Unmanned aerial vehicle group task planning method based on large language model

Abstract

The invention discloses an unmanned aerial vehicle group task planning method based on a large language model, and the method comprises the steps: collecting and preprocessing the multi-level input information of an unmanned aerial vehicle group, transmitting the multi-level input information to a large language model in a specific coding format, and enabling the large language model to understand a natural language instruction, generating a high-level task target; performing recursive decomposition on the high-level task target by using a traditional task planner to form a task tree, and allocating sub-tasks after task decomposition to the unmanned aerial vehicle according to the capability of the unmanned aerial vehicle; and constructing an unmanned aerial vehicle action function library, understanding a task planning result generated by the traditional task planner by using a large language model, matching and calling related action functions in the unmanned aerial vehicle action function library, generating an action sequence which can be directly executed by the unmanned aerial vehicle, and checking and verifying the action sequence. According to the invention, the task planning capability of the unmanned aerial vehicle group in a dynamic environment is improved, and efficient task distribution and path planning are realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicle (UAV) mission planning, and in particular relates to a UAV swarm mission planning method based on a large language model. Background Art

[0002] Drones (UAVs) have been widely used in various fields, including regional exploration, environmental monitoring, and logistics. Compared to single UAVs, multi-UAV collaboration can significantly improve mission efficiency by parallelizing task execution. Mission planning is one of the most important issues in multi-UAV operations, directly impacting mission efficiency. Traditional mission planning methods typically rely on predefined rule systems to address different mission scenarios, including algorithms such as auctions and negotiations. These methods can assign tasks and plan paths for UAV swarm systems through structured logical rules, providing a basic framework for multi-UAV collaboration. However, with the increasing complexity and diversity of mission environments, the limitations of these traditional methods are becoming increasingly apparent.

[0003] While traditional mission planning methods demonstrate some effectiveness in structured mission environments, they exhibit significant limitations when dealing with highly variable mission environments and natural language instructions. Specifically, traditional methods require the pre-determination of precise planning domains and logical rules, making them difficult to adapt to new mission scenarios and undefined instruction sets. This framework, which relies on expert knowledge, is already struggling to adapt to complex tasks. Furthermore, traditional methods lack flexibility and dynamism, making it impossible to adjust mission plans in real time to accommodate environmental changes. With the emergence of large language models (LLMs), mission planning has made significant progress in incorporating natural language understanding and multi-scenario generalization. However, direct application of LLMs to UAV swarm mission planning also faces numerous challenges, such as limited ability to analyze complex causal relationships and a lack of closed-loop feedback systems and explicit feasibility checks, resulting in potential risks and uncertainties in the generated mission plans. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes a UAV swarm task planning method based on a large language model to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for UAV swarm mission planning based on a large language model, comprising:

[0006] Collect and pre-process multi-level input information from the drone swarm, and pass the multi-level input information in a specific encoding format to the large language model as prompt information, so that the large language model can understand natural language instructions and generate high-level task objectives;

[0007] Using a traditional mission planner, the high-level mission objectives are recursively decomposed to form a task tree, and the subtasks after the task decomposition are assigned to the UAVs according to their capabilities;

[0008] Build a drone action function library, use a large language model to understand the task planning results generated by the traditional task planner, match and call the relevant action functions in the drone action function library, generate action sequences that can be directly executed by the drone, and check and verify the action sequences.

[0009] Preferably, the multi-level input information includes: mission information, drone role capability information, environmental information, and thought chain process;

[0010] Among them, the task information includes task objectives, task scope, and task priority; the drone role capability information includes the drone's load capacity, flight time, and sensor configuration; the environmental information includes the geographical information and meteorological information of the task scene; the thinking chain process is used to guide the large language model to think according to the preset logic.

[0011] Preferably, the process of recursively decomposing the high-level task objectives using a traditional task planner to form a task tree and allocating the decomposed subtasks to the UAVs according to the capabilities of the UAVs includes:

[0012] Decomposing the high-level task objectives using a traditional task planner according to task dependencies, priorities, or complexity to form a hierarchical structure of a task tree;

[0013] The subtasks after task decomposition are assigned to drones through pre-defined capability parameters.

