Unmanned system swarm intelligence large model task planning method based on virtual-real migration and feedback evolution

Abstract

This invention provides a task planning method for unmanned system swarm intelligence based on virtual-real migration and feedback evolution. It constructs a local environmental perception and sharing distribution mechanism for unmanned system swarm task scenarios, extracting local perception features; builds an airborne task planning model based on unified state information and globally shared features, forming a lightweight task planning model suitable for airborne platform deployment; constructs a distributed collaborative decision-making and conflict resolution mechanism, a virtual-real feature offset analysis mechanism, and a closed-loop result hierarchical comprehensive evaluation mechanism to obtain the evaluation results required for model optimization; and constructs a collaborative evolution update mechanism for the large model based on the evaluation results to adaptively update the airborne task planning model, realizing the continuous evolution of unmanned system swarm intelligence task planning capabilities in dynamic environments. This invention can improve the collaboration, robustness, and environmental adaptability of unmanned system swarm task planning under conditions of limited communication, limited computing power, and dynamic environmental changes.

Description

Technical Field

[0001] This invention relates to the field of unmanned system swarm intelligence collaborative control and large model task planning, and particularly to a large model-driven unmanned system swarm intelligence distributed collaboration and virtual-real evolution task planning method; Background Technology

[0002] Unmanned system swarms are widely used in military reconnaissance, coordinated strikes, emergency rescue and complex industrial operations. Their mission planning capabilities directly affect mission execution efficiency and system coordination effects. In complex and dynamic environments, unmanned system swarms usually face problems such as limited local perception, insufficient communication bandwidth, limited onboard computing power and strong environmental uncertainty. Traditional centralized mission planning methods are difficult to meet the requirements of real-time performance, autonomy and robustness for unmanned system swarm collaborative tasks.

[0003] With the development of large model technology, using large models for unmanned system mission planning has become a new research direction. However, most existing large models rely on high-computing platforms and are difficult to directly adapt to airborne restricted environments. At the same time, the mission planning capabilities formed by large models in simulation environments are prone to performance degradation when transferred to real environments due to differences between the characteristics of real environments and virtual mapping environments. In addition, existing methods generally lack closed-loop evaluation and continuous evolution update mechanisms that incorporate real-world feedback, making it difficult to achieve long-term adaptive improvement of the swarm intelligence mission planning capabilities of unmanned systems.

[0004] Therefore, it is necessary to propose a large model-driven swarm intelligence distributed collaboration and virtual-real evolution task planning method for unmanned systems to solve problems such as limited local perception, difficulty in multi-machine collaboration, limited airborne deployment, virtual-real migration mismatch and insufficient model continuous evolution capability. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a task planning method for a large-scale swarm intelligence model of unmanned systems that incorporates virtual-real migration and feedback evolution.

[0006] A task planning method for a large-scale swarm intelligence model of unmanned systems involving virtual-real migration and feedback evolution includes the following steps:

[0007] Step 1: Construct a local environment perception and sharing distribution mechanism for unmanned system cluster task scenarios, perform unified state representation of the position information, velocity information, resource status information and environmental constraint information of each agent, and extract local perception features to provide basic input for multi-agent collaborative planning;

[0008] Step 2: Tensor quantization, compression coding and spatiotemporal registration are performed on the local perception features of each agent. Unified global shared features are formed through time synchronization and spatial coordinate alignment to reduce communication overhead and improve the consistency of multi-source information fusion.

[0009] Step 3: Based on the unified state information and globally shared features, construct a large airborne mission planning model, and combine model compression and model quantization to form a lightweight mission planning model suitable for airborne platform deployment, so as to output the mission planning results corresponding to each intelligent agent.

[0010] Step 4: Based on the task planning results, construct a distributed collaborative decision-making and conflict resolution mechanism, apply safe distance constraints, resource allocation constraints and time sequence coordination constraints to the planning actions of each agent, and filter and correct the action results to obtain collaborative decision results that meet the collaborative safety requirements;

[0011] Step 5: Based on the collaborative decision-making results, construct a virtual-real feature offset analysis mechanism to characterize the differences between real environment features and virtual mapped environment features, and obtain virtual-real migration difference information;

[0012] Step Six: Based on the virtual-real migration difference information and task execution results, construct a closed-loop result hierarchical comprehensive evaluation mechanism. Use the analytic hierarchy process (AHP)-entropy weight method to perform subjective and objective weighting on the evaluation indicators, and comprehensively score the inference and decision-making results in the real environment and the virtual mapping environment to obtain the evaluation results required for model optimization.

