Control methods, devices, equipment, and media for unmanned aerial vehicle (UAV) swarm intent-guided missions

Abstract

This application belongs to the field of unmanned aerial vehicle (UAV) control technology, and relates to control methods, devices, equipment, and media for UAV swarm intent-oriented tasks. The method includes: acquiring instruction information for the intent-oriented task; analyzing natural language and gesture commands in the instruction information to obtain dynamic parameters and generate a preliminary spatial intent vector; acquiring a knowledge graph and generating an intent data package; acquiring the state of the UAV swarm, constructing internal and external constraints to obtain a dynamically changing environment; assigning roles to each UAV in the swarm based on the intent data package and the environment, and planning corresponding sub-tasks for different roles to plan collaborative behavior paths for the UAV swarm and generate a four-dimensional action sequence; and obtaining a collaborative planning scheme based on different roles, corresponding sub-tasks, and the four-dimensional action sequence to achieve control of the UAV swarm. This application enables a deep understanding of UAV swarm intent-oriented tasks and the control of UAVs.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) control technology, and in particular to control methods, apparatus, equipment and media for UAV swarm intent-oriented missions. Background Technology

[0002] At the forefront of the deep intersection of human-computer interaction and artificial intelligence, focusing on the collaborative control of multi-agent systems, with the application of drone swarms as a typical scenario, this field is at a critical stage of evolution from automation to intelligence and autonomy, and is an important indicator of a country's core competitiveness in the fields of robotics, unmanned systems and artificial intelligence.

[0003] With the rapid development of drone technology, the application of single drones has matured, while swarm systems composed of a large number of heterogeneous drones are becoming a transformative force in fields such as military reconnaissance and strike, disaster emergency response, public safety monitoring, large-scale precision agriculture, and 3D mapping due to their huge advantages in coverage, mission efficiency, functional redundancy, and damage resistance.

[0004] However, the exponential growth of bee colony capabilities has created an increasingly acute contradiction with the linear development of its control methods, constituting the main pain point and bottleneck in current technological development.

[0005] Existing drone swarm control technologies, whether graphical programming, limited command control, or tactical database-based systems, largely remain within the realm of command-based or programmable methods. Operators typically need to pre-plan detailed flight paths and mission maneuvers for each drone or swarm through complex graphical user interfaces. Even with the introduction of voice control, it is mostly limited to recognizing fixed commands such as takeoff, landing, and executing Plan A, and cannot handle ambiguous or complex intentions.

[0006] This control mode has the following drawbacks: 1. High cognitive load: Operators need to process massive amounts of information and perform tedious micro-operations, making it difficult to cope with rapidly changing and complex environments.

[0007] 2. Low adaptability: The preset scripts and solutions are difficult to cope with sudden situations in the battlefield or disaster area, and the system lacks the ability to adjust and repair itself.

[0008] 3. Human-machine collaboration gap: There is a huge gap between human strategic intuition and machine execution capabilities, making it impossible to form smooth and efficient collaborative decision-making.

[0009] 4. Cognitive Rigidity: The system lacks true understanding and can only mechanically execute pre-set scripts or recognize fixed instructions. This results in extremely poor adaptability when facing dynamic, unknown, and complex real-world environments. It is unable to autonomously adjust its strategies to cope with unexpected situations, exhibiting severe fragility.

[0010] 5. Human-Machine Disconnection: These systems place operators in a cognitively burdensome micromanagement role, forcing them to expend energy on tedious low-level operations. Valuable human strategic intuition, experiential judgment, and creative thinking are artificially excluded from the decision-making loop at critical stages of task execution, failing to efficiently integrate with the machine's computing power. Summary of the Invention

[0011] Therefore, it is necessary to provide a control method, apparatus, device, and medium for UAV swarm intent-oriented tasks to address the aforementioned technical problems, which can deeply understand UAV swarm intent-oriented tasks and control the UAVs.

[0012] Control methods for drone swarm intent-driven missions include: Obtain instruction information for intent-driven tasks, analyze the natural language and gesture instructions in the instruction information, obtain dynamic parameters, and generate a preliminary spatial intent vector; The knowledge graph is acquired and, combined with a language parsing engine, the initial spatial intent vector is parsed to obtain keywords and generate intent data packages. Obtain the state of the drone swarm, construct internal and external constraints, and obtain the dynamically changing environment; Based on the intent data packet and the environment, each drone in the drone swarm is assigned a role, and corresponding sub-tasks for different roles are planned. Based on different roles and corresponding sub-tasks, a collaborative behavior path is planned for the drone swarm, generating a four-dimensional action sequence. Based on different roles, corresponding sub-tasks, and the four-dimensional action sequence, a collaborative planning scheme is obtained to achieve the control of the drone swarm.

[0013] In one embodiment, after obtaining the collaborative planning scheme, the method further includes: The drones begin executing their missions according to the collaborative planning scheme. When an emergency occurs, the system identifies the failed nodes and the valid missions of the relevant drones, and generates different repair plans from the perspectives of different roles. Select one of the repair options as the optimal solution and perform distributed adaptive replanning for the drone swarm.

[0014] In one embodiment, after obtaining the collaborative planning scheme or performing distributed adaptive replanning for the drone swarm, the method further includes: Based on the results of the intent-driven task, different regions of the target area are colored to achieve semantic coloring of the regions; Based on the progress of the intent-driven task, the normal status of the drone swarm is displayed to achieve a story-like narrative of the task progress. Based on emergencies and optimal solutions, the abnormal status of the drone swarm is displayed to enable the repair of critical events.

