Unmanned aerial vehicle cluster task dynamic allocation method and system based on heterogeneous capability modeling and semantic evaluation

Abstract

The invention relates to the technical field of unmanned aerial vehicle cluster cooperative control, and discloses an unmanned aerial vehicle cluster task dynamic allocation method and system based on heterogeneous capability modeling and semantic evaluation. Modeling and quantifying the heterogeneous capability of the unmanned aerial vehicle from three aspects of sensing capability, maneuvering capability and cruising capability; environment images are collected based on a sensor, the semantic correlation between each sub-region and a task target is evaluated through a visual language alignment mechanism, and the exploration value of each region is calculated in combination with the semantic integrating degree, the obstacle density and the path accessibility; modeling multi-unmanned aerial vehicle task allocation as a path optimization problem with capacity constraint, and establishing a mapping relationship between unmanned aerial vehicle capabilities and task demands; and constructing task clusters based on the unmanned aerial vehicle capability score and the regional exploration value, generating a preliminary task allocation scheme by adopting a saving algorithm, and triggering dynamic task reallocation when the unmanned aerial vehicle capability or environmental semantics changes. According to the invention, intelligent task allocation and efficient collaborative operation of the unmanned aerial vehicle cluster in a complex environment are realized.

Description

Technical Field

[0001] This invention belongs to the field of intelligent robot autonomous navigation and multi-agent cooperative control technology, and in particular, it is a method and system for dynamic allocation of unmanned aerial vehicle (UAV) swarm tasks based on heterogeneous capability modeling and semantic evaluation. Background Technology

[0002] With the continuous development of drone technology, drones have been widely used in disaster relief, environmental monitoring, and resource exploration. Autonomous drone exploration refers to the process by which drones, without prior maps or human intervention, independently complete environmental modeling, target identification, and information collection through perception, localization, and path planning. However, single drones are limited in terms of perception range, endurance, and computing power, making it difficult to efficiently complete exploration tasks in complex or large-scale environments.

[0003] To improve mission efficiency, multi-UAV collaborative exploration technology has gradually become a research hotspot. Existing methods typically achieve task division through mission area partitioning, frontier point allocation, or information gain-based planning strategies. While these methods have achieved some success in improving overall coverage, they still have several shortcomings in practical applications.

[0004] First, existing collaborative exploration methods are mostly based on the homogeneity assumption, which assumes that all UAVs have the same performance in terms of perception, maneuverability, and endurance. Some schemes allocate tasks through task location matching or auction algorithms, but do not consider individual performance differences. Due to differences in payload configuration, sensor type, and energy state, the execution capabilities of different UAVs often vary significantly. If the task allocation is unbalanced, some UAVs may be overloaded and their missions may be interrupted, while others may be underloaded and their resources may be underutilized, thereby reducing the overall efficiency of the system.

[0005] Secondly, most methods lack effective analysis and utilization of environmental semantic information. Existing exploration algorithms mostly rely on geometric features or information gain metrics to assess the importance of regions, while giving less consideration to the semantic hierarchical features of the scene. For example, doorways, passageways, or gathering areas in a disaster site usually have higher information value and task priority, but algorithms based on geometry or information gain struggle to identify the semantic differences in these areas, easily leading to excessive resource allocation to secondary areas and insufficient allocation to key areas.

[0006] Beyond this, while some studies have attempted to introduce semantic-driven mechanisms, most have focused on single-machine platforms or static models. Existing methods often use semantic information for local path selection or map annotation, lacking comprehensive application at the task planning and allocation level. Such solutions struggle to combine semantic understanding with cluster collaboration, and the system remains at the level of local perception, unable to perform global optimization based on task objectives.

[0007] Furthermore, existing systems generally lack the ability to dynamically adjust task allocation strategies. Most algorithms employ fixed region partitioning or path planning models based on the traveling salesman problem, making it difficult to adapt to changes in UAV status and environmental information during task execution, resulting in untimely resource scheduling and suboptimal path exploration. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a method and system for dynamic task allocation in UAV swarms based on heterogeneous capability modeling and semantic evaluation. The method introduces two types of factors simultaneously during the task planning process: capability modeling and semantic value guidance. It can adjust task allocation in real time according to changes in UAV status and updates in environmental semantic information, enabling the system to maintain high resource utilization and information acquisition efficiency in complex scenarios.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] According to a first aspect of the present invention, a method for dynamic task allocation in UAV swarms based on heterogeneous capability modeling and semantic evaluation is provided, comprising the following steps:

[0011] Step 1: Heterogeneous capability modeling and quantification; Model and quantify the heterogeneous capabilities of the UAV swarm from three aspects: perception capability, maneuverability and endurance, and give a comprehensive capability score for each UAV.

[0012] Step 2: Semantic Value Assessment and Region Ranking; Environmental perception data is collected using UAV sensors to divide the environment into regions. The semantic relevance of each region to the task objective is analyzed based on a visual-language alignment mechanism. The exploration value of each region is calculated by combining obstacle density and path accessibility, resulting in a region value ranking.

[0013] Step 3: Constrained optimization modeling steps: With the goal of minimizing the total flight time, the task allocation problem of multiple UAVs is modeled as a path optimization problem with capacity constraints (Capacitated Vehicle Routing Problem, or CVRP for short), in which the UAV capability score is used as the capacity constraint and the area exploration value is used as the demand constraint;

[0014] Step 4: Optimize the solution and dynamically reallocate tasks; construct task clusters based on UAV capability scores and area exploration value, generate initial paths using a cost-saving algorithm, and obtain initial task allocation schemes through local optimization; when UAV capability parameters or task area semantic information change, the system triggers dynamic task reallocation.

[0015] Furthermore, in step 1, the drone is evaluated. Overall performance, its perception capabilities Mobility and battery life Perform a weighted geometric average; let the weights of each capability be respectively... , and And satisfy drones The formula for calculating the comprehensive ability score is as follows:

[0016]

[0017] Furthermore, the perceptual ability in step 1 Horizontal field of view based on sensor Vertical field of view With maximum sensing distance Quantify at least one of the following: mobility Based on maximum linear velocity Maximum linear acceleration Maximum angular velocity With maximum angular acceleration At least one of the four dynamic parameters is quantified; range Quantification is based on a combination of energy reserves and flight energy consumption trends.

[0018] The steps for constructing a perceptual ability scoring model are as follows:

[0019] Let the vertex of the square pyramid be the sensor position, the base be a rectangle, and the distance from the vertex to the base be D; therefore, the length and width of the base are:

[0020]

[0021] Therefore, the field volume can be obtained:

[0022]

[0023] Based on cluster reference volume Based on this, the scores are normalized as follows:

[0024]

[0025] in, Indicates perceptual ability score The horizontal and vertical field of view of the sensor are in radians. Maximum sensing distance; It can be set by the cluster baseline configuration or by statistical analysis of typical scenarios;

[0026] The overall maneuverability scoring model for drones is constructed as follows:

[0027] Maneuverability is divided into two dimensions: translational capability and steering capability, which respectively reflect the UAV's linear motion response and attitude adjustment capability in space.

[0028] Translational capability takes into account the time required for a drone to accelerate from a standstill to its maximum speed. Compare it with the reference value Normalize the ratio to obtain the translational capability. .

[0029] Steering capability corresponds to attitude adjustment speed, and is approximated by the time required to reach maximum angular velocity. The formula for calculating steering response time is: Compare it with the reference value Normalize the ratio to obtain the steering capability. .