[0013] Step 7: Based on the evaluation results, construct a large-scale model collaborative evolution update mechanism. Through local case fine-tuning, sparse gradient sharing, and confidence-driven federated aggregation, the airborne mission planning model is adaptively updated to realize the continuous evolution of the unmanned system's swarm intelligence mission planning capability in a dynamic environment.

[0014] This invention forms a complete technical closed loop through the above seven steps: local perception fusion, airborne lightweight planning, distributed collaborative decision-making, virtual-real offset calibration, and closed-loop evaluation and evolution. This effectively solves core problems in existing technologies, such as difficulties in airborne deployment of large models, weak global collaborative capabilities under communication constraints, a sharp drop in performance when migrating from simulation to reality, and a lack of continuous adaptive capabilities in dynamic environments. Compared with traditional centralized task planning methods and large model planning methods using single simulation training, this invention can significantly improve the autonomous collaboration level, task execution robustness, and environmental adaptability of unmanned system swarms in complex task scenarios with insufficient communication bandwidth, limited airborne computing power, and highly dynamic environments. Simultaneously, through a continuous evolution mechanism, it achieves long-term iterative improvement of task planning capabilities, making it widely adaptable to the task requirements of unmanned system swarms in various fields such as military reconnaissance, collaborative strikes, emergency rescue, and complex industrial operations. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the task planning method for the unmanned system swarm intelligence large model based on virtual-real migration and feedback evolution of the present invention;

[0016] Figure 2 This is a schematic diagram comparing the errors before and after spatiotemporal registration.

[0017] Figure 3 A schematic diagram illustrating the trajectory and time sequence recovery verification before and after conflict resolution;

[0018] Figure 4 This is a schematic diagram of the convergence of the virtual-real closed-loop evolution proposed in this invention. Detailed Implementation

[0019] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings:

[0020] like Figure 1 As shown, a task planning method for a large-scale swarm intelligence model of unmanned systems based on virtual-real migration and feedback evolution includes the following steps:

[0021] Step 1: The agent's state information includes position vector, velocity vector, and resource vector. The environmental constraint information includes static obstacles, dynamic threats, and communication restrictions. A unified state space representation is constructed using vectorization.

[0022] Suppose that the unmanned system cluster includes The first intelligent agent, the... An intelligent agent at time The unified state vector is represented as:

[0023]

[0024] in, Indicates the first An intelligent agent at time The unified state vector, Represents a position vector. Represents the velocity vector. Represents a resource vector. Indicates a static obstacle. Indicates a dynamic threat. Indicates communication constraints. This represents the total number of intelligent agents. Indicates the agent index. Indicates the current moment;

[0025] Position vector Represented as:

[0026]

[0027] in, Indicates the first An intelligent agent at time The position vector, , , Indicates the first An intelligent agent at time Three-dimensional spatial coordinates;

[0028] velocity vector Represented as:

[0029]

[0030] in, Indicates the first An intelligent agent at time The velocity vector, , , They represent the first An intelligent agent at time The velocity components along the three coordinate axes;

[0031] Resource Vector Represented as:

[0032]

[0033] in, Indicates the first An intelligent agent at time resource vector, Indicates the first An intelligent agent at time The remaining energy, Indicates the first An intelligent agent at time Remaining load capacity Indicates the first An intelligent agent at time The remaining task resources;

[0034] Static obstacle information vector Represented as:

[0035]

[0036] in, Indicates the first The static obstacle information vector perceived by an agent. Indicates the number of static obstacles. Indicates the first The first agent and the second The distance between static obstacles Indicates the static obstacle index;

[0037] Dynamic Threat Information Vector Represented as:

[0038]

[0039] in, Indicates the first An intelligent agent at time The perceived dynamic threat information vector, Indicates the number of dynamic threat sources. Indicates the first An intelligent agent at time With the The distance between dynamic threat sources Indicates the first The first dynamic threat source for the first An intelligent agent at time The intensity of the threat, Indicates a dynamic threat source index;

[0040] Communication constraint vector Represented as:

[0041]

[0042] in, Indicates the first An intelligent agent at time The communication constraint vector, Indicates the first An intelligent agent at time Available communication bandwidth Indicates the first An intelligent agent at time Communication latency, Indicates the first An intelligent agent at time The stability coefficient of the communication link.