[0015] In one embodiment, after obtaining the collaborative planning scheme, the method further includes: Acquire query information, analyze the natural language and gesture commands in the query information, and focus on the region based on the query information; Acquire new intent information, analyze the natural language and gesture commands in the new intent information, and combine them with the current collaborative planning scheme to deduce alternative suggestions; Based on regional focus and alternative suggestions, make dialogical intent adjustments.

[0016] In one embodiment, after generating the four-dimensional action sequence, the method further includes: Different emergency plans are generated based on different unexpected situations; Based on different roles, corresponding sub-tasks, four-dimensional action sequences, and emergency plans, a collaborative planning scheme is obtained to achieve the control of the drone swarm.

[0017] In one embodiment, the initial spatial intent vector includes: geographic location, range, and initial weights related to the mode of action, including: speed and stealth.

[0018] In one embodiment, the intent data packet includes: target level, key constraints, execution resources, geometric and environmental features of the work area, and a comprehensive confidence score. The target hierarchy includes primary and secondary targets, and the key constraints include concealment and timeliness.

[0019] Control devices for drone swarm intent-guided missions include: The first module is used to acquire instruction information for intent-driven tasks, analyze the natural language and gesture instructions in the instruction information, obtain dynamic parameters, and generate a preliminary spatial intent vector. The second module is used to acquire the knowledge graph, combine it with the language parsing engine to parse the initial spatial intent vector, obtain keywords, and generate intent data packages; The third module is used to obtain the state of the drone swarm, construct internal and external constraints, and obtain the dynamically changing environment. The fourth module is used to assign roles to each drone in the drone swarm based on the intent data packet and the environment, and plan the corresponding sub-tasks for different roles; based on different roles and corresponding sub-tasks, it plans the collaborative behavior path for the drone swarm and generates a four-dimensional action sequence; based on different roles, corresponding sub-tasks and the four-dimensional action sequence, it obtains a collaborative planning scheme to achieve the control of the drone swarm.

[0020] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the above-described method.

[0021] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0022] The aforementioned control method for drone swarm intent-oriented tasks breaks through existing technological bottlenecks by constructing an intent-oriented human-machine interaction and collaborative decision-making system. This frees operators from tedious micro-management, allowing them to focus on high-level strategic decision-making. At the same time, it endows the swarm system with a higher level of autonomy and intelligence, realizing true intelligent collaboration between humans and drone swarms, thereby fully unleashing the enormous potential of the swarm system in future applications. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a control method for an intent-driven task involving a drone swarm, as shown in one embodiment. Figure 2 This is a structural block diagram of a control device for an intent-guided mission of an unmanned aerial vehicle swarm in one embodiment; Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0025] Furthermore, the use of terms such as "first" and "second" in this application is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of those features. In the description of this application, "multiple sets" means at least two sets, such as two sets, three sets, etc., unless otherwise explicitly specified.

[0026] In this application, unless otherwise expressly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection, an electrical connection, a physical connection, or a wireless communication connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two elements or the interaction between two elements, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0027] Furthermore, the technical solutions of the various embodiments of this application can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0028] This application provides a control method for unmanned aerial vehicle (UAV) swarm intent-oriented tasks, such as... Figure 1 The flowchart shown, in one embodiment, includes: Step 101: Obtain instruction information for the intent-driven task, analyze the natural language and gesture instructions in the instruction information, obtain dynamic parameters, and generate a preliminary spatial intent vector.

[0029] Specifically: The system acquires instruction information for intent-driven tasks, analyzes the natural language and gesture instructions in the instruction information to obtain dynamic parameters, compares and verifies the natural language and gesture instructions, updates the weights and confidence of the dynamic parameters, and generates a preliminary spatial intent vector based on the dynamic parameters, weights, and confidence.

[0030] The initial spatial intent vector includes: geographic location, range, and initial weights related to the mode of action, which includes: speed and stealth.

[0031] Gesture commands encompass the gesture itself, as well as its "speed," "trajectory coherence," "trajectory smoothness," and "force," to more accurately understand the subtext of intent. Speed ​​determines the rhythm of the task, while force or pointing gestures represent priority. This fusion mechanism can detect contradictions within commands. Data acquisition doesn't rely on simple mouse clicks but rather on high-precision, high-refresh-rate input devices to collect gesture data. These could be tactical tablets supporting multi-touch and pressure sensing, interactive sand tables, or computer vision (such as ToF depth cameras) tracking operator hand movements in space. The device records a series of time-series data points with millisecond-level precision. Gesture recognition employs machine learning-based classifiers (e.g., using recurrent neural networks (RNNs) or long short-term memory networks (LSTMs)) to identify specific tactical gestures. For example, "two-finger zoom" represents "focus / zoom," "wave" represents "cancel / clear," and a decisive "chop" gesture might represent "segmenting an area" or "establishing a no-fly zone." When analyzing velocity, the velocity and acceleration system calculates the first and second derivatives at each point on the trajectory to extract key dynamic features and map tactical semantics. When analyzing trajectory smoothness, trajectory smoothness and jitter-passing algorithms (such as Kalman filtering) can be used.

[0032] Generating vectors from features of natural language and gesture commands involves: capturing the complete dynamic process of natural language and gesture commands through sensors (such as cameras and touchscreens), quantifying the behavior and mapping it to a preset parameter space, and decomposing it into computable features. Specifically: 1) Static features → Spatial vector (geographical location and range) Gesture trajectory: The vertex coordinates of circles or polygons drawn on the map directly constitute the geographical location and range information in the vector.

[0033] 2) Dynamic characteristics → Behavioral weights (action patterns) Gesture speed / acceleration: High speed, high acceleration (rapid, hasty delineation of movements) → Increase the weight of "speed" and decrease the weight of "concealment," which is interpreted as "urgent" and "efficiency first." Low speed and low acceleration (slow, cautious depiction of movement) → Increase the weight of "stealth" and decrease the weight of "speed," which is interpreted as "caution" and "risk avoidance."

