constrained causal task graph and residual risk subset Residual risk cost The aforementioned counterfactual programming Relative to the constrained causal task graph and residual risk subset Residual risk cost The aforementioned counterfactual programming Relative to the candidate plan Distance of disturbance And prevent zero positive numbers Calculate the first according to the following formula The planning element of the first Repair benefit density of each replacement item : ， in, For the replacement item index, For the first The replacement item replaces the first one. Counterfactual planning is formed after planning elements. A positive number is set to prevent the denominator from being zero. The repair benefit density of the corresponding replacement item; removing subsets that do not meet the hard constraints. The replacement item, and according to the repair benefit density. The remaining replacement items, selected from high to low, constitute the constraint conflict repair subset. The replacement items include adjusting the route, adjusting the time window, adjusting the altitude, adjusting the speed, replacing the action, or adding a communication maintenance action; a repair responsibility label is written for each selected replacement item, and the repair responsibility label includes the corresponding conflict edge type, constraint source, affected variable, and repair action.

8. The method for large-scale intelligent agent UAV mission planning and instruction generation according to claim 1, characterized in that, The executable plan Convert to a typed instruction queue And generate a format repair subset. This includes: for the executable plan Middle and the first Typed instructions Corresponding planning segment A conversion function from planning primitives to flight control command templates is used. According to the planning output format constraints and mapping table Generate the first according to the following formula Typed instructions And based on the total number of typed instructions The typed instruction queue is composed of : ， in, For typed instruction indexes, In order to be with the first The planning segment corresponding to each typed instruction, To plan the conversion function from primitives to flight control command templates, The total number of typed instructions; each typed instruction includes instruction type, execution object, spatial parameters, time parameters, speed parameters, communication parameters, and rollback conditions; generate a queue of typed instructions. A shadow command queue, corresponding to but not sent to the flight control system, is generated, and a command template fingerprint is created that records the planning primitives, flight control command template version, and field integrity verification results. If the format compliance check fails, the output format constraints are applied according to the planning parameters. Mapping table The format repair subset is generated by generating the instruction template fingerprint. The format repair subset Used to complete fields, unify field units, correct field types, field enumeration values ​​or field order, and add / revert conditions, without changing the executable plan. The mission objectives, flight route intentions, and targets of execution.

9. The method for large-scale intelligent agent UAV mission planning and instruction generation according to claim 8, characterized in that, The typed instruction queue The system performs checks on task consistency, format compliance, and constraint compliance, and applies a gating mismatch mask. Form a first reprojection set This includes: embedding based on task state frame consistency. Consistent embedding with typed instruction queues Task consistency similarity between Task consistency threshold The typed instruction queue Relative to the planning output format constraints Formatting error Formatting error threshold The typed instruction queue Relative to the hard constraint subset Constraint violation degree Constraint violation threshold and indicator functions Generate semantic mismatch tags according to the following formula. Format mismatch marker Restricting violation markings and the gated mismatch mask : ， ， ， ， in, Embed a function for consistency of task state frames. Embedded functions for consistency of typed instruction queues. For task consistency similarity function, This is the task consistency threshold. For format error rate function, For formatting errors, To constrain the violation degree function, To constrain the violation threshold, , and These indicate whether semantic mismatch, format mismatch, and constraint violation have occurred, respectively. When the shadow command queue experiences route out-of-bounds errors, altitude exceeding limits, communication hold failures, incorrect action sequence, or missing rollback conditions, at least one mismatch flag is set to 1 based on the failure type. When all three mismatch flags are 0 and the shadow command queue verification passes, the typed command queue is issued. When the semantic mismatch marker At that time, based on the task status frame Candidate planning And planning anchor point table Generate the semantic anchoring repair subset This is used to restore the correspondence between task sub-goals, execution objects, planning primitives, or action sequences and task fields; to repair items. For set elements, the constraint conflict-repaired subset is determined according to the following formula based on three mismatch markers. Format Repair Subset and semantic anchoring repair subset The first reprojection set is composed of the middle : ， in, Indicates the repair item, the indicator function The value is 1 if the condition within the parentheses is true, and 0 otherwise; the correction item corresponding to the constraint violation is applied to the executable plan. The repair items corresponding to semantic mismatches will be determined based on the planned anchor point table. Acting on the executable plan With typed instruction queues The task anchoring relationship between them; the executable plan after completing constraint repair and semantic anchoring repair. and the task anchoring relationship regenerates the typed instruction queue. Apply the format mismatch correction to the regenerated typed instruction queue. After a reprojection, the same judgment conditions are applied again. If the reprojection passes, a typed instruction queue is issued; otherwise, the reprojection fails or the set of reprojections is re-evaluated. If the value is empty, a downgrade instruction is output, and the process of second reprojection or candidate program regeneration is not initiated.

10. A large-scale intelligent agent unmanned aerial vehicle (UAV) mission planning and instruction generation system, characterized in that, include: The data graph building module is used during the planning cycle. Freeze the task status frame with source identifier Constraint Sets Planning output format constraints and mapping table The timestamp and version number will constrain the set. Divided into hard-constrained subsets and residual risk subset Construct a constrained causal task graph and image embedding The system writes the constraint source, affected variables, and constraint responsibility chain list into the conflict edges; a planning projection module is used to embed the graph. Constraint Sets and planning output format constraints Input a large model agent and generate only candidate plans. And planning anchor point table From the planning anchor point table Associative candidate planning elements and constraint causal task graph Nodes, edges, and task status frames The task field reads conflict edges and combines them with semantically anchorless elements to form a conflict mask. In hard-constrained subsets and conflict mask Within the limited feasible planning space, based on the aforementioned subset of remaining risks The candidate plans Projection as an executable plan ; Command gating module, used to control the conflict mask. Generate constraint conflict repair subset The executable plan Convert to a typed instruction queue Generate a format repair subset and semantic anchoring repair subset It performs checks on task consistency, format compliance, and constraint compliance; if any check fails, it generates a gating mismatch mask. Based on this, a reprojection set is formed. Without invoking the large model agent or adding new task objectives, only one reprojection is performed. If the reprojection still fails the judgment, a reprojection set is obtained. If the feasible planning space is empty or empty, output a downgrade instruction and terminate the current planning cycle; The feedback isolation module is used to isolate execution feedback. Only write to the input buffer for the next planning period, do not write back to the typed instruction queue for the current planning period. .

Images