[0030] The final maneuverability score is expressed as the geometric mean of translational and steering capabilities:

[0031]

[0032] in, This represents the overall maneuverability score of the drone, a reference value. Data can be taken from a drone swarm baseline platform or historical samples. This scoring method is derived based on response time, avoiding subjective weighting.

[0033] The steps for constructing the battery life rating model are as follows:

[0034] Introducing the derivative of acceleration, jerk, the formula is as follows: (unit ), and defined within the task time interval. Cumulative fluctuations within:

[0035]

[0036] Considering the current battery state, let For remaining battery power, If the battery is at full capacity, then the percentage of remaining charge is: To eliminate the dimensionlessness and suppress the range uncertainty caused by violent maneuvers, a reference constant is introduced. The battery life rating is defined as follows:

[0037]

[0038] in, Calibration can be performed using standard operating conditions or historical mission data.

[0039] Furthermore, environmental images are acquired using an RGB-D sensor, and semantic segmentation is performed. A text-guided object detection model is then employed, using natural language prompts to detect task-related objects in the images, resulting in a target set.

[0040]

[0041] Each object It contains category tags and location information.

[0042] Subsequently, using an image segmentation model, contour segmentation was performed on the detection results to obtain object mask information. Based on spatial proximity and semantic feature correlation, the environment was clustered to obtain several semantically consistent sub-regions.

[0043]

[0044] Each sub-region It corresponds to a function or semantic scenario.

[0045] The visual language alignment mechanism employs a visual language model. The cluster master node receives the task instruction and distributes it to each UAV to perform semantic matching calculations. The UAVs utilize the visual language model to perform semantic matching calculations on the regional images. and task text prompts Perform feature encoding to obtain image feature vectors. With text feature vectors The cosine similarity between the two is calculated in a unified semantic embedding space.

[0046]

[0047] income Indicates the area The degree of matching with the task semantics is used to linearly normalize the results:

[0048]

[0049] Normalized semantic fit A higher value indicates a better semantic fit between the region and the task. Furthermore, in step 2, the comprehensive exploration value of each region is calculated based on three indicators: semantic fit, obstacle density, and path reachability.

[0050] Obstacle density is defined as follows:

[0051]

[0052] in, Indicates the area The number of barrier voxels within, Indicates the area The total number of elements.

[0053] Path reachability Indicates the area Accessibility is used to measure how well the area is accessible to other key locations; first, the area is calculated on the topological map. Average shortest path cost to all critical nodes Then, its reachability is defined as: ,

[0054] in To prevent the small constant from being divided by zero, further... Normalization is performed:

[0055]

[0056] Normalized path reachability The larger the value, the easier the area is to be accessed by drones;

[0057] The final exploration value score of a region is defined as follows:

[0058]

[0059] in, , and Representing the normalized semantic fit, obstacle density, and path reachability, respectively, the weight parameters are: satisfy ;

[0060] Comprehensive Exploration Value of Each Region The regions are sorted to form a value map for regional exploration.

[0061] Furthermore, step 3 models the multi-UAV task allocation problem as a capacity-constrained path optimization problem (CVRP), specifically including:

[0062] Let the set of drones be Each drone The overall ability score is recorded as the capacity limit. ;

[0063] Let the set of exploration regions be Each region Corresponding exploration value This indicates the task demand in that area;

[0064] The system introduces node r0 as the unified start and end point for all UAV paths. A centralized take-off and landing area or supply center for corresponding UAVs, used to achieve unified scheduling and closed-loop execution of missions;

[0065] Define a binary variable When drones From the region Fly to the area hour, Otherwise, it is 0; the flight time between regions is defined as:

[0066]

[0067] in, The distance between regions. The maximum speed of the drone, This is an environmental correction factor used to account for the effects of factors such as wind speed and obstacles.

[0068] Furthermore, the specific steps for optimizing the CVRP problem in step 4 are as follows:

[0069] Assume there is a total The drones were deployed, and the mission points were assembled at the following locations: The exploration value of each task point is The capability rating of each drone is: Path cost Indicates the drone departs from the mission point Arrive at the mission point The overall flight cost is calculated as follows. The optimization objective is to minimize the overall cost function, which is expressed as follows, under the condition that each mission point is visited only once and the load of each UAV does not exceed its capacity constraints:

[0070]

[0071] in, For drones The set of access paths must satisfy the following constraints. .

[0072] First, tasks are clustered, and the system determines the exploration value of each task point. With drone capability rating The ratio of the value density of task points is used to categorize task points according to their value density. The task points are sorted and divided into several clusters based on geographical proximity and load balancing principles, with each cluster corresponding to the task range of one drone. Then, path construction is performed: for each task cluster, starting from the drone's initial position, an initial path is generated using a cost-saving algorithm. The algorithm first assumes that each task point is independently round-tripped, and then calculates the cost savings after merging the task points. The paths are merged in descending order of savings until the capacity constraints are met. Finally, path optimization is performed. The system executes local adjustments to the initial path, including node swapping, node insertion, and path reversal. Node swapping is used to balance the load across different paths, node insertion is used to optimize the access order within the same path, and path reversal is used to reduce redundant tracks. The cost change is calculated for each operation. ,like If so, then accept the new plan. Through multiple iterations, the initial task allocation scheme is obtained when the total cost stabilizes or the maximum number of iterations is reached. .

[0073] Furthermore, in step 4, task reassignment is judged from three aspects: changes in UAV capabilities, changes in regional semantics, and task execution deviations. When any one of the three conditions meets the threshold requirement, the system automatically starts the task reassignment module to resolve and adjust the affected task points and UAV paths.

[0074] In response to changes in drone capabilities, the system monitors the capability score of each drone in real time. When the current ability value is significantly lower than the initial ability value When this occurs, it is considered that the drone's payload capacity or endurance has decreased; the degree of change in capability is expressed as follows:

[0075]

[0076] when When the system determines that the drone's capability has declined beyond a preset threshold, it triggers a task reassignment; among these, The threshold for capability changes can be set based on factors such as the characteristics of the drone;

[0077] In response to changes in the semantic value of a region, the system dynamically updates the exploration value of task points based on environmental perception results. When the real-time exploration value of a task point differs significantly from its initial value, it indicates a significant change in the importance, accessibility, or risk factors of that area. The magnitude of semantic change can be defined as:

[0078]

[0079] when When this occurs, the system determines that the semantic features of the task region have changed significantly, thus triggering task reallocation. The semantic change threshold is determined by the task scenario parameters;

[0080] To address mission execution deviations, the system compares the actual flight costs of the drone. Cost of Model Prediction If the two deviate significantly, it indicates that the path planning result does not match the actual execution. The execution deviation is defined as:

[0081]

[0082] when When the system determines that the current task allocation scheme is no longer applicable, it enters the task reallocation phase. The deviation threshold can be set according to the accuracy requirements of the task;

[0083] Once the triggering condition is met, the system enters the dynamic task reallocation phase. An incremental optimization strategy is employed to adjust the affected task nodes and related UAV paths. Specific steps include: first, freezing the currently executing path segments and including only the incomplete parts in the replanning set; second, calculating the incremental cost of reinserting the affected task points into different UAV paths. The system selects the insertion combination that minimizes the overall cost and performs the update accordingly; then, it sequentially performs node migration and path correction operations to generate a new task allocation scheme. Finally, when the difference in total cost between the old and new solutions satisfies At that point, the plan was considered to have reached a stable state, and no further adjustments were made.

[0084] Another aspect of the present invention provides a dynamic task allocation system for UAV swarms based on heterogeneous capability and semantic evaluation, including a UAV capability modeling module, an environment perception module, an environment semantic value evaluation module, a constraint optimization modeling module, and a dynamic reallocation and task execution module. The above modules of the system are used to perform the method described in any one of claims 1-9.