[0034] Gesture strength / pressure: A sudden increase in emphasis or a tapping action in the trajectory → Marks a "point of interest" or "area of ​​focus" in the vector, affecting the priority of subsequent tasks.

[0035] 3) Natural Language Features → Behavioral Weights (Semantic Correction) Keywords: Words like "quickly" and "immediately" in language instructions will further amplify the weight of "speed"; while "quietly" and "avoid being discovered" will amplify the weight of "concealment".

[0036] Entity association: When "ghost group" is mentioned, the knowledge graph shows that its characteristic is concealment, so "concealment" will be automatically given a high initial weight.

[0037] Fusion Generation: The above-mentioned quantitative features are weighted and fused. If the gestures are slow and the language requires "concealment," the "concealment" weight will be very high if both are consistent. If the two are contradictory (e.g., the gestures are fast but the language requires "concealment"), the system may lower the confidence of the vector or even proactively ask clarifying questions.

[0038] Weight updates can be understood as a reward and punishment system: Final concealment weight = Base weight × Consistency factor; The base weight comes from the direct analysis of language instructions. For example, the phrase "avoid being detected" itself gives stealth a high base weight value.

[0039] Consistency factor: Reward (factor > 1): When the gesture is "slow and cautious", this factor will be greater than 1, which will enhance the base weight and eventually result in a very high "stealth" weight; Neutral (Factor = 1): When gesture features are not obvious, the factor is 1, and the weight remains unchanged; Penalty (factor < 1): When the gesture is "quick and sloppy", this factor will be less than 1, which will weaken the base weight and indicate that the certainty of this "covert" instruction is not high.

[0040] The confidence level determines how certain one is about this weighting assessment: Final confidence level = Initial confidence level × Consistency factor; Initial confidence level refers to an initial assessment of the accuracy of language parsing.

[0041] Consistency factor: When language and gestures are highly consistent, the factor is greater than 1, indicating that the system is very confident in its judgment of high concealment weight. When language and gestures contradict each other, the factor is much less than 1, and the system has very low confidence in the weight assessment.

[0042] The core logic of the above settings is that when the final confidence level is lower than a preset threshold, a clarifying question will be triggered. This mechanism ensures that the application will not perform the task based on a low-confidence, potentially erroneous understanding, thereby elevating human-computer interaction to a more intelligent and reliable level.

[0043] For example, an operator can engage in natural, intuitive interactions with a highly immersive interface (such as a tactical sand table, a large touchscreen, or a holographic display). Imagine a typical scenario: an operator draws an irregular circle in the air above a mountainous area on the sand table with their finger, while calmly saying, "I suspect there's a target hiding here. Send a 'ghost' team to investigate, but avoid detection." Traditional systems respond disjointedly to this: they might identify a geographical area and a voice command, then attempt a brute-force match. This solution, however, treats this as a unified, semantically rich whole, performing in-depth analysis. The core of this process lies in a parallel, cross-referencing analytical flow. The system acts like an experienced advisor, simultaneously analyzing two distinct information streams.

[0044] On the one hand, it deeply captures the dynamic semantics of gestures. The system focuses on much more than just the two-dimensional or three-dimensional geographical boundaries defined by gestures; it analyzes the complete dynamic process of gestures in real time: Is the circling speed fast or slow? Is the trajectory continuous or discontinuous? Is the pressure gentle or decisive? These dynamic characteristics are given tactical meaning; a fast, hasty circle (high average speed, high acceleration change) will be initially marked by the system as a quick overview or urgent intent, which may mean that the operator needs a quick overview scan, thus increasing the speed weight in the initial intent vector; while a slow, cautious trajectory precisely drawn along the ridgeline (low and steady speed) means "detailed planning" or "covert infiltration", strongly suggesting a covert infiltration operation that needs to utilize terrain cover, thus increasing the "covertness" weight; a continuous, smooth curve may represent a defined, complete area, while a discontinuous, jittery trajectory may indicate the operator's hesitation or uncertainty about the area, which the system may interpret as an area that needs to be focused on but whose nature is unknown; a sudden pointing or tapping action, meaning pressing or tapping on a certain point, represents a point of interest or "high-value target" that needs to be focused on, and the system will assign a very high attention weight to the coordinates of that point. The system analyzes these dynamic parameters to generate a preliminary spatial intent vector, which includes geographic location, range, and preliminary weights related to modes of action (such as speed and stealth).

[0045] On the other hand, there is the deep tactical deconstruction of natural language. This is not a simple keyword matching. The system has a built-in language parsing engine that has been trained with a large amount of military and emergency domain knowledge. (The language parsing engine is based on existing Transformer architecture language models (such as BERT or GPT series). On this basis, it is fine-tuned by fine-tuning the domain knowledge through a military and emergency response domain corpus. This corpus includes military regulations, tactical manuals, historical battle case analysis reports, communication records of simulated exercises, etc. Through fine-tuning, the model has learned the domain's "jargon", specific terms (such as "getting a clear picture of the situation" is tactically equivalent to "close reconnaissance" or "finding out the enemy situation"), and contextual information. The customized entity and intent recognition system [unit name] identifies "avoid being discovered" as a [core constraint] and "south side of the valley" as a [tactical geographical location]. Similarly, the intent classification is far beyond the conventional "check the weather" or "play music", but rather complex tactical intents such as [conducting reconnaissance], [establishing a blockade], and [conducting fire strikes], which can penetrate the ambiguity of language and extract core tactical elements.) In the sentence "Send a 'ghost' team to gather intelligence, but avoid detection," the engine identifies the actor (Who) as a ghost team. The system, through its internal resource library, knows this refers to a swarm of drones carrying high-resolution electro-optical / infrared pods and possessing low-observable characteristics. The core task (What) is to gather intelligence; this is a typical ambiguous objective, and the engine will interpret it as information reconnaissance rather than attack or occupation—this characterization is crucial. The core constraint (How) is to avoid detection; this is the highest principle of mission execution, its priority potentially even exceeding the comprehensiveness of reconnaissance. It directly relates to risk avoidance, path planning, and sensor usage patterns in subsequent strategy generation. Additionally, implicit emotion (Emotion / Urgency) is determined by analyzing the operator's tone and speaking speed to initially assess the mission's urgency, providing a basis for subsequent resource allocation and time planning.