[0014] Preferably, the process of building a drone action function library includes:

[0015] Constructing basic drone actions; wherein the basic actions include takeoff, landing, and hovering;

[0016] Construct complex operations of drones; wherein the complex operations include trajectory tracking, target recognition, and path planning.

[0017] Preferably, before generating an action sequence that can be directly executed by the drone, the process further includes:

[0018] Appropriate actions are called from the action library to form a sequence, and the generation process always follows a priori rules; wherein the a priori rules include the dependency relationship and execution order between actions.

[0019] Preferably, before generating an action sequence that can be directly executed by the drone, the process further includes:

[0020] The action sequence is accurately mapped to the functions in the function tool library through similarity matching, wherein the similarity matching includes vector space analysis and word meaning relevance screening.

[0021] Preferably, the process of generating an action sequence that can be directly executed by the drone includes:

[0022] Evaluate the mission planning results generated by the traditional mission planner using a large language model to identify potential planning errors, including logical errors and dependency conflicts;

[0023] Using a large language model to understand and fill in the slots of the subtasks, the subtasks are parsed into structured task descriptions; wherein the task descriptions include standardized fields such as action type, height, target object, and motion parameters;

[0024] Leverage the contextual understanding capabilities of large language models to infer implicit parameters and resolve references.

[0025] Preferably, the process of checking and verifying the action sequence includes:

[0026] Evaluate the feasibility and safety of action sequences;

[0027] A simulation test is performed on the execution environment of the action sequence to obtain a simulation test result, wherein the simulation test result is used to adjust the action sequence.

[0028] In a second aspect, the present invention further discloses a computer-readable storage medium having a computer program stored thereon, which implements the steps of the method described in the first aspect when executed by a processor.

[0029] Compared with the prior art, the present invention has the following advantages and technical effects:

[0030] The present invention provides a method for unmanned aerial vehicle (UAV) swarm task planning based on a large language model, comprising the following steps: first, collecting and preprocessing multi-level input information, and passing the multi-level input information to the large language model in a specific encoding format, so that the large language model understands natural language instructions and generates task planning results; second, using a traditional task planner to recursively decompose the high-level task objectives to form a task tree, and assigning the decomposed subtasks to the UAVs according to the capabilities of the UAVs; finally, constructing a UAV action function library, using the large language model to understand the task planning results generated by the traditional task planner, and matching and calling relevant action functions in the UAV action function library to generate an action sequence that can be directly executed by the UAVs, and checking and verifying the action sequence.

[0031] This invention combines a large language model with a traditional mission planner, effectively addressing multiple scenarios while ensuring feasible output. Leveraging the semantic understanding and generative capabilities of the large language model, it overcomes the shortcomings of traditional mission planners in real-time dynamic decision-making. This integrated approach enables a more reliable and efficient drone mission execution system that adapts to changing environmental demands and enables real-time, flexible adjustments.

[0032] The present invention constructs a UAV action function library and avoids the danger of directly using a large language model for UAV mission planning by matching and calling related action functions.

[0033] This invention utilizes a large language model to achieve intelligent planning of drone swarm missions, providing an efficient task allocation strategy for multi-drone systems. The resulting planning scheme is not only efficient and accurate, but also ensures feasibility and optimization through the establishment of multiple evaluation metrics. Ultimately, this achieves interpretable and optimal task planning, effectively improving the reliability and overall operational efficiency of drone swarm mission execution.

[0034] In the face of changing mission environments, this invention combines the potential of large language models with the reliability and security of traditional mission planning to effectively increase the flexibility and adaptability of drone swarm mission planning methods. This integrated approach not only improves the accuracy and stability of mission execution but also reduces the need for human intervention, promoting the practicality and efficiency of drone mission planning for complex missions in multiple scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0036] Figure 1 This is a schematic diagram of the task planning framework of "task understanding - task decomposition and allocation - action sequence generation - inspection and feedback" in an embodiment of the present invention;

[0037] Figure 2 A schematic diagram of an example of drone swarm mission planning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0039] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0040] Example 1

[0041] The existing technology has at least one of the following problems:

[0042] 1. Traditional rule-based planning algorithms require the pre-definition of precise planning domains and logical rules, which makes them difficult to adapt to new task scenarios and deal with undefined instruction sets. They have obvious limitations when dealing with ever-changing task environments and natural language instructions.