[0085] The UAV capability modeling module is used to establish a comprehensive capability evaluation model for individual UAVs from three dimensions: perception capability, maneuverability, and endurance capability, and generate a comprehensive capability score for each UAV.

[0086] The environmental perception module acquires RGB images, depth maps, and IMU data collected by sensors on the UAV. The RGB images provide texture information about the environment to support subsequent target detection and semantic segmentation; the depth maps reflect the spatial structure of the scene, used to identify obstacle locations and passable areas; and the IMU data is used to calculate the UAV's attitude and motion state. By fusing information from these multi-source sensors, the environmental perception module forms fundamental perception data describing the environment's appearance and spatial structure, providing reliable environmental input for subsequent semantic value assessment and task allocation.

[0087] The environmental semantic value assessment module includes an environmental analysis unit and a visual-language alignment unit. The environmental analysis unit receives data from the environmental perception module and performs target detection, semantic segmentation, and semantic assessment operations. The visual-language alignment unit calculates the semantic relevance between environmental regions and task objectives. The environmental semantic value assessment module combines semantic fit, obstacle density, and path reachability to calculate the comprehensive exploration value of each region. Sorting is performed to create a regional exploration value map;

[0088] The constrained optimization modeling module is used to model the allocation of multiple UAV missions as a path optimization problem with capacity constraints. The UAV capability score serves as the capacity constraint, the area exploration value serves as the demand constraint, and the optimization objective is to minimize the total flight time.

[0089] The dynamic redistribution and task execution module includes an initial allocation unit and an optimization adjustment unit. The initial allocation unit is used to construct task clusters based on UAV capability scores and regional exploration value, and generate an initial task allocation scheme. The optimization adjustment unit is used to locally optimize the initial scheme according to path costs and constraints. When changes are detected in UAV capability parameters or task area semantic information, dynamic task redistribution is triggered to achieve real-time optimization and global balance of task planning. The optimized task allocation scheme is distributed to each UAV terminal for execution by the task execution unit, and the task execution status is monitored and feedback is provided.

[0090] Compared with the prior art, the present invention has the following beneficial effects:

[0091] First, this invention quantifies the heterogeneous performance of UAVs by establishing a three-dimensional capability model encompassing perception, maneuverability, and endurance, and matches tasks based on comprehensive capability scores. This ensures that task load is coordinated with platform capabilities, avoiding situations where some UAVs are overloaded while others are idle, thereby improving resource utilization and overall system efficiency.

[0092] Secondly, this invention utilizes a visual language model to perform semantic understanding of environmental images, identify regions with task-relevant significance, and comprehensively evaluate the value of these regions by combining semantic fit, obstacle density, and path accessibility. By introducing semantic hierarchical information, the system can prioritize exploration based on the semantic importance of the task objective, avoiding problems such as unidentified target regions or unreasonable resource allocation.

[0093] Furthermore, this invention does not rely on preset semantic categories or heuristic rules, but achieves cross-modal semantic reasoning through a visual language model. It can autonomously identify potentially high-value areas in unknown environments and directly integrate semantic results into the task planning and allocation process, thereby achieving true semantic-driven collaborative exploration.

[0094] Finally, this invention models task allocation as a path optimization problem with capacity constraints, simultaneously considering UAV capabilities and regional semantic value during the optimization process to achieve dynamic updates and global optimization of task planning. The system can adjust the allocation results based on real-time information, making the exploration process more flexible and efficient, and avoiding the resource scheduling lag problem in traditional methods.

[0095] In summary, this invention has made systematic improvements in UAV capability modeling, environmental semantic understanding, in-depth utilization of semantic information, and task allocation optimization, enabling multi-UAV swarms to achieve more reasonable resource allocation and more efficient collaborative exploration in complex environments. Attached Figure Description

[0096] Figure 1 This is a method and system flowchart for dynamic task allocation in UAV swarms based on heterogeneous capability modeling and semantic evaluation.

[0097] Figure 2 This is a framework diagram of a UAV swarm task dynamic allocation system based on heterogeneous capabilities and semantic evaluation.

[0098] Figure 3 This is a schematic diagram of UAV perception capability modeling.

[0099] Figure 4 This is a schematic diagram of semantic value assessment and regional ranking.

[0100] Figure 5 The simulation scene used in the experiment and the 3D map generated by the method of this invention Detailed Implementation

[0101] Example 1

[0102] Combination Figure 1-4 The present invention proposes a method for dynamic task allocation in UAV swarms based on heterogeneous capabilities and semantic evaluation, comprising the following steps:

[0103] Step 1: Heterogeneous capability modeling and quantification; Model and quantify the heterogeneous capabilities of the UAV swarm from three aspects: perception capability, maneuverability and endurance, and give a comprehensive capability score for each UAV.

[0104] Step 2: Semantic Value Assessment and Region Ranking; Environmental perception data is collected using UAV sensors to divide the environment into regions. The semantic relevance of each region to the task objective is analyzed based on a visual-language alignment mechanism. The exploration value of each region is calculated by combining obstacle density and path accessibility, resulting in a region value ranking.

[0105] Step 3: Constrained optimization modeling: With the goal of minimizing the total flight time, the task allocation problem of multiple UAVs is modeled as a path optimization problem with capacity constraints, where the UAV capability score is used as the capacity constraint and the area exploration value is used as the demand constraint.

[0106] Step 4: Optimize the solution and dynamically reallocate tasks; construct task clusters based on UAV capability scores and area exploration value, generate initial paths using a cost-saving algorithm, and obtain initial task allocation schemes through local optimization; when UAV capability parameters or task area semantic information change, the system triggers dynamic task reallocation.

[0107] The system of this invention mainly comprises four functional modules: a UAV capability modeling module, an environmental perception module, a constraint optimization modeling module, and a dynamic reallocation and task execution module. Compared with existing multi-UAV collaborative exploration methods, this invention introduces both capability constraints and semantic value guidance during the task planning process. It can adjust task allocation in real time based on changes in UAV status and updates to environmental semantic information, enabling the system to maintain high resource utilization and information acquisition efficiency in complex scenarios. This makes it suitable for applications such as disaster relief, resource surveying, and target search.

[0108] Step 1: Heterogeneous capability modeling and quantification; Model and quantify the heterogeneous capabilities of the UAV swarm from three aspects: perception capability, maneuverability and endurance, and give a comprehensive capability score for each UAV.

[0109] This embodiment quantifies the heterogeneous capabilities of a UAV swarm through multi-dimensional capability modeling, providing a quantitative basis for subsequent task allocation. Capability modeling is conducted from three aspects: perception capability, maneuverability, and endurance capability, and a capability score for each UAV is generated through comprehensive analysis. To unify the units of measurement, each capability score in this section is first normalized to [0,1], and then a weighted geometric average is performed to obtain the comprehensive score. Through this capability model, the system can rationally allocate task loads based on the performance differences of each platform, achieving balanced resource utilization and improving overall exploration efficiency in a heterogeneous swarm.

[0110] The following details the evaluation methods for each dimension and the overall score calculation process.

[0111] 1.1.1 Perceptual ability

[0112] Sensing capability measures the range and accuracy of environmental information acquired by an unmanned aerial vehicle (UAV) in space exploration missions. To uniformly describe the sensing performance of heterogeneous platforms, this invention employs three typical parameters: horizontal field of view... Vertical field of view With maximum sensing distance A perception capability scoring model was constructed. These parameters are applicable to depth cameras, LiDAR, and multimodal sensors, and have good versatility and availability.