[0046] The most crucial step is the fusion and semantic enrichment of cross-modal information. This is not a simple information overlay, but a reasoning process of mutual verification and reinforcement. The system compares the slow, cautious dynamic features captured in gestures with the core constraints of avoiding detection identified in language for cross-validation. For example, when the two (the "slow, cautious" features of gestures and the "avoid detection" constraint in language) are highly consistent, the system significantly increases the weight and confidence score of the "stealth" dimension in the initial spatial intent vector, greatly enhancing the confidence of the stealthy infiltration tactical mode, making the final generated intent more explicit and reliable. Conversely, if the operator's gestures are rapid, but the language requires stealth, the system may determine that there is a contradiction and may proactively ask to eliminate ambiguity: Your instruction requires rapid action, but also requires stealth; please confirm which should be prioritized.

[0047] In this step, how to analyze the natural language and gesture instructions in the instruction information to obtain dynamic parameters is existing technology and will not be elaborated here.

[0048] Step 102: Obtain the knowledge graph, combine it with the language parsing engine to parse the preliminary spatial intent vector, obtain keywords, and generate intent data packages.

[0049] The intent data package includes: target level, key constraints, execution resources, geometric and environmental features of the work area, and confidence score of comprehensive evaluation. The target level includes: primary and secondary targets, and the key constraints include: concealment and timeliness.

[0050] In this step, the knowledge graph acts like the brain of a seasoned commander, storing countless tactical concepts and their relationships. It transforms from an "information retrieval tool" into a "tactical decision-making reasoning engine." Using the knowledge graph for tactical reasoning, it identifies the paths between concepts, enriching the operator's vague intentions into concrete, machine-executable strategies. It solves not the "what" problem, but the complex planning problem of "how to do it." For example, the knowledge graph pre-stores knowledge that the best practice for covert reconnaissance in mountainous environments is to utilize terrain cover for low-altitude flight and prioritize the use of passive sensors. The deep coupling of the knowledge graph and the language parsing engine is a process of automatically filling in tactical details. It transforms high-level human intentions into detailed plans containing specific execution strategies through knowledge reasoning.

[0051] For example, after the system parses the elements of mountainous terrain, reconnaissance, and concealment, it activates a complex network in the knowledge graph. This network connects specific tactical knowledge such as terrain-matching flight, passive sensor priority, intermittent communication, and small-team dispersed infiltration. It is this graph that allows the system to enrich and concretize the operator's vague intentions into tactically feasible alternatives. The generated intent data packet fully encapsulates the operator's true intentions, including a clear target hierarchy (primary objective: reconnaissance; secondary objective: survival), key constraints (concealment > timeliness), execution resources (ghost team), the geometric and environmental features of the operational area, and a comprehensive confidence score. For another example, when the entity "'ghost team'" is parsed, the knowledge graph is immediately queried to obtain the team's detailed attributes (e.g., composed of three small UAVs, equipped with high-resolution electro-optical / infrared pods, possessing low detectability characteristics). This information, in turn, greatly helps the engine accurately understand the actual meaning and feasibility of the constraint "avoiding detection," thereby generating intent data packets more accurately.

[0052] The generation of initial spatial intent vectors and intent data packets allows for the accurate listening to and understanding of the operator's true intent, rather than merely translating their surface-level instructions.

[0053] It should be noted that knowledge graphs and language parsing engines are existing technologies and will not be discussed further here.

[0054] Step 103: Obtain the state of the drone swarm, construct internal and external constraints, and obtain the dynamically changing environment.

[0055] Specifically: The system acquires the status of each drone in the drone swarm in real time, including: precise location, remaining battery power, mounted equipment (e.g., drone #1 has a high-definition zoom camera, drone #2 has an infrared thermal imager, and drone #3 has an electronic signal monitoring device), as well as the quality and bandwidth of the communication link, and constructs internal constraints. Load all external conditions for the mission area, including: high-precision 3D terrain map, real-time meteorological data (wind speed, wind direction, visibility), known enemy air defense fire coverage, and airspace to be avoided such as civil aircraft routes, and construct external constraints; Based on internal and external constraints, a dynamically changing environment is obtained, which serves as the basis for strategy generation.

[0056] In this step, how to obtain the dynamically changing environment based on constraints is existing technology and will not be elaborated here.

[0057] Step 104: Assign roles to each drone in the drone swarm based on the intent data packet and the environment, and plan corresponding sub-tasks for different roles; based on different roles and corresponding sub-tasks, plan interrelated and cooperative behavior paths for the drone swarm in a unified spatiotemporal dimension, and generate a four-dimensional action sequence (four-dimensional refers to three-dimensional space and time); based on different roles, corresponding sub-tasks and four-dimensional action sequences, obtain a cooperative planning scheme to achieve control of the drone swarm.

[0058] Preferably, after generating the four-dimensional action sequence, the method further includes: generating different emergency plans based on different emergencies (such as: the assassin being discovered, the sentry losing contact); and obtaining a collaborative planning scheme based on different roles, corresponding sub-tasks, the four-dimensional action sequence, and the emergency plans to achieve the control of the drone swarm.