[0043] In step two: Zhang quantitative representation includes multi-dimensional feature encoding of multi-agent perception information, compression encoding reduces communication overhead through dimensionality reduction or feature filtering, and spatiotemporal registration achieves consistency of multi-source information through time synchronization and spatial coordinate alignment;

[0044] Let the first An intelligent agent at time The local perceptual feature tensor is:

[0045]

[0046] in, Indicates the first An intelligent agent at time The local perceptual feature tensor. This represents a multidimensional feature encoding function. Indicates the first An intelligent agent at time The unified state vector, Represents the third-order tensor space over the real number field. This represents the length of the first dimension of the feature tensor. This represents the length of the second dimension of the feature tensor. This represents the length of the third dimension of the feature tensor;

[0047] To reduce the information transmission overhead between multiple agents, the local perceptual feature tensor is compressed and encoded to obtain a compressed feature tensor:

[0048]

[0049] in, Indicates the first An intelligent agent at time The compressed feature tensor, This represents a compression coding function, which is used to reduce the feature transmission dimension and communication overhead through dimensionality reduction or feature filtering.

[0050] Furthermore, spatiotemporal registration is performed on the compressed feature tensor to obtain the registered feature tensor:

[0051]

[0052] in, Indicates the first An intelligent agent at time The registration feature tensor Represents the spatiotemporal registration function. Indicates the first The time offset of an agent relative to a unified reference time. Indicates the first The spatial transformation matrix from the local coordinate system of an agent to the global reference coordinate system;

[0053] After completing the spatiotemporal registration, for all The registration feature tensors of each agent are fused to construct a globally shared feature tensor:

[0054]

[0055] in, Indicates time The globally shared feature tensor. This represents the feature fusion function of multiple agents. This represents the total number of intelligent agents.

[0056] To verify the effectiveness of tensor quantization representation, compression encoding, and spatiotemporal registration processing in step two of this invention, a comparative experiment on the fusion of perceptual information across multiple scenarios and agents was conducted. The experimental results are as follows: Figure 2 As shown. Figure 2 This is a schematic diagram comparing the errors before and after spatiotemporal registration. Figure 2 (a) shows the spatiotemporal registration dynamic response curve under a typical task scenario. The horizontal axis represents the number of task execution steps, the left side of the vertical axis represents the platform state comprehensive error, and the right side represents the registration accuracy. Figure 2 Figure (b) shows the box plots of the average registration error for 10 scenarios with different levels of complexity, comparing the error distribution characteristics of three methods: initial unregistered methods, traditional single spatiotemporal registration, and the registration enhancement model proposed in this invention. Experimental results show that the average registration error of the initial unregistered method is approximately 1.6, while the average error using only the traditional time synchronization and spatial coordinate alignment method is approximately 0.16. The registration enhancement model proposed in this invention, which combines tensor quantization feature encoding, further reduces the average registration error to 0.12, significantly improving the spatiotemporal consistency of multi-source heterogeneous sensing information and providing a reliable data foundation for the accurate construction of subsequent globally shared features and distributed collaborative decision-making.

[0057] Step 3: Construct a large-scale airborne task planning model based on the unified state information and globally shared features of each agent, execute task planning reasoning on the airborne platform, and output the task planning results corresponding to each agent;

[0058] Construct the first An intelligent agent at time Task planning results:

[0059]

[0060] in, The parameter is A large-scale model for airborne mission planning. Indicates the first An intelligent agent at time The unified state vector, Indicates time The globally shared feature tensor. Indicates the first An intelligent agent at time The target waypoint vector, Indicates the first An intelligent agent at time Task execution instructions, Indicates the first An intelligent agent at time Task priority coefficient, This represents the set of parameters for the large-scale airborne mission planning model.