[0113] The field of view of a common sensor can be approximated as a square pyramid or a cone. This embodiment uses a square pyramid approximation, with the vertex of the pyramid representing the sensor position. Figure 3 As shown, the base is a rectangle, and the distance from the vertex to the base is... Therefore, the length and width of the base are:

[0114]

[0115] Therefore, the field volume can be obtained:

[0116]

[0117] Based on cluster reference volume Based on this, the scores are normalized as follows:

[0118]

[0119] in, Indicates a rating of perceptual ability. The horizontal and vertical field of view of the sensor (in radians). This represents the maximum sensing distance (in meters). If a fan-shaped / conical field of view is used, the corresponding volume formula can be used for calculation. The scoring interface remains unchanged. It can be configured based on a cluster baseline or statistically set for typical scenarios. Based on field-of-view volumetric quantization sensing capabilities, this model enables objective comparisons of heterogeneous platforms without relying on specific sensor models.

[0120] 1.1.2 Mobility

[0121] Maneuverability measures an unmanned aerial vehicle's ability to respond to mission changes, dynamically avoid obstacles, and rapidly deploy, directly impacting its operational stability and adaptability in complex environments. This implementation selects the maximum linear velocity. Maximum linear acceleration Maximum angular velocity With maximum angular acceleration Four dynamic parameters were used to construct a maneuverability scoring model applicable to heterogeneous platforms.

[0122] Maneuverability is divided into two dimensions: translational capability and steering capability, which respectively reflect the UAV's linear motion response and attitude adjustment capability in space.

[0123] Translational capability takes into account the time required for a drone to accelerate from a standstill to its maximum speed. Compare it with the reference value Normalize the ratio to obtain the translational capability. .

[0124] Steering capability corresponds to attitude adjustment speed, and is approximated by the time required to reach maximum angular velocity. The formula for calculating steering response time is: Compare it with the reference value Normalize the ratio to obtain the steering capability. .

[0125] The final maneuverability score is expressed as the geometric mean of translational and steering capabilities:

[0126]

[0127] in, This represents the overall maneuverability score of the drone, a reference value. The data can be taken from a baseline platform or historical samples of the drone swarm. This scoring method is based on response time derivation, avoiding subjective weight setting. Furthermore, this model relies only on basic dynamic performance parameters, making it suitable for performance evaluation of different types of drone platforms.

[0128] 1.1.3 Battery life

[0129] Endurance measures a drone's ability to operate continuously during mission execution, and is influenced by both its current energy reserves and dynamic energy consumption during flight. Unlike static energy estimation, this implementation method incorporates energy consumption fluctuations during flight to construct a more dynamic and accurate endurance scoring model.

[0130] During flight, frequent acceleration, deceleration, and path changes cause significant fluctuations in motor output power, thus accelerating energy consumption. To characterize this energy consumption fluctuation trend, the derivative of acceleration, jerk, is introduced, and the calculation formula is as follows: (unit ), and defined within the task time interval. Cumulative fluctuations within:

[0131]

[0132] Considering the current battery state, let For remaining battery power, If the battery is at full capacity, then the percentage of remaining charge is: To eliminate the dimensionlessness and suppress the range uncertainty caused by violent maneuvers, a reference constant is introduced. The battery life rating is defined as follows:

[0133]

[0134] in, Calibration can be performed using standard operating conditions or historical mission data. For example, it can be based on energy consumption fluctuations measured during typical missions. The statistical analysis results are used, and the representative mean or median of the range is taken as the calibration reference. This model combines energy margin with flight energy consumption trend, which can effectively identify platforms with strong endurance and stable flight, and is suitable for quantitative evaluation of the performance stability of UAV platforms in dynamic mission environments.

[0135] 1.2 Comprehensive Ability Assessment

[0136] To evaluate drones Overall performance, its perception capabilities Mobility and battery life Perform a weighted geometric mean. Let the weights of each capability be as follows: , and And satisfy Then the drone The formula for calculating the comprehensive ability score is as follows:

[0137]

[0138] Because the geometric mean formula is sensitive to weaker metrics, this method avoids overestimating the performance of drones that excel in only a few areas. The weighting parameters are automatically calculated by the mission-type adaptive optimization module, requiring no manual setting. For example, improving performance in long-endurance missions... In tasks primarily based on perception, improving .

[0139] Step 2: Semantic Value Assessment and Region Ranking; Environmental perception data is collected using UAV sensors to divide the environment into regions. The semantic relevance of each region to the task objective is analyzed based on a visual-language alignment mechanism. The exploration value of each region is calculated by combining obstacle density and path accessibility, resulting in a region value ranking.

[0140] This embodiment proposes a regional semantic value assessment method based on a visual language alignment mechanism. This method is used to quantify the semantic relevance and explorable value of each sub-region of the environment in autonomous exploration tasks of UAV swarms, thereby providing a quantitative basis for swarm task allocation. Figure 3 As shown, the method includes the following three steps: semantic segmentation and region division, semantic fit calculation, and region exploration value calculation.

[0141] 2.1 Semantic Segmentation and Region Division

[0142] Each drone in the cluster acquires environmental images using its onboard RGB-D sensors and performs semantic segmentation. Specifically, a text-guided object detection model can be used, such as, but not limited to, the Grounding DINO model. By inputting natural language prompts, the model detects task-related objects in the images, obtaining a target set.

[0143]

[0144] Each object It contains category tags and location information.

[0145] Subsequently, image segmentation models, such as but not limited to the SAM model, are used to perform contour segmentation on the detection results to obtain object mask information. Based on spatial proximity and semantic feature correlation, the environment is clustered to obtain several semantically consistent sub-regions.

[0146]

[0147] Each sub-region It corresponds to a function or semantic scenario, such as a kitchen, office, or storage room.

[0148] 2.2 Semantic fit calculation

[0149] To assess the semantic relevance of each sub-region to the task objective, this embodiment introduces a visual language alignment mechanism. The cluster master node receives the task instruction (e.g., "Find the burning gas cylinder") and distributes the prompt to each UAV to perform semantic matching calculations.

[0150] The drone utilizes visual language models, such as, but not limited to, the BLIP-2 model, to process regional images separately. and task text prompts Perform feature encoding to obtain image feature vectors. With text feature vectors The two are compared using cosine similarity in a unified semantic embedding space:

[0151]

[0152] income Indicates the area The degree of matching with the task semantics. To facilitate consistent comparison, the results are linearly normalized:

[0153]

[0154] Normalized semantic fit The higher the value, the better the semantic fit of the region with the task. For example, when the task instruction text is "find the burning gas cylinder", the model can infer that gas cylinders usually appear in kitchen scenes. Therefore, the cosine similarity of the kitchen region is the highest, that is, the semantic fit is the greatest.

[0155] 2.3 Regional Exploration Value Calculation

[0156] To prioritize sub-regions, this embodiment calculates region exploration value based on three indicators: semantic fit, obstacle density, and path reachability.

[0157] Obstacle density is defined as follows:

[0158]

[0159] in, Indicates the area The number of barrier voxels within, Indicates the area The total number of elements.

[0160] Path reachability Indicates the area Accessibility is used to measure how well an area is connected to other key locations. First, the area is calculated on the topology map. Average shortest path cost to all critical nodes Then, its reachability is defined as: ,

[0161] in To prevent the use of tiny constants that divide by zero. To standardize dimensions and facilitate comparison, further... Normalization is performed:

[0162]

[0163] Normalized path reachability The larger the value, the easier the area is to be accessed by drones.