[0059] For example, based on the core intent of covert infiltration and reconnaissance in the previous step, the engine might make the following decisions: 1) Sentinel role allocation: It will select the No. 1 aircraft with the most power and equipped with a high-definition zoom camera, and plan a flight path for it that can reach the high ground on the periphery of the target area and remain airborne for a long time; its sub-task is: to conduct perimeter control observation of the target area, provide early warning for the infiltration team, and act as a relay communication node when necessary. 2) Assassin role allocation: It will select the No. 2 aircraft, which is smaller in size, equipped with infrared payloads and has the lowest radar cross-section, as the main infiltration unit; its sub-task is: to use terrain cover to penetrate deep into the core of the target area along the planned lowest detectable path and perform close-range reconnaissance. 3) Support role allocation: The No. 3 aircraft, equipped with electronic listening equipment, is assigned a support role; its sub-task is: to fly at a safe distance on the flank of the infiltration path, passively listen to electromagnetic signals in the area to detect potential radar or communication activities, and provide threat warnings to the assassins.

[0060] When planning infiltration routes for the assassin, the engine ensures that the flight path always remains within the radar blind spots of mountains or terrain. Simultaneously, it precisely calculates the positions of sentries, ensuring that when the assassin passes through a key pass, the sentry's camera is perfectly positioned to cover the area, providing cover. It also coordinates the positions of support aircraft, maximizing their surveillance range over unknown areas ahead of the assassin. This entire network of routes is filled with sophisticated tactical coordination: time synchronization, spatial complementarity, and functional synergy, all tightly centered around the core constraint of avoiding detection.

[0061] If the assassin is exposed, the backup strategy might be to immediately climb and disengage at maximum speed, while the sentries and support aircraft become feints to draw attention. These contingency plans are loaded into each sub-unit (one drone per sub-unit), providing the basis for subsequent distributed adaptive replanning.

[0062] In this step, the final generated coordinated action plan is not a simple list of waypoints, but a complete, multi-layered coordinated action plan. This plan includes task descriptions for each role, four-dimensional spatiotemporal paths, sensor usage rules (how to activate / deactivate active detection), communication protocols (how to maintain radio silence), and a series of if-then-if contingency response rules. This plan is like a meticulously choreographed symphony score, detailing when, where, and how each instrument (the drone) will play, to collectively present the grand symphony (tactical intent) desired by the conductor (operator). It is distributed to every member of the swarm, guiding their highly intelligent and coordinated actions in the real world.

[0063] It should be noted that how to plan interconnected and cooperative behavioral paths for drone swarms in a unified spatiotemporal dimension based on different roles and corresponding sub-tasks, and generate four-dimensional action sequences, is existing technology and will not be elaborated here.

[0064] In this embodiment, after obtaining the collaborative planning scheme, the method further includes: the UAV starts to execute the task according to the collaborative planning scheme; when a sudden event occurs (e.g., the signal may be suddenly interrupted, a new threat may appear out of nowhere, or the target may move unexpectedly), the method determines the failed node and the valid tasks of the relevant UAV, and generates different repair schemes from the perspective of different roles; selects one of the repair schemes as the optimal scheme, and performs distributed adaptive replanning for the UAV swarm.

[0065] The aforementioned contingency response setup elevates the swarm from a passive executor to an organic whole with autonomous repair and adaptation capabilities. Its core idea lies in the inheritance and decentralization of intent. In step 104, the system issues not only specific waypoints and action commands to each drone or subswarm, but more importantly, each drone inherits a copy of the upper-level tactical intent or its key summary. This means that the assassin drone performing the infiltration mission not only knows it is heading towards coordinate A, but also understands that its purpose is to ascertain the situation undetected. Similarly, the sentry drone responsible for cover understands that the highest principle of the entire operation is concealment. This shared understanding of higher-level intent is the cognitive foundation for achieving distributed adaptive replanning. This mechanism demonstrates its power when contingencies occur.

[0066] For example, imagine a scenario: Assassin Unit 2, while deep in a valley, suddenly encounters strong electromagnetic interference, causing a temporary loss of communication with the swarm and its rear. In a traditional system, it might hover due to its inability to receive the next command, or return to base as planned, leading to mission failure; while other drones would fall into chaos due to the loss of a critical node. However, in this solution, a series of intelligent, distributed chain reactions are automatically triggered: 1) Rapid Detection and Information Sharing of Local Failures. Sentinel Unit 1, responsible for communications relay and peripheral observation, is the first to detect the loss of contact with Unit 2. Instead of simply labeling it a communication failure, it immediately assesses the impact of this event on the entire covert reconnaissance mission, considering its own mission and understanding of the overall intent. It determines that the core reconnaissance unit has been disrupted, and the mission is at risk of failure. This assessment is immediately shared via encrypted data link with other surviving members in the swarm network, such as Unit 3, which is responsible for electronic support.

[0067] 2) Distributed Collaborative Replanning Based on Shared Intent. Decision-making is no longer solely centralized in the rear command center. The swarm itself, especially the sentry drones with superior situational awareness, will initiate a localized, rapid tactical consultation. It will assess the current situation: While drone #2 is out of contact, the original intent to avoid detection and ascertain the situation remains valid. Therefore, it will not recklessly order other aircraft to search for drone #2, as this would increase exposure risk. Instead, it may autonomously generate several candidate remediation plans and conduct a rapid evaluation: Plan A (Risk Avoidance): Temporarily halt infiltration, order drone #3 to utilize its more powerful electronic eavesdropping equipment to attempt to analyze the nature and location of the interference source, while simultaneously using its telephoto lens to conduct long-range observation of the area where drone #2 lost contact, awaiting communication restoration. Plan B (Mission Compensation): Assess its own position and sensor capabilities, evaluating whether, without significantly increasing risk, it can slightly adjust its flight path to attempt to observe a portion of the core area originally designated for drone #2's responsibility from the flank, thus partially achieving the reconnaissance objective. Option C (Role Reorganization): If there are spare reconnaissance units in the swarm (e.g., Unit 4), the sentinel can assess the feasibility of activating it and plan a completely new infiltration path that avoids known interference areas to take over the mission of Unit 2. This decision-making process is distributed and may be led by the sentinel or jointly calculated and evaluated by multiple units, ultimately selecting the optimal option through a rapid consensus mechanism.