[0061] A lightweight deployment of the large-scale airborne mission planning model yields a lightweight planning model suitable for airborne platform operation:

[0062]

[0063] in, This represents the lightweight airborne mission planning model. This represents the original large-scale model of airborne mission planning. This represents the lightweight deployment mapping function of the model. Represents the set of model compression operations. Represents the set of model quantization operations. This represents the set of parameters for the lightweighted model;

[0064] The output of the lightweight airborne mission planning model is as follows: An intelligent agent at time The task planning results are represented as follows:

[0065]

[0066] in, Indicates the first An intelligent agent at time Task planning results;

[0067] By using the above method, the unified state information and globally shared features of each agent are input into the airborne task planning model. Combined with model compression and model quantization, task planning inference is performed on the airborne platform, thereby outputting task planning results suitable for distributed collaborative execution.

[0068] Step 4: Based on the task planning results, construct a distributed collaborative decision-making and conflict resolution mechanism, apply safe distance constraints, resource allocation constraints and time sequence coordination constraints to the planning actions of each agent, and filter and correct the action results to obtain collaborative decision results that meet the collaborative safety requirements;

[0069] Based on the An intelligent agent at time Based on the task planning results, construct its collaborative action vector:

[0070]

[0071] in, Indicates the first An intelligent agent at time The cooperative action vector, Indicates the first An intelligent agent at time The target waypoint vector, Indicates the first An intelligent agent at time Task execution instructions, Indicates the first An intelligent agent at time Task priority coefficient;

[0072] A uniform constraint is applied to the cooperative action vector, expressed as follows:

[0073]

[0074] in, Indicates the first An intelligent agent at time The position vector, Indicates the first An intelligent agent at time The position vector, Describing the Euclidean norm, This represents the minimum safe distance threshold between agents. Indicates the first An intelligent agent at time For the first The amount of resources used. Indicates the first An intelligent agent at time For the first The amount of resources used. Indicates the first The maximum total amount of such resources available. Indicates the first Each agent corresponds to the end time of the task. Indicates the first Each agent corresponds to the start time of the task. Indicates the first The first agent and the second The minimum time interval required for task coordination between agents;

[0075] Based on the aforementioned unified constraints, the cooperative action vectors of each agent are filtered and modified to obtain the first... An intelligent agent at time Collaborative decision-making results:

[0076]

[0077] in, Indicates the first An intelligent agent at time The collaborative decision-making results after satisfying the constraints. This represents the action correction function. Indicates the first An intelligent agent at time The collaborative action vector;

[0078] By applying the above methods to the planning actions of each agent, safety distance constraints, resource allocation constraints, and timing coordination constraints are imposed. The action results are then filtered and corrected using an action correction function, thereby obtaining a collaborative decision result that meets the requirements of distributed collaborative security.

[0079] To verify the security and collaborative accuracy of the distributed collaborative decision-making and conflict resolution mechanism in step four of this invention, an experiment was designed involving four unmanned aerial vehicles (UAV-1 to UAV-4) in a collaborative target area assault mission. The experimental results are as follows: Figure 3 As shown. Figure 3 This is a schematic diagram illustrating the trajectory and time-series recovery verification before and after conflict resolution, where... Figure 3 (a) shows the initial planned trajectory without cooperative constraints, indicating that there are multiple trajectory intersection and collision risks between UAV-1 and UAV-2, and between UAV-3 and UAV-4. Figure 3 (b) shows the flight trajectory after being corrected by the triple constraint conflict resolution mechanism of this invention. All UAV trajectories do not intersect and maintain a uniform safe distance. Figure 3 The curve in (c) shows the comparison of synchronous arrival error in the target area. The horizontal axis represents the mission execution time, and the vertical axis represents the time difference between the arrival of each UAV at the target point. After the conflict is resolved, the synchronous arrival error of the cluster is controlled within 2.5s, which meets the timing requirements of the coordinated strike mission. Figure 3 In the middle (d), the curve of minimum inter-machine distance changing over time is marked, indicating the system's preset 5m safety distance threshold. After conflict resolution, the inter-machine distance at all task times is higher than this threshold, eliminating the risk of collision and verifying the effectiveness and robustness of the distributed collaborative decision-making mechanism of the present invention under the condition of no central node.

[0080] Step 5: Based on the collaborative decision-making results, construct a virtual-real feature offset analysis mechanism to characterize the differences between the simulation expected features and the actual execution feedback features, and obtain virtual-real migration difference information;

[0081] Build Time The difference between real and virtual values:

[0082]

[0083] in, Indicates time The difference between the virtual and real quantities, Indicates time The simulation expected feature tensor generated based on the collaborative decision-making results Indicates time The real feedback feature tensor obtained during actual execution. The norm of a vector or tensor is used to measure the degree of deviation between the expected features of the simulation and the actual feedback features.