[0164] All indicators have been normalized for easier comparison. The final exploration value score for a region is defined as follows:

[0165]

[0166] in, , and Representing the normalized semantic fit, obstacle density, and path reachability, respectively, the weight parameters are: satisfy .

[0167] Comprehensive Exploration Value of Each Region The regions are sorted to form an exploration value map. This value map reflects the exploration potential of each sub-region, providing a quantitative basis for task allocation. By introducing semantic fit, this invention achieves a shift from a geometry-driven to a semantic-driven exploration mechanism, enabling UAVs to autonomously identify and prioritize access to high-value regions.

[0168] Step 3: Constrained Optimization Modeling: With the objective of minimizing the total flight time, the multi-UAV task allocation is modeled as a path optimization problem with capacity constraints, where UAV capability scores serve as capacity constraints and area exploration value serves as demand constraints; the specific steps are as follows:

[0169] 3.1 Defining the Model

[0170] In this implementation, the multi-UAV task allocation problem is modeled as a Capacitated Vehicle Routing Problem (CVRP). In this model, UAVs are treated as vehicles with different task-carrying capacities, and the exploration area is treated as customer points with different task requirements. By establishing a mapping between capabilities and requirements, balanced task allocation and global path optimization are achieved.

[0171] Let the set of drones be Each drone The overall ability score is recorded as the capacity limit. Let the set of exploration regions be... Each region Corresponding exploration value This indicates the task demand for that region. The system introduces nodes. This serves as the unified start and end point for all drone paths. In the model, These are considered virtual nodes, used to define the start and return positions of a path; in real-world scenarios, This corresponds to a centralized take-off and landing area or supply center for UAVs, used to achieve unified scheduling and closed-loop execution of missions.

[0172] To describe path relationships, define binary variables. When drones From the region Fly to the area hour, Otherwise, it is 0. The flight time between regions is defined as:

[0173]

[0174] in, The distance between regions. The maximum speed of the drone, This is an environmental correction factor used to account for the effects of factors such as wind speed and obstacles.

[0175] 3.2 Determine the optimization objective

[0176] The goal of the model is to minimize the total flight time of all drones while satisfying various constraints:

[0177]

[0178] This objective ensures that the system completes all exploration tasks while minimizing the total operation time.

[0179] 3.3 Determine the constraints

[0180] This optimization problem needs to satisfy the following 5 constraints:

[0181] 1) Capacity Constraint: The total mission requirements of each drone must not exceed its capacity limit. To allow for a safety margin, a proportionality coefficient is introduced. :

[0182]

[0183] This constraint ensures that each drone will not experience mission interruption or insufficient energy due to excessive load during operation.

[0184] 2) Unique access constraint: Each sub-region must be accessed by one and only one drone once.

[0185]

[0186] This constraint prevents repeated exploration or omission of areas, ensuring that the task allocation is reasonable.

[0187] 3) Flow balance constraint: This constraint ensures path continuity. After a UAV enters a certain area, it must leave that area once more; that is, the number of entries must equal the number of exits.

[0188]

[0189] For each drone It is in any task region The number of entries must equal the number of exits. This condition ensures that all access nodes are connected to form a complete task path, avoiding breakpoints or dangling nodes.

[0190] 4) Start and end point constraints: Each drone starts from the same start and end point once and eventually returns once.

[0191]

[0192] This constraint specifies the start and end points of each mission path, ensuring a complete closed loop for the operation of each UAV.

[0193] 5) Sub-loop elimination constraint: To prevent independent small loops from appearing even after the aforementioned constraints are met, this implementation uses a sub-loop elimination condition. For each UAV... and each task node Define access order variables Its constraints are:

[0194]

[0195] This constraint prevents the generation of closed sub-paths that are independent of a unified start and end point, thus ensuring the overall connectivity of each task route.

[0196] Step 4: Optimize the solution and dynamically reallocate

[0197] This invention, based on a constrained optimization model, transforms the UAV task allocation problem into a capacity-constrained vehicle routing problem (CVRP) and solves this model. Through optimized solutions, all task areas can be efficiently covered while satisfying UAV capability constraints. Furthermore, when environmental or system states change, a dynamic reallocation mechanism enables real-time updates of the solution, thereby ensuring the stability and balance of overall task execution.

[0198] 4.1 Solving the CVRP Problem

[0199] In this invention, it is assumed that there is a total The drones were deployed, and the mission points were assembled at the following locations: The exploration value of each task point is The capability rating of each drone is: Path cost Indicates the drone departs from the mission point Arrive at the mission point The overall flight cost is calculated as follows. The optimization objective is to minimize the overall cost function, which is expressed as follows, under the condition that each mission point is visited only once and the load of each UAV does not exceed its capacity constraints:

[0200]

[0201] in, For drones The set of access paths must satisfy the following constraints. .

[0202] CVRP is a typical combinatorial optimization problem, whose computational complexity increases exponentially with the number of task points, making direct exhaustive search difficult to complete within the task time limit. To balance solution efficiency and solution quality, this invention employs a heuristic algorithm to obtain an approximately optimal task allocation scheme by constructing an initial solution and gradually optimizing it.

[0203] The solution process consists of three stages. The first stage is task clustering. The system clusters tasks based on their exploration value. With drone capability rating The ratio of the value density of task points is used to categorize task points according to their value density. The task points are sorted and divided into clusters based on geographical proximity and load balancing principles, with each cluster corresponding to the task range of one drone. The second stage is path construction. For each task cluster, an initial path is generated using the drone's starting position as the starting point and the Savings Algorithm. The algorithm first assumes that each task point is independently round-trip, and then calculates the cost savings after merging the task points. The paths are merged in descending order of savings until the capacity constraints are met. The third stage is path optimization. The system performs local adjustments to the initial path, including node swapping, node insertion, and path reversal. Node swapping is used to balance the load across different paths, node insertion is used to optimize the access order within the same path, and path reversal is used to reduce redundant tracks. The cost change is calculated for each operation. ,like If so, then accept the new plan. Through multiple iterations, the initial task allocation scheme is obtained when the total cost stabilizes or the maximum number of iterations is reached. .

[0204] 4.2 Triggering conditions for task reassignment

[0205] During mission execution, the capabilities of the UAV, the semantic features of the mission area, and the execution results may change over time, causing the original mission plan to lose its optimality or become partially infeasible. To ensure the real-time performance and stability of the system, this invention sets triggering conditions for mission reassignment, judging from three aspects: changes in UAV capabilities, changes in regional semantics, and mission execution deviations.

[0206] First, in response to changes in drone capabilities, the system monitors the capability score of each drone in real time. When the current ability value is significantly lower than the initial ability value At this point, the drone's payload capacity or endurance is considered to have decreased. The degree of change in capability can be expressed as:

[0207]

[0208] when When the system determines that the drone's capability has declined beyond a preset threshold, it triggers a task reassignment. The threshold for capability changes can be set based on factors such as the characteristics of the drone.

[0209] Secondly, in response to changes in the semantic value of a region, the system dynamically updates the exploration value of task points based on the environmental perception results. When the real-time exploration value of a task point differs significantly from its initial value, it indicates a significant change in the importance, accessibility, or risk factors of that area. The magnitude of semantic change can be defined as:

[0210]

[0211] when When this occurs, the system determines that the semantic features of the task region have changed significantly, thus triggering task reallocation. The semantic change threshold is determined by the task scenario parameters.

[0212] Finally, to address mission execution deviations, the system compares the actual flight costs of the drone. Cost of Model Prediction If the two deviate significantly, it indicates that the path planning result does not match the actual execution. Execution deviation is defined as:

[0213]

[0214] when When the system determines that the current task allocation scheme is no longer applicable, it enters the task reallocation phase. The deviation threshold can be set according to the accuracy requirements of the task.