[0068] In generating repair solutions, the underlying algorithms (such as rule engines and search algorithms) fall within the scope of existing technologies. However, this application creates an intent-driven framework that starts with tactical roles and high-level intents, using a cost function to guide the search direction of the planning algorithm. This framework seamlessly integrates multiple AI technologies (rules, planning, constraint solving) to achieve fast, autonomous, and tactically sound online decision-making. The cost function is a multi-objective dynamic evaluation system, specifically: Highest Priority: To what extent can the new solution continue to accomplish the original top priority (e.g., "getting a clear picture of the situation")? Key constraint compliance: New side; risk assessment: Survival risk: What is the probability that implementing the new plan will result in other drones being detected or lost? Risk of mission failure: New approach; Resource consumption: Energy costs: New method; Time cost: How long will it take to complete the new plan? Will it miss the critical time window? Role Adaptability: Is it appropriate to assign a Sentinel an Assassin mission? Evaluate how well the drone performing the new mission matches its assigned "role" and its own abilities.

[0069] Workflow summary: When an emergency occurs, the system will generate multiple possible solutions (such as Solution A: all units return to base; Solution B: sentries take over some reconnaissance; Solution C: activate backup drones). The system will then use the aforementioned cost function to quickly score each solution, and the solution with the best score (e.g., the solution that maximizes the completion of the mission objective while adhering to the core constraints with acceptable risk and resource consumption) will be automatically selected as the optimal solution and executed.

[0070] 3) Smooth Autonomous Action Adjustment. Once the optimal plan (e.g., Plan B) is selected, Sentinel 1 will fine-tune its attitude and flight path, and its onboard gimbal will automatically turn to the new observation area. Simultaneously, it will notify Sentinel 3 to pay closer attention to the electromagnetic spectrum in that direction. The entire swarm's formation and behavior undergo dynamic adjustments, but these adjustments are not chaotic; rather, they are an orderly, self-healing process closely aligned with the initial tactical intent. Their goal never changes; only the paths and methods to achieve that goal evolve adaptively under the intelligent drive within the swarm. Through this intent-driven distributed replanning mechanism, the swarm evolves from a fragile chain structure into a resilient, highly resistant mesh-like intelligent agent. It can gracefully handle battlefield contingencies without human intervention, ensuring mission progresses to the maximum extent possible under various adversities. This endows the entire system with unprecedented robustness and autonomy, enabling it to steadfastly execute the operator's will even in the harshest environments.

[0071] Furthermore, after obtaining the collaborative planning scheme or performing distributed adaptive replanning for the drone swarm, the process also includes: Based on the results of intent-driven tasks, different areas of the target region are colored to achieve semantic coloring of the region; for example, the mountainous area that the bee colony is scouting will dynamically change color according to the progress of the task. Gray represents the unknown, light blue represents that no anomalies have been found after scanning, and if suspicious signs are found, the area will be flashing orange or red. The color intensity and flashing frequency directly correspond to the threat level assessment. Based on the progress of the intent-driven mission, the normal status of the drone swarm is displayed to achieve a story-like narrative of the mission progress; for example, the information panel on the interface will not display that drone A has reached waypoint 5, but will summarize it in concise natural language: the 'Ghost' team has completed 70% of the infiltration mission and is currently conducting reconnaissance on the south side of the canyon. The stealth assessment of the entire operation is 'high'. Based on emergencies and optimal solutions, the abnormal status of the drone swarm is displayed to facilitate the repair of critical events. For example, when an Assassin drone loses contact and triggers autonomous replanning, the system will not disturb the operator with jarring alarms and technical logs. Instead, it will wait for the initial adaptive adjustment within the swarm to complete before providing a calm but clear voice or text prompt: "Commander, drone number 2 has lost contact due to strong interference in area C. The swarm has autonomously adjusted, with drone number 1 compensating for part of the reconnaissance area from the flank. Preliminary assessment indicates that the mission objective can still be partially achieved, but intelligence in the core area is missing. Please provide instructions."

[0072] The aforementioned setup represents abstract information feedback, freeing operators from the burdensome role of data monitors. This allows them to focus their valuable attention entirely on their core tactical decision-making duties, maintaining a constant grasp of the overall battlefield picture and key moments. Traditional human-machine interfaces are often overwhelmed by massive amounts of raw data: real-time coordinates, speeds, altitudes, battery percentages, camera feeds, etc., of dozens of drones. These data streams not only quickly deplete the operator's cognitive resources, but more importantly, they are not directly equivalent to valuable tactical information. This solution's system, however, acts as a high-level intelligence analyst. It doesn't simply dump raw data on the operator; instead, it deeply processes, refines, and abstracts it, ultimately presenting the crucial tactical situation in the most intuitive and easily understood way.