[0043] 2. Due to problems such as hallucinations inherent in large language models, the method of using large language models alone for task planning has great uncertainty.

[0044] 3. Existing task planning methods based on large language models are mostly applied to single intelligent agents, and the planning methods do not have a check and feedback mechanism, making it difficult to solve complex tasks in dynamic scenarios.

[0045] To address the above issues, this embodiment provides a UAV swarm mission planning method based on a large language model. Figure 1-2 The main goal of this invention is to solve the task planning problem of drone swarms in different mission scenarios. Unlike a single drone executing a mission, the task planning process of a drone swarm needs to consider the matching capabilities of the drones to allocate tasks and ultimately achieve safe control of the drones.

[0046] The present invention first uses the powerful common sense reasoning ability of the large language model to understand complex and ambiguous natural language instructions. Then, it generates inputs that can be accepted by the basic task planner to obtain the subtask allocation results for each drone in the drone swarm. After receiving the corresponding subtask, the drone calls the relevant function actions in the library according to the relevant action definitions in the tool library to form the drone's action sequence. Finally, the large language model acts as a checker to ensure that the generated action sequence is acceptable and executable by the drone, thereby avoiding the uncertainty of the large language model and completing the drone's task planning. On the one hand, the present invention can realize rich downstream functions by supplementing the advanced drone function library. On the other hand, different prompts can be designed according to different task environments, thereby forming a universal execution unit.

[0047] A method for UAV swarm mission planning based on a large language model, comprising the following steps:

[0048] S1. Collect and pre-process multi-level input information of the drone swarm, and pass the multi-level input information to the large language model in a specific encoding format as prompt information, so that the large language model can understand natural language instructions and generate high-level task objectives;

[0049] Furthermore, the multi-level input information includes: mission information, drone role capability information, environmental information, and thought chain process;

[0050] Among them, the task information includes task objectives, task scope, and task priority; the drone role capability information includes the drone's load capacity, flight time, and sensor configuration; the environmental information includes the geographical information and meteorological information of the task scene; the thinking chain process is used to guide the large language model to think according to the preset logic.

[0051] Specifically, the mission understanding phase involves initially collecting and preprocessing multi-level input information. The acquired mission information, drone role capability information, environmental information, and thought chain process information are then passed to the large language model via a specific encoding format (JSON) as system prompts. This allows the model to understand complex and ambiguous natural language instructions, form an overall mission plan, and convert it into a standardized input format that can be used by the basic mission planner.

[0052] During the mission understanding phase, a prompting process is designed that includes human instructions, information about the UAV's role capabilities, environmental information, and thought chain processes. This phase does not directly output the mission planning results, but rather understands the complex and ambiguous natural language instructions and converts them into an input format that can be directly used by the basic mission planner.

[0053] In one example, a human instruction is "Conduct reconnaissance of the southern industrial area of ​​the city, focusing on suspicious vehicle activity." The drone role capability information includes payload information for each drone, environmental information including imagery of the mission scenario, and real-time imagery captured by the drones. The thought chain process guides the large language model to think and output in a targeted manner. Based on this information, the large language model understands the natural language instruction, forms an overall mission plan, and outputs it to the basic planner. In this example, the overall plan output from the mission understanding phase is "target reconnaissance - target positioning - continuous surveillance - data collection."

[0054] S2. Recursively decompose the high-level mission objectives using a traditional mission planner to form a mission tree, and assign the decomposed subtasks to the UAVs based on their capabilities to generate a mission planning result.

[0055] Furthermore, the process of recursively decomposing the high-level mission objectives using a traditional mission planner to form a mission tree and allocating the decomposed subtasks to the UAVs according to the capabilities of the UAVs includes:

[0056] Decomposing the high-level task objectives using a traditional task planner according to task dependencies, priorities, or complexity to form a hierarchical structure of a task tree;

[0057] The subtasks after task decomposition are assigned to drones through pre-defined capability parameters.

[0058] Specifically, in the task decomposition and assignment phase, after receiving the input information, the basic planner decomposes the task into multiple executable task sequences and assigns them to appropriate drones. The entire process is carried out under the framework of the Hierarchical Task Network (HTN) and includes the following steps:

[0059] S201: Constructing a Hierarchical Task Network. This invention uses HTN as the basic planner for UAV swarm task planning. By analyzing task requirements, the overall task is recursively decomposed to form a task tree.