[0215] When any one of the above three conditions meets the threshold requirement, the system automatically starts the task reassignment module to re-solve and adjust the affected task points and UAV paths.

[0216] 4.3 Dynamic Task Reassignment

[0217] Once the triggering condition is met, the system enters the dynamic task reassignment phase. To improve real-time performance and computational efficiency, this invention employs an incremental optimization strategy, adjusting only the affected task nodes and related UAV paths without resolving the entire CVRP model. Specific steps include: first, freezing the currently executing path segments and including only the incomplete parts in the replanning set; second, calculating the incremental cost of reinserting the affected task points in different UAV paths. The system selects the insertion combination that minimizes the overall cost and performs the update accordingly; then, it sequentially performs node migration and path correction operations to generate a new task allocation scheme. Finally, when the difference in total cost between the old and new solutions satisfies When the solution reaches a stable state, no further adjustments are made. This dynamic redistribution method achieves lightweight solution while maintaining task continuity. It can quickly redistribute tasks when UAV capabilities change, environmental semantics are updated, or task execution deviations occur, thus ensuring that the system can continuously and efficiently complete multi-UAV collaborative tasks in dynamic environments.

[0218] Example 2

[0219] 1. Combination Figure 1 This embodiment provides a dynamic task allocation system for UAV swarms based on heterogeneous capability and semantic evaluation, specifically including: a UAV capability modeling module, an environmental perception module, an environmental semantic value evaluation module, a constraint optimization modeling module, and a dynamic reallocation and task execution module. The above modules of the system are used to execute the method described in any one of claims 1-9.

[0220] The UAV capability modeling module is used to establish a comprehensive capability evaluation model for individual UAVs from three dimensions: perception capability, maneuverability, and endurance capability, and generate a comprehensive capability score for each UAV.

[0221] The environmental perception module acquires RGB images, depth maps, and IMU data collected by sensors on the UAV. The RGB images provide texture information about the environment to support subsequent target detection and semantic segmentation; the depth maps reflect the spatial structure of the scene, used to identify obstacle locations and passable areas; and the IMU data is used to calculate the UAV's attitude and motion state. By fusing information from these multi-source sensors, the environmental perception module forms fundamental perception data describing the environment's appearance and spatial structure, providing reliable environmental input for subsequent semantic value assessment and task allocation.

[0222] The environmental semantic value assessment module includes an environmental analysis unit and a visual-language alignment unit. The environmental analysis unit receives data from the environmental perception module and performs target detection, semantic segmentation, and semantic assessment operations. The visual-language alignment unit calculates the semantic relevance between environmental regions and task objectives. The environmental semantic value assessment module combines semantic fit, obstacle density, and path reachability to calculate the comprehensive exploration value of each region. Sorting is performed to create a regional exploration value map;

[0223] The constrained optimization modeling module is used to model the allocation of multiple UAV missions as a path optimization problem with capacity constraints. The UAV capability score serves as the capacity constraint, the area exploration value serves as the demand constraint, and the optimization objective is to minimize the total flight time.

[0224] The dynamic redistribution and task execution module includes an initial allocation unit and an optimization adjustment unit. The initial allocation unit is used to construct task clusters based on UAV capability scores and regional exploration value, and generate an initial task allocation scheme. The optimization adjustment unit is used to locally optimize the initial scheme according to path costs and constraints. When changes are detected in UAV capability parameters or task area semantic information, dynamic task redistribution is triggered to achieve real-time optimization and global balance of task planning. The optimized task allocation scheme is distributed to each UAV terminal for execution by the task execution unit, and the task execution status is monitored and feedback is provided.

[0225] Compared to existing multi-UAV collaborative exploration systems, this invention simultaneously introduces capability modeling and semantic value assessment modules during the mission planning process. The system can acquire and update UAV status and environmental semantic features in real time, and complete problem solving and task allocation through dynamic reallocation and task execution modules. After generating the initial task allocation plan, the system can continuously monitor changes in UAV status and environmental semantics, and automatically trigger the task reallocation process when necessary, achieving dynamic task adjustment. This mechanism enables UAV swarms to work efficiently in complex scenarios, effectively utilize system resources, and maintain high information acquisition efficiency, making it suitable for applications such as disaster relief, resource surveying, and target search.

[0226] Experimental data verification:

[0227] To verify the effectiveness and technical advantages of the system described in this invention, multiple sets of performance verification experiments based on simulation environments were conducted. The experiments consisted of two parts: comparative experiments and ablation experiments. The comparative experiments were used to verify the performance improvement of the method described in this invention compared to existing multi-UAV cooperative exploration algorithms, while the ablation experiments were used to analyze the independent roles of key functional modules in the system and their impact on overall performance.

[0228] The experiments were conducted on ROS, Gazebo, and PX4 Autopilot simulation platforms. The experimental environments included two types: a complex maze scenario with high spatial irregularity and a typical indoor office scenario with rich semantic structure and obstacle distribution. To verify the system's adaptability and stability under different semantic structures and spatial layouts, independent tests were performed in both scenarios.

[0229] During the experiment, the UAV performed a collaborative exploration task using the task allocation algorithm of the present invention, and generated a corresponding 3D environmental map during the exploration process to reflect the spatial coverage and path planning effect of the group exploration. This 3D map, generated by the environmental modeling and path recording module of the present invention, can intuitively display the exploration coverage and map construction accuracy, as shown in Figure 5.

[0230] The participating drone platforms exhibited significant differences in perception capabilities, maneuverability, and endurance to verify the system's load distribution effectiveness under heterogeneous platform conditions. The mission terminated when the proportion of unexplored voxels in the environment fell below 3%. During the experiment, all drones completed their exploration tasks without human intervention.

[0231] The experimental evaluation metrics include three items: total task time, total path length, and task time variance. Total task time measures the overall time required for the entire cluster to complete the exploration task; total path length reflects the effectiveness of path optimization during the planning and task allocation phases; and task time variance measures the degree of task load balance among different UAVs, with a smaller variance indicating a more even load distribution.

[0232] (2) Comparative experiment

[0233] In the comparative experiment, the performance of the method of this invention was compared with that of the existing representative algorithm RACER (Rapid Collaborative Exploration) under different numbers of UAVs. The experimental results are shown in Table 1.

[0234] Table 1 Comparative Experiment Results with Different Numbers of Drones

[0235] Scene Number of drones method Total time (s) Path length (m) Time variance (s²) maze 4 RACER 410.3 1279.2 111.5 maze 4 This invention 360.8 1120.5 70.7 maze 6 RACER 368.1 1278.9 128.3 maze 6 This invention 312.5 1188.1 69.3 maze 8 RACER 340.1 1430.7 102.0 maze 8 This invention 298.4 1258.6 65.1 office 4 RACER 70.6 237.6 101.6 office 4 This invention 62.3 221.5 60.3 office 6 RACER 73.1 242.9 97.0 office 6 This invention 54.2 218.1 30.4 office 8 RACER 59.4 278.7 65.5 office 8 This invention 48.9 230.6 24.6

[0236] As shown in Table 1, the method of this invention exhibits significant performance advantages under different drone scales and environmental conditions. Compared with the RACER method, this invention can reduce the total task time by approximately 15% on average, decrease the path length by approximately 7%, and reduce the task time variance by approximately 46% in complex maze scenarios. The performance improvement is even more pronounced in semantically significant office environments, particularly in load balancing. As the number of drones increases from 4 to 8, the overall performance of the method of this invention maintains a stable improvement, while the performance of the comparative algorithm fluctuates significantly, indicating that this invention has good scalability and robustness when scaling up to multiple drones.