[0073] Furthermore, after obtaining the collaborative planning scheme, it also includes: acquiring inquiry information, analyzing the natural language and gesture commands in the inquiry information, and focusing on the area based on the inquiry information; for example, when the system reports that Unit 2 has lost contact and there is a lack of intelligence, the operator can point to the C area marked in orange on the sand table and say directly: zoom in on this area and take over the picture that Unit 1 can see. The system will immediately understand this focusing intention, seamlessly switch the display content, and may add an explanation: this is the lowest resolution picture of Unit 1's extreme observation at 6 kilometers away; The system acquires new intent information, analyzes the natural language and gesture commands within it, and combines this with the current collaborative planning scheme to deduce alternative suggestions. For example, after observing a blurry image, an operator might form a new judgment and intent, saying, "I believe this interference source itself is the target. Change the task and prioritize identifying the location and nature of the interference source." This statement completes a fundamental evolution of intent, and the system immediately captures this significant shift from reconnaissance of a suspicious target to reverse-locating the interference source. Upon receiving a new intent, the system will not execute it blindly; it will immediately initiate a new round of tactical planning and may proactively propose alternatives. Suggestion: "Understood. The new mission is 'electronic countermeasures reconnaissance.' Based on available resources, the best option is to send Unit 3 (electronic support type) close in, but this is risky. The alternative is to guide long-range firepower to conduct a suppressive strike on the area. Which option would you like to simulate?" The operator can say: Simulate the success rate and risk of the first option. The system will run the simulation within seconds and visualize Unit 3's possible penetration paths, estimated success probability, and potential loss risks on the sand table. Based on this clear decision support information, the operator will ultimately decide: Execute Option 1, allowing Unit 1 to continue surveillance and provide cover. Based on regional focus and alternative suggestions, make dialogical intent adjustments.

[0074] The above settings belong to dialogic intent correction. Through this back-and-forth, iterative dialogue process, the initial vague intent to understand the situation is continuously clarified, corrected, and deepened in the continuous interaction between humans and machines, and eventually evolves into a series of highly specific and highly adaptable tactical actions.

[0075] The aforementioned abstract information feedback and conversational intent correction establish an elegant and powerful human-machine interaction loop, ensuring seamless integration of human strategic wisdom and machine tactical efficiency, forming an enhanced decision-making entity where 1+1 is far greater than 2. Human intuition, experience, and strategic foresight are perfectly coupled with the machine's powerful computational, simulation, and autonomous execution capabilities within this loop. This continuously collaborative loop not only guarantees the success of individual tasks but also constitutes a learning system itself. The system records each human-machine collaborative decision-making process, analyzing which decisions the operator tends to make under different circumstances, thus providing more personalized and predictive suggestions for the future. The operator also continuously improves their understanding of swarm capabilities and battlefield situations through interaction with the system. Humans and machines are no longer in a master-slave relationship but rather tactical partners who grow together in the face of challenges. The construction of this loop overturns the information overload mode of traditional monitoring interfaces, replacing it with two innovative interaction paradigms: abstract information feedback and conversational intent correction. Based on the above-mentioned abstract feedback, the interaction between humans and machines has entered a higher-dimensional dialogue level. Operators can now make precise and efficient secondary interventions based on the insights extracted by the system. This intervention is not a micro-management of the underlying actions, but a continuous injection and evolution of the top-level intentions.

[0076] The aforementioned control method for drone swarm intent-oriented tasks breaks through existing technological bottlenecks by constructing an intent-oriented human-machine interaction and collaborative decision-making system. This frees operators from tedious micro-management, allowing them to focus on high-level strategic decision-making. At the same time, it endows the swarm system with a higher level of autonomy and intelligence, realizing true intelligent collaboration between humans and drone swarms, thereby fully unleashing the enormous potential of the swarm system in future applications.

[0077] Specifically, the following technical problems were solved: 1) Break down interaction barriers and establish cognitive channels: The core is to solve the problem of machines' deep understanding of human ambiguous intentions, overcome cross-modal (gesture, language) semantic fusion, dynamic context perception and intention reasoning based on domain knowledge graphs, and enable the system to accurately understand the operator's strategic intentions and the constraints and expectations behind them, just like a human advisor.

[0078] 2) Empowering bee colonies with intelligence to achieve autonomous emergence: The core is to solve the problem of generative autonomous decision-making and collaboration of bee colonies based on intent. It has the ability to dynamically transform abstract intents into specific, heterogeneous collaborative strategies, and can perform adaptive replanning based on shared intents in a distributed environment. The ultimate goal is to evolve bee colonies from passive tools into intelligent partners that can perform creative tasks under human guidance.

[0079] It has the following beneficial effects: 1) A cross-modal tactical intent parsing and fusion system was designed, which can not only identify commands, but also understand intents. It can simultaneously analyze the dynamic features of gestures (such as speed and force) and the deep semantics of natural language (entities, constraints, and emotions). It can eliminate cross-modal input ambiguity and use tactical knowledge graphs for semantic enrichment, and finally generate structured intent objects.

[0080] 2) A generative swarm collaborative strategy planning was designed, which can create tactics instead of calling scripts. It can autonomously allocate heterogeneous roles (such as sentinels and assassins) based on the input tactical intention objects and real-time dynamic constraints (swarm state, environment). It can generate spatiotemporally synchronized and functionally complementary four-dimensional collaborative paths, and the generated collaborative planning scheme includes contingency plans.

[0081] 3) An intent-driven distributed adaptive and human-machine collaborative interaction mechanism was designed, which realizes the autonomous resilience of the bee colony and the deep dialogue between humans and machines. The high-level intent is pushed down to the individual bee colony, enabling it to have distributed autonomous replanning capabilities. It extracts massive amounts of raw data into abstract and semantic tactical situations and has a closed-loop collaborative mechanism that supports operators to evolve intents through natural interaction and provides decision support such as wargaming simulations.