[0084] Construct virtual-physical migration difference information composed of environment offset, perception offset, and execution offset:

[0085]

[0086] in, Indicates time Information on the difference between virtual and real migration Indicates the environmental feature offset. This represents the offset of the perceived result. Indicates the offset of the action execution;

[0087] Furthermore, the virtual-to-real feature offset and the virtual-to-real migration difference information are combined as the virtual-to-real offset analysis result, expressed as:

[0088]

[0089] in, Indicates time The results of the virtual-real offset analysis, Indicates time Offset of virtual and real features;

[0090] The differences between real-world environmental features and virtual-mapped environmental features are characterized by the above methods. The virtual-real offset analysis results are constructed by combining environmental offset, perception offset and execution offset, thereby providing input for the subsequent hierarchical comprehensive evaluation of closed-loop results.

[0091] Step Six: Based on the virtual-real migration difference information and task execution results, construct a closed-loop result hierarchical comprehensive evaluation mechanism. Use the analytic hierarchy process (AHP)-entropy weight method to integrate subjective and objective weights on the evaluation indicators. Feed real-world environmental data and virtual-mapped environmental data into the current version of the airborne mission planning big model in parallel to obtain the corresponding real-world comprehensive score and virtual comprehensive score, so as to provide a basis for the subsequent collaborative evolution and update of the big model.

[0092] Construct the first The combined weights of the evaluation indicators:

[0093]

[0094] in, Indicates the first The combined weights of the evaluation indicators Indicates the first Subjective weighting of each evaluation indicator Indicates the first The objective weight of each evaluation indicator This indicates the total number of evaluation indicators. Indicates the index of evaluation indicators;

[0095] By feeding real-world environmental data and virtual-mapped environmental data in parallel into the current version of the large-scale airborne mission planning model, two sets of corresponding inference decision outputs are obtained:

[0096]

[0097] in, Indicates time Real-world scenario-based decision-making output Indicates time The virtual mapping environment inference decision output, Indicates time Input data from the real-world environment Indicates time Input data to the virtual mapping environment;

[0098] Based on the combined weights, the decision outputs from the real-world environment simulation and the virtual-mapped environment simulation are comprehensively scored to obtain the real-world comprehensive score and the virtual comprehensive score:

[0099]

[0100] in, Indicates time The actual overall score, Indicates time The virtual composite score, Indicates the first The combined weights of the evaluation indicators Indicates the first The standardization operator for each evaluation indicator is used to standardize the corresponding inference decision output;

[0101] Construct a closed-loop evaluation difference:

[0102]

[0103] in, Indicates time The difference between the virtual and real comprehensive scores, Indicates time The virtual composite score, Indicates time The actual overall score;

[0104] Based on the system's allowed normal disturbance tolerance threshold Construct closed-loop result evaluation criteria:

[0105]

[0106] in, This represents the system's permissible normal disturbance tolerance threshold. Indicates time The difference between the virtual and real combined scores;

[0107] By using the above method, the inference decision results in the real environment and the virtual mapping environment are scored in parallel and compared with the difference to obtain the closed-loop result hierarchical comprehensive evaluation result, thereby providing a basis for the subsequent collaborative evolution update of the airborne mission planning large model or the correction of the digital twin mapping.

[0108] Step 7: Based on the evaluation results, construct a large-scale model co-evolution update mechanism. When the closed-loop result hierarchical comprehensive evaluation determines that the comprehensive score of the airborne mission planning large model in the virtual environment is significantly higher than the comprehensive score in the real environment, the large-scale model evolution update is triggered. The large-scale model co-evolution update mechanism includes local case fine-tuning, sparse gradient sharing, and confidence-driven federated aggregation.

[0109] When the An intelligent agent at time When an evolutionary update is triggered, the low-scoring edge scenario data encountered in the real environment is stored in a local case memory, and the airborne mission planning model is fine-tuned locally based on the local case memory, as shown below:

[0110]

[0111] in, Indicates the first An intelligent agent at time The local model parameter set, Indicates the first An intelligent agent at time The set of model parameters after fine-tuning for local cases. This indicates local fine-tuning of the learning rate. Indicates the first Each agent uses a case fine-tuning loss function built based on its local case memory. This represents the gradient of the fine-tuned loss function relative to the local model parameters in the given case;

[0112] Based on the locally fine-tuned model gradient, extract the top gradient values ​​with the largest gradient magnitudes. The components form a sparse shared gradient, represented as:

[0113]

[0114] in, Indicates the first An intelligent agent at time sparse shared gradient, This indicates that the gradient with the largest magnitude is retained. Sparsity operators for each component, This indicates the number of gradient components retained. Indicates the first Fine-tuning the gradient of the loss function with respect to the model parameters for each agent in local cases;

[0115] The actual comprehensive score obtained from the hierarchical comprehensive evaluation of the closed-loop results is converted into the evolutionary confidence weight of the nodes, expressed as:

[0116]

[0117] in, Indicates the first An intelligent agent at time The evolutionary confidence weights, Indicates the first An intelligent agent at time The actual overall score, Indicates the smoothing temperature coefficient. Indicates the first The set of neighboring nodes of an agent in an ad hoc network Indicates the neighbor node index;

[0118] Based on the evolutionary confidence weights, a decentralized federated aggregation of the sparse gradients shared by neighboring nodes is performed to update the [number]th [node]. The local model parameters of each agent are represented as follows:

[0119]

[0120] in, Indicates the first An intelligent agent at time The set of model parameters after decentralized federated aggregation Indicates the first An intelligent agent at time The set of model parameters after fine-tuning for local cases. Indicates the federated learning rate. Indicates the first The neighboring nodes at time... The evolutionary confidence weights, Indicates the first The neighboring nodes at time... sparse shared gradient, Indicates the first The set of neighboring nodes of an agent.

[0121] To verify the superiority of the large-scale model collaborative evolution update mechanism of this invention, four sets of control experiments were set up: no adaptation, local feedback update, federated shared adaptation, and the virtual-real closed-loop evolution of this invention. The performance differences of 30 rounds of evolution were compared, and the results are as follows: Figure 4 As shown. Figure 4 Image (a) illustrates the convergence process of the virtual-real offset. Figure 4 Figure (b) illustrates the changes in the overall task reward. Experiments show that the performance of the non-adaptive method continues to degrade, and although local updates and traditional federated learning improve performance, they converge slowly. The method of this invention reduces the virtual-real offset to 0.38 and converges after 25 rounds, achieving an overall score of 85. Its performance is consistently the best, and the gap widens with iterations, verifying that it can simultaneously utilize local and global experience to correct the virtual-real bias, achieving faster convergence and better long-term adaptive capability.

[0122] It should be noted that the above description of the embodiments is only for the purpose of helping to understand the method and core idea of ​​this application. For those skilled in the art, several improvements and modifications can be made to this application without departing from the principle of this application, and these improvements and modifications are also within the protection scope of the claims of this application.

Claims

1. A task planning method for a large-scale swarm intelligence model of unmanned systems based on virtual-real migration and feedback evolution, characterized in that... Includes the following steps: Step 1: Construct a local environment perception and sharing distribution mechanism for unmanned system cluster task scenarios, perform unified state representation of the agent state information and environmental constraint information of each agent, and extract local perception features to provide basic input for multi-agent collaborative planning; Step 2: Tensor quantization, compression coding and spatiotemporal registration are performed on the local perception features of each agent. Unified global shared features are formed through time synchronization and spatial coordinate alignment to reduce communication overhead and improve the consistency of multi-source information fusion. Step 3: Based on the unified state information and globally shared features, construct a large airborne mission planning model, and combine model compression and model quantization to form a lightweight mission planning model suitable for airborne platform deployment, so as to output the mission planning results corresponding to each intelligent agent. Step 4: Based on the task planning results, construct a distributed collaborative decision-making and conflict resolution mechanism, apply safe distance constraints, resource allocation constraints and time sequence coordination constraints to the planning actions of each agent, and filter and correct the action results to obtain collaborative decision results that meet the collaborative safety requirements; Step 5: Based on the collaborative decision-making results, construct a virtual-real feature offset analysis mechanism to characterize the differences between real environment features and virtual mapped environment features, and obtain virtual-real migration difference information; Step Six: Based on the virtual-real migration difference information and task execution results, construct a closed