[0237] (3) Ablation test

[0238] To further verify the independent function of each key module of the system, an ablation experiment was conducted. The experiment fixed the number of UAVs at four, removing the capability modeling module and the environmental semantic value assessment module, and setting up a traditional scheme based solely on geometric partitioning as a control. The experimental results are shown in Table 2.

[0239] Table 2 Ablation Experiment Results (Number of UAVs = 4)

[0240] Scene method Total time (s) Path length (m) Time variance (s²) maze Full system 360.8 1120.5 70.7 maze No-Cap (No-Cap) 385.2 1203.6 97.4 maze No-Semantic Modules (No-Sem) 373.9 1176.3 85.1 maze Geometric partitioning (Equal) 412.7 1292.1 113.6 office Full system 62.3 221.5 60.3 office No-Cap (No-Cap) 66.8 233.4 78.6 office No-Semantic Modules (No-Sem) 70.5 239.1 102.4 office Geometric partitioning (Equal) 71.4 242.7 110.2

[0241] As shown in Table 2, removing the capability modeling module significantly reduced the overall system performance, with both total task time and time variance increasing substantially. This indicates that ignoring performance differences between UAVs will lead to uneven resource utilization and task delays. Removing the environmental semantic value assessment module resulted in a more pronounced performance decline in scenarios with significant semantic features, particularly in an office environment where the task time variance increased from 60.3 to 102.4. This verifies the importance of semantic value assessment for task allocation accuracy in scenarios with high semantic complexity. When both modules were removed simultaneously, relying solely on geometric partitioning for task allocation resulted in the worst system performance, with significantly reduced task efficiency and path planning effectiveness. This further demonstrates that the two core modules in this invention play crucial roles in achieving dynamic task balancing and global path optimization.

[0242] The experimental results above demonstrate that this invention, by combining UAV heterogeneous capability modeling with environmental semantic value assessment, achieves globally optimal allocation and dynamic real-time adjustment of exploration tasks, significantly improving the efficiency of group exploration and resource utilization. Compared with existing geometry-driven algorithms, this invention maintains stable performance in complex, dynamic, and semantically diverse environments, effectively avoiding single-unit overload and redundant flight, demonstrating good system robustness and engineering application value.

[0243] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic task allocation in UAV swarms based on heterogeneous capabilities and semantic evaluation, comprising the following steps: Step 1: Modeling and quantification of the heterogeneous capabilities of UAVs; The heterogeneous capabilities of each drone in the drone swarm are modeled and quantified from three aspects: perception capability, maneuverability and endurance, and a comprehensive capability score is given for each drone. Step 2: Semantic value assessment and region ranking; Use UAV sensors to collect environmental perception data and divide the environment into regions; Analyze the semantic relevance of each region to the task objective based on the visual language alignment mechanism, and calculate the exploration value of each region by combining obstacle density and path accessibility to form a region value ranking. Step 3: Constraint optimization modeling; With the goal of minimizing total flight time, the task allocation problem of multiple UAVs is modeled as a path optimization problem with capacity constraints (CVRP), where UAV capability scores serve as capacity constraints and area exploration value serves as demand constraints. Step 4: Optimize the solution and dynamically reallocate tasks; construct task clusters based on UAV capability scores and regional exploration value, use a cost-saving algorithm to generate initial paths and obtain preliminary task allocation schemes through local optimization, and trigger dynamic task reallocation when UAV capabilities or environmental semantics change.

2. The method according to claim 1, characterized in that, The evaluation of the drone in step 1 Overall performance, its perception capabilities Mobility and battery life Perform a weighted geometric average; let the weights of each capability be respectively... , and And satisfy ; drones The formula for calculating the comprehensive ability score is as follows: 。 3. The method according to claim 2, characterized in that, The perception ability in step 1 Horizontal field of view based on sensor Vertical field of view With maximum sensing distance Quantify at least one of the following: mobility Based on maximum linear velocity Maximum linear acceleration Maximum angular velocity With maximum angular acceleration At least one of the four dynamic parameters is quantified; range Quantification is based on a combination of energy reserves and flight energy consumption trends.

4. The method according to claim 3, characterized in that, The steps for constructing the perception ability scoring model are as follows: Let the vertex of the square pyramid be the sensor position, the base be a rectangle, and the distance from the vertex to the base be D; therefore, the length and width of the base are: ， Therefore, the field volume can be obtained: ， Based on cluster reference volume Based on this, the scores are normalized as follows: ， in, Indicates perceptual ability rating The horizontal and vertical field of view of the sensor are in radians. Maximum sensing distance; It can be set by the cluster baseline configuration or by statistical analysis of typical scenarios; The overall maneuverability scoring model for the UAV is constructed as follows: Maneuverability is divided into two dimensions: translational capability and steering capability, which respectively reflect the linear motion response and attitude adjustment capability of the UAV in space. Translational capability takes into account the time required for a drone to accelerate from a standstill to its maximum speed. Compare it with the reference value Normalize the ratio to obtain the translational capability. ; Steering capability corresponds to attitude adjustment speed, and is approximated by the time required to reach maximum angular velocity. The formula for calculating steering response time is: Compare it with the reference value Normalize the ratio to obtain the steering capability. ; The final maneuverability score is expressed as the geometric mean of translational and steering capabilities: ， in, This represents the overall maneuverability score of the drone, a reference value. The data can be taken from the drone swarm baseline platform or historical samples; this scoring method is based on response time derivation, avoiding subjective weight setting; The steps for constructing the battery life rating model are as follows: Introducing the derivative of acceleration, jerk, the formula is as follows: (unit ), and defined within the task time interval. Cumulative fluctuations within: ， Considering the current battery state, let For remaining battery power, If the battery is at full capacity, then the percentage of remaining charge is: To eliminate the dimensionlessness and suppress the range uncertainty caused by violent maneuvers, a reference constant is introduced. The battery life rating is defined as follows: ， in, Calibration can be performed using standard operating conditions or historical mission data.

5. The method according to claim 1, characterized in that, The semantic segmentation in step 2 involves: acquiring environmental images using an RGB-D sensor and performing semantic segmentation; employing a text-guided object detection model, which detects task-related objects in the image by providing natural language prompts, thus obtaining a target set. ， Each object Contains category tags and location information; Subsequently, using an image segmentation model, contour segmentation was performed on the detection results to obtain object mask information. Based on spatial proximity and semantic feature correlation, the environment was clustered to obtain several semantically consistent sub-regions. ， Each sub-region Corresponding to a functional or semantic scenario; The visual language alignment mechanism employs a visual language model. The cluster master node receives the task instruction and distributes it to each UAV to perform semantic matching calculations. The UAVs utilize the visual language model to perform semantic matching calculations on the regional images. and task text prompts Perform feature encoding to obtain image feature vectors. With text feature vectors The cosine similarity between the two is calculated in a unified semantic embedding space. ， income Indicates the area The degree of matching with the task semantics is used to linearly normalize the results: ， Normalized semantic fit The larger the value, the better the region matches the semantics of the task.