[0082] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0083] This application also provides a control device for unmanned aerial vehicle (UAV) swarm intent-guided missions, such as... Figure 2 As shown, in one embodiment, it includes: a first module 201, a second module 202, a third module 203, and a fourth module 204, wherein: The first module 201 is used to acquire instruction information for intent-oriented tasks, analyze the natural language and gesture instructions in the instruction information, obtain dynamic parameters, and generate a preliminary spatial intent vector. The second module 202 is used to acquire the knowledge graph, combine it with the language parsing engine to parse the preliminary spatial intent vector, obtain keywords, and generate intent data packages; The third module 203 is used to obtain the state of the drone swarm, construct internal and external constraints, and obtain the dynamically changing environment. The fourth module 204 is used to assign roles to each drone in the drone swarm based on the intent data packet and the environment, and plan the corresponding sub-tasks for different roles; plan the collaborative behavior path for the drone swarm based on different roles and corresponding sub-tasks, and generate a four-dimensional action sequence; and obtain a collaborative planning scheme based on different roles, corresponding sub-tasks and four-dimensional action sequence to achieve the control of the drone swarm.

[0084] Specific limitations regarding the control device for UAV swarm intent-guided tasks can be found in the limitations of the control method for UAV swarm intent-guided tasks described above, and will not be repeated here. Each module in the aforementioned device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0085] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a control method for an unmanned aerial vehicle (UAV) swarm intent-guided mission. The display screen can be an LCD screen or an e-ink display screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0086] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0087] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0088] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0089] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0090] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0091] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0092] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended application documents.

Claims

1. A control method for unmanned aerial vehicle (UAV) swarm intent-oriented tasks, characterized in that, include: Obtain instruction information for intent-driven tasks, analyze the natural language and gesture instructions in the instruction information, obtain dynamic parameters, and generate a preliminary spatial intent vector; The knowledge graph is acquired and, combined with a language parsing engine, the initial spatial intent vector is parsed to obtain keywords and generate intent data packages. Obtain the state of the drone swarm, construct internal and external constraints, and obtain the dynamically changing environment; Based on the intent data packet and the environment, assign roles to each drone in the drone swarm and plan corresponding sub-tasks for different roles; Based on different roles and corresponding sub-tasks, a collaborative behavior path is planned for the drone swarm, generating a four-dimensional action sequence; based on different roles, corresponding sub-tasks, and the four-dimensional action sequence, a collaborative planning scheme is obtained to achieve the control of the drone swarm.

2. The control method for unmanned aerial vehicle swarm intent-guided missions according to claim 1, characterized in that, After obtaining the collaborative planning solution, the following is also included: The drones begin executing their missions according to the collaborative planning scheme. When an emergency occurs, the system identifies the failed nodes and the valid missions of the relevant drones, and generates different repair plans from the perspectives of different roles. Select one of the repair options as the optimal solution and perform distributed adaptive replanning for the drone swarm.

3. The control method for unmanned aerial vehicle swarm intent-guided missions according to claim 2, characterized in that, After obtaining the collaborative planning scheme or performing distributed adaptive replanning for the drone swarm, the following are also included: Based on the results of the intent-driven task, different regions of the target area are colored to achieve semantic coloring of the regions; Based on the progress of the intent-driven task, the normal status of the drone swarm is displayed to achieve a story-like narrative of the task progress. Based on emergencies and optimal solutions, the abnormal status of the drone swarm is displayed to enable the repair of critical events.

4. The control method for unmanned aerial vehicle swarm intent-guided missions according to claim 3, characterized in that, After obtaining the collaborative planning solution, the following is also included: Acquire query information, analyze the natural language and gesture commands in the query information, and focus on the region based on the query information; Acquire new intent information, analyze the natural language and gesture commands in the new intent information, and combine them with the current collaborative planning scheme to deduce alternative suggestions; Based on regional focus and alternative suggestions, make dialogical intent adjustments.

5. The control method for unmanned aerial vehicle swarm intent-guided missions according to claim 4, characterized in that, After generating the four-dimensional action sequence, the following is also included: Different emergency plans are generated based on different unexpected situations; Based on different roles, corresponding sub-tasks, four-dimensional action sequences, and emergency plans, a collaborative planning scheme is obtained to achieve the control of the drone swarm.

6. The control method for unmanned aerial vehicle swarm intent-guided missions according to any one of claims 1 to 5, characterized in that, The initial spatial intent vector includes: geographic location, range, and initial weights related to the mode of action, which includes: speed and stealth.

7. The control method for unmanned aerial vehicle swarm intent-guided missions according to any one of claims 1 to 5, characterized in that, The intent data packet includes: target level, key constraints, execution resources, geometric and environmental features of the work area, and a confidence score for comprehensive evaluation; The target hierarchy includes primary and secondary targets, and the key constraints include concealment and timeliness.

8. A control device for unmanned aerial vehicle (UAV) swarm intent-guided missions, characterized in that, include: The first module is used to acquire instruction information for intent-driven tasks, analyze the natural language and gesture instructions in the instruction information, obtain dynamic parameters, and generate a preliminary spatial intent vector. The second module is used to acquire the knowledge graph, combine it with the language parsing engine to parse the initial spatial intent vector, obtain keywords, and generate intent data packages; The third module is used to obtain the state of the drone swarm, construct internal and external constraints, and obtain the dynamically changing environment. The fourth module is used to assign roles to each drone in the drone swarm based on the intent data packet and the environment, and to plan the corresponding sub-tasks for different roles. Based on different roles and corresponding sub-tasks, a collaborative behavior path is planned for the drone swarm, generating a four-dimensional action sequence; based on different roles, corresponding sub-tasks, and the four-dimensional action sequence, a collaborative planning scheme is obtained to achieve the control of the drone swarm.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images