[0060] S202: Task Decomposition and Subtask Assignment. Leveraging the decomposition capabilities of the HTN, high-level, complex tasks are gradually broken down into easily executable subtasks. At each task node, drone capabilities are matched to ensure that the task is assigned to a drone with the appropriate capabilities. The capability matching process considers drone attributes such as payload capacity, flight time, and sensor configuration, all of which are managed using predefined capability parameters.

[0061] In the task decomposition and assignment phase, a hierarchical task network is constructed, and the overall task is recursively decomposed to form a task tree. First, the input task set is organized into a hierarchical structure based on the task dependency, priority or complexity, laying the foundation for subsequent task decomposition and assignment. Then, each layer of the task hierarchy is traversed from top to bottom to obtain all subtasks of the current level and all available drone individuals that can handle the tasks of the current level. For each subtask, the most suitable drone is selected to execute based on factors such as the drone's payload. If a suitable drone is found, the task is assigned and the drone is removed from the available list; if no suitable drone is found, the task is further decomposed into smaller subtasks, and these new subtasks are added to the current level for subsequent re-assignment attempts. Finally, each drone generates an execution plan based on the task assigned to it. The generated action sequence is as follows:

[0062] 1) Drone 1: Take off

[0063] 2) Drone 1: Plan a path and fly to <Industrial Zone>

[0064] 3) Drone 1: Target Identification <Car>

[0065] 4) Drone 1: Target Positioning <Car>

[0066] 5) Drone 2: Takeoff

[0067] 6) Drone 2: Plan a path and fly <car>

[0068] 7) Drone 2: Tracking <Car>

[0069] 8) Drone 1: Plans a path and flies to <starting point>

[0070] 9) Drone 1: Land.

[0071] S3. Build a drone action function library, use a large language model to understand the task planning results generated by the traditional task planner, match and call the relevant action functions in the drone action function library, generate an action sequence that can be directly executed by the drone, and check and verify the action sequence.

[0072] Furthermore, the process of building a drone action function library includes:

[0073] Constructing basic drone actions; wherein the basic actions include takeoff, landing, and hovering;

[0074] Construct complex operations of drones; wherein the complex operations include trajectory tracking, target recognition, and path planning.

[0075] Furthermore, before generating an action sequence that can be directly executed by the drone, the following steps are also included:

[0076] Appropriate actions are called from the action library to form a sequence, and the generation process always follows a priori rules; wherein the a priori rules include the dependency relationship and execution order between actions.

[0077] Furthermore, before generating an action sequence that can be directly executed by the drone, the following steps are also included:

[0078] The action sequence is accurately mapped to the functions in the function tool library through similarity matching, wherein the similarity matching includes vector space analysis and word meaning relevance screening.

[0079] Furthermore, the process of generating an action sequence that can be directly executed by the drone includes:

[0080] Evaluate the mission planning results generated by the traditional mission planner using a large language model to identify potential planning errors, including logical errors and dependency conflicts;

[0081] Using a large language model to understand and fill in the slots of the subtasks, the subtasks are parsed into structured task descriptions; wherein the task descriptions include standardized fields such as action type, height, target object, and motion parameters;

[0082] Leverage the contextual understanding capabilities of large language models to infer implicit parameters and resolve references.

[0083] Furthermore, the process of checking and verifying the action sequence includes:

[0084] Evaluate the feasibility and safety of action sequences;

[0085] A simulation test is performed on the execution environment of the action sequence to obtain a simulation test result, wherein the simulation test result is used to adjust the action sequence.

[0086] Specifically, during the inspection execution phase, we integrate a series of basic actions and complex operations required for the drone to perform tasks, encapsulate the algorithms involved in these actions into a standardized function tool library for the large language model to call, and form an action sequence that the drone can directly execute. This avoids the uncertainty caused by the large language model directly generating code. The steps are as follows:

[0087] S301: Build a function tool library. The tool library includes basic drone actions such as takeoff, landing, and hovering. It also encapsulates advanced functions such as trajectory tracking, target recognition, and path planning, tailored to the specific mission scenarios. Each action is precisely defined, including its functional description, required parameters, and execution conditions.

[0088] S302: Task plan review. Leveraging the analytical and reasoning capabilities of the large language model, the subtask sequence is re-evaluated to identify potential planning errors, such as logical errors. Completeness checks confirm the coherence and operability of the task chain.

[0089] S303: Function Call. Based on the task planning results, appropriate actions are called from the action library to form a sequence. The generation process follows a priori rules, including dependencies between actions and execution order, to ensure the rationality and coherence of the action sequence.

[0090] S304: Similarity Matching and Correction. During the parsing process, similarity matching analysis is performed between the action sequence and the function names in the tool library to ensure that the generated sequence accurately matches the reliable functions in the tool library. Through vector space analysis and word semantic relevance screening, the execution sequence is ensured to be safe and usable, and possible solution deviations and parameter conflicts are corrected in advance.

[0091] S305: Parameter Mapping and Behavior Generation. The large language model is used to understand and fill in the slots of subtasks, parsing them into structured task descriptions. Instructions are parsed into a structured representation containing standardized fields such as action type, altitude, target object, and motion parameters. Simultaneously, the large language model's contextual understanding capabilities are leveraged to infer implicit parameters and resolve references. The parameters parsed by the large language model are injected into the corresponding motion planning module through a unified algorithm call interface, completing the development of the specific action sequence for the drone. This integration ensures the controllability of the drone's behavior.

[0092] In this embodiment, a comprehensive task execution framework is constructed in the inspection and execution phase. By deeply integrating the reasoning capabilities of the large language model with the specialized UAV control system, it ensures that the task intent is efficiently, safely, and accurately converted into the actual execution effect of the physical world. First, a layered encapsulation architecture is used to build a UAV action function library. The basic flight control layer integrates the basic flight control actions of the UAV (such as takeoff, landing, hovering, and fixed-point flight), and the task execution layer encapsulates the UAV perception and decision-making function modules (such as target detection and recognition, path planning, and obstacle avoidance). Each functional module is standardized and includes detailed interface specifications and functional descriptions. The above information is also described in the prompt project in the task understanding phase.

[0093] The large language model further verifies and audits the subtask action sequence results obtained during the task decomposition and assignment phase. Using its reasoning capabilities, the large language model analyzes the action sequence's coherence and logical integrity, detecting dependency conflicts within the task chain to ensure smooth transitions between subtasks. For example, in this example, target tracking must occur after target discovery and positioning. After confirming that the action sequence can be executed, the appropriate function is called from the function library. By calculating the semantic similarity between each action description in the action sequence and the function library, a mapping from fuzzy actions to precise functions is achieved. Furthermore, fine-grained semantic analysis is applied to extract explicit and implicit parameters from the action sequence. For example, from the command "Inspect 50 meters above the factory, focusing on chimney emissions," the flight altitude, inspection object, and focus are extracted. Based on the task type and scenario characteristics, reasonable default values ​​are filled in for unspecified parameters. For example, when receiving the action command "Return to the origin from here," a path to the starting point is automatically planned. Context-aware reference resolution is also implemented to handle ambiguous references in task descriptions, such as accurately identifying the specific objects or locations referred to by terms like "it" and "there" in a multi-target environment.

[0094] See also Figure 1 The task planning example in , the final planned executable function sequence is as follows.

[0095] 1) take_off(UAV1)

[0096] 2)path_plan(start_point,Industrial zone)

[0097] 3)track(UAV1,path_points_1)

[0098] 4)target_detect(UAV1,car)

[0099] 5) locate(UAV1,car)

[0100] 6) take_off(UAV2)

[0101] 7)track(UAV2,path_points_1)

[0102] 8) monitor (UAV2, car)

[0103] 9)path_plan(Industrial zone,start_point)

[0104] 10)track(path_points_2)

[0105] 11)land(UAV1)

[0106] To verify the effectiveness of this embodiment, experiments were conducted in a simulation environment. The payloads of the drones included optoelectronic and infrared sensors, proving the feasibility of this embodiment. Specifically, the experiment designed six typical mission scenarios ranging from simple to complex, and randomly initialized the positions of each drone in these scenarios. For the generated mission plan, this embodiment introduced various criteria including mission success rate, collision probability during flight, aircraft crash probability, and trajectory stability to better evaluate the performance of the mission plan. The experimental results are as follows:

[0107] Table 1

[0108]

[0109] This embodiment is named MUTP-LLM, and the bold values ​​represent the best indicators among all methods. It can be seen that this embodiment outperforms other methods in various tasks, including large model direct planning methods.

[0110] It should be noted that this embodiment provides a technical framework for multi-UAV mission planning that combines a large language model with a traditional planner. Specifically, the large language model is used to provide top-level understanding of the natural language command input from the human commander, and formalized parameters are passed to the mission planner. By designing prompts, a mechanism is further designed to check the generated subtask sequences using the large language model. Different prompts can be designed for different mission scenarios and heterogeneous UAVs to achieve more efficient results.

[0111] Beneficial effects of this embodiment:

[0112] This embodiment provides a drone swarm task planning method based on a large language model. It uses the large language model in combination with a traditional task planner to solve the task planning problem of the drone swarm system. It provides a planning mechanism of task understanding - task decomposition and allocation - action sequence generation - inspection and feedback. By combining the extensive world knowledge of the large language model with the reliability of the traditional task planner, it has good generalization for different types of tasks and scenarios, and has been experimentally verified in a simulation environment.

[0113] Example 2

[0114] This embodiment further discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in the first embodiment are implemented.

[0115] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A UAV swarm mission planning method based on a large language model, characterized in that: The following steps are involved: Collect and pre-process multi-level input information from the drone swarm, and pass the multi-level input information in a specific encoding format to the large language model as prompt information, so that the large language model can understand natural language instructions and generate high-level task objectives; Using a traditional mission planner, the high-level mission objectives are recursively decomposed to form a mission tree, and the subtasks after the mission decomposition are assigned to the UAVs according to their capabilities to generate a mission planning result; Build a drone action function library, use a large language model to understand the task planning results generated by the traditional task planner, match and call the relevant action functions in the drone action function library, generate action sequences that can be directly executed by the drone, and check and verify the action sequences.

2. The method according to claim 1, characterized in that The multi-level input information includes: mission information, UAV role capability information, environmental information, and thought chain process; Among them, the task information includes task objectives, task scope, and task priority; the drone role capability information includes the drone's load capacity, flight time, and sensor configuration; the environmental information includes the geographical information and meteorological information of the task scene; the thinking chain process is used to guide the large language model to think according to the preset logic.

3. The method according to claim 1, characterized in that The process of using a traditional mission planner to recursively decompose the high-level mission objectives to form a task tree and assigning the decomposed subtasks to the UAVs according to their capabilities includes: Decomposing the high-level task objectives using a traditional task planner according to task dependencies, priorities, or complexity to form a hierarchical structure of a task tree; The subtasks after task decomposition are assigned to drones through pre-defined capability parameters.

4. The method according to claim 1, wherein The process of building a drone motion function library includes: Constructing basic drone actions; wherein the basic actions include takeoff, landing, and hovering; Construct complex operations of drones; wherein the complex operations include trajectory tracking, target recognition, and path planning.

5. The method according to claim 1, wherein Before generating an action sequence that can be directly executed by the drone, it also includes: Appropriate actions are called from the action library to form a sequence, and the generation process always follows a priori rules; wherein the a priori rules include the dependency relationship and execution order between actions.

6. The method according to claim 1, characterized in that Before generating an action sequence that can be directly executed by the drone, it also includes: The action sequence is accurately mapped to the functions in the function tool library through similarity matching, wherein the similarity matching includes vector space analysis and word meaning relevance screening.

7. The method according to claim 1, characterized in that The process of generating an action sequence that the drone can directly execute includes: Evaluate the mission planning results generated by the traditional mission planner using a large language model to identify potential planning errors, including logical errors and dependency conflicts; Using a large language model to understand and fill in the slots of the subtasks, the subtasks are parsed into a structured task description, wherein the task description includes standardized fields such as action type, height, target object, and motion parameters; Leverage the contextual understanding capabilities of large language models to infer implicit parameters and resolve references.

8. The method according to claim 1, characterized in that The process of checking and verifying the action sequence includes: Evaluate the feasibility and safety of action sequences; A simulation test is performed on the execution environment of the action sequence to obtain a simulation test result, wherein the simulation test result is used to adjust the action sequence.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images