6. The method according to claim 1, characterized in that, In step 2, the comprehensive exploration value of each region is calculated based on three indicators: semantic fit, obstacle density, and path accessibility. Obstacle density is defined as follows: ， in, Indicates the area The number of barrier voxels within, Indicates the area The total number of elements; Path reachability Indicates the area Accessibility is used to measure how well the area is accessible to other key locations; first, the area is calculated on the topological map. Average shortest path cost to all critical nodes Then, its reachability is defined as: , in To prevent the small constant from being divided by zero, further... Normalization is performed: ， Normalized path reachability The larger the value, the easier the area is to be accessed by drones; The final exploration value score of a region is defined as follows: ， in, , and Representing the normalized semantic fit, obstacle density, and path reachability, respectively, the weight parameters are: satisfy ; Comprehensive Exploration Value of Each Region The regions are sorted to form a value map for regional exploration.

7. The method according to claim 1, characterized in that, Step 3, which models the multi-UAV task allocation problem as a capacity-constrained path optimization problem (CVRP), specifically includes: Let the set of drones be Each drone The overall ability score is recorded as the capacity limit. ; Let the set of exploration regions be Each region Corresponding exploration value This indicates the task demand in that area; The system introduces node r0 as the unified start and end point for all UAV paths. A centralized take-off and landing area or supply center for corresponding UAVs, used to achieve unified scheduling and closed-loop execution of missions; Define a binary variable When drones From the region Fly to the area hour, Otherwise, it is 0; the flight time between regions is defined as: ， in, The distance between regions. The maximum speed of the drone, This is an environmental correction factor used to account for the effects of factors such as wind speed and obstacles; The objective function of the CVRP model is: minimize the total flight time. ， And satisfy the following constraints Capacity constraints: the total number of tasks undertaken by each drone must not exceed its capacity limit; and a proportionality coefficient is introduced. : ， Each sub-region must be visited by one drone only once: ， Flow balance constraint: The number of entries equals the number of exits. ， Start and end point constraints: Each drone departs from a unified start and end point once and eventually returns once. ， Sub-loop constraint elimination: for each UAV and each task node Define access order variables Its constraints are: 。 8. The method according to claim 1, characterized in that, The optimized solution to the CVRP problem in step 4 is as follows: Assume there is a total The drones were deployed, and the mission points were assembled at the following locations: The exploration value of each task point is The capability rating of each drone is: Path cost Indicates the drone departs from the mission point Arrive at the mission point The overall flight cost is calculated as follows; the optimization objective is to minimize the overall cost function, expressed as follows, under the condition that each mission point is visited only once and the load of each UAV does not exceed its capacity constraints: ， in, For drones The set of access paths must satisfy the following constraints. ; First, tasks are clustered, and the system determines the exploration value of each task point. With drone capability rating The ratio of the value density of task points is used to categorize task points according to their value density. The task points are sorted and divided into several clusters based on geographical proximity and load balancing principles, with each cluster corresponding to the task range of one drone. Then, path construction is performed: for each task cluster, starting from the drone's initial position, an initial path is generated using a cost-saving algorithm. The algorithm first assumes that each task point is independently round-tripped, and then calculates the cost savings after merging the task points. The paths are merged in descending order of savings until the capacity constraints are met. Finally, path optimization is performed. The system executes local adjustments to the initial path, including node swapping, node insertion, and path reversal. Node swapping is used to balance the load across different paths, node insertion is used to optimize the access order within the same path, and path reversal is used to reduce redundant tracks. The cost change is calculated for each operation. ,like If so, then accept the new plan. Through multiple iterations, the initial task allocation scheme is obtained when the total cost stabilizes or the maximum number of iterations is reached. .

9. The method according to claim 1, characterized in that, The task reassignment in step 4 is judged from three aspects: changes in UAV capabilities, changes in regional semantics, and task execution deviations. When any one of the three conditions meets the threshold requirement, the system automatically starts the task reassignment module to resolve and adjust the affected task points and UAV paths. In response to changes in drone capabilities, the system monitors the capability score of each drone in real time. When the current ability value is significantly lower than the initial ability value When this occurs, it is considered that the drone's payload capacity or endurance has decreased; the degree of change in capability is expressed as follows: ， when When the system determines that the drone's capability has declined beyond a preset threshold, it triggers a task reassignment; among these, The threshold for capability changes can be set based on factors such as the characteristics of the drone; In response to changes in the semantic value of a region, the system dynamically updates the exploration value of task points based on environmental perception results. When the real-time exploration value of a task point differs significantly from its initial value, it indicates a significant change in the importance, accessibility, or risk factors of that area; the magnitude of semantic change can be defined as: ， when When the system detects a significant change in the semantic features of the task region, it triggers task reassignment; among which... The semantic change threshold is determined by the task scenario parameters; To address mission execution deviations, the system compares the actual flight costs of the drone. Cost of Model Prediction If the two deviate significantly, it indicates that the path planning result does not match the actual execution. The execution deviation is defined as: ， when When the system deems the current task allocation scheme no longer applicable, it enters the task reallocation phase; among which, The deviation threshold can be set according to the accuracy requirements of the task; Once the triggering condition is met, the system enters the dynamic task reallocation phase. An incremental optimization strategy is employed to adjust the affected task nodes and related UAV paths. Specific steps include: first, freezing the currently executing path segments and including only the incomplete parts in the replanning set; second, calculating the incremental cost of reinserting the affected task points into different UAV paths. The system selects the insertion combination that minimizes the overall cost and performs the update accordingly; then, it sequentially performs node migration and path correction operations to generate a new task allocation scheme. Finally, when the difference in total cost between the old and new solutions satisfies At that point, the plan was considered to have reached a stable state, and no further adjustments were made.

10. A dynamic task allocation system for UAV swarms based on heterogeneous capabilities and semantic evaluation, characterized in that, The system includes a UAV capability modeling module, an environmental perception module, an environmental semantic value assessment module, a constraint optimization modeling module, and a dynamic reallocation and task execution module. The above modules of the system are used to perform the method described in any one of claims 1-9. The UAV capability modeling module is used to establish a comprehensive capability evaluation model for individual UAVs from three dimensions: perception capability, maneuverability, and endurance capability, and generate a comprehensive capability score for each UAV. The environment perception module is used to acquire RGB images, depth maps, and IMU data collected by sensors on the UAV. The RGB images provide texture information about the environment to support subsequent target detection and semantic segmentation. The depth map reflects the spatial structure of the scene to identify obstacle locations and passable areas. The IMU data is used to calculate the UAV's attitude and motion state. By fusing the information from these multi-source sensors, the environment perception module forms basic perception data describing the appearance and spatial structure of the environment, providing reliable environmental input for subsequent semantic value assessment and task allocation. The environmental semantic value assessment module includes an environmental analysis unit and a visual-language alignment unit. The environmental analysis unit receives data from the environmental perception module and performs target detection, semantic segmentation, and semantic assessment operations. The visual-language alignment unit calculates the semantic relevance between environmental regions and task objectives. The environmental semantic value assessment module combines semantic fit, obstacle density, and path reachability to calculate the comprehensive exploration value of each region. Sorting is performed to create a regional exploration value map; The constrained optimization modeling module is used to model the multi-UAV task allocation problem as a path optimization problem with capacity constraints. The UAV capability score is used as the capacity constraint, the area exploration value is used as the demand constraint, and the optimization objective is to minimize the total flight time. The dynamic redistribution and task execution module includes an initial allocation unit and an optimization adjustment unit. The initial allocation unit is used to construct task clusters based on UAV capability scores and regional exploration value, and generate an initial task allocation scheme. The optimization adjustment unit is used to locally optimize the initial scheme according to path costs and constraints. When changes are detected in UAV capability parameters or task area semantic information, dynamic task redistribution is triggered to achieve real-time optimization and global balance of task planning. The optimized task allocation scheme is distributed to each UAV terminal for execution by the task execution unit, and the task execution status is monitored and feedback is provided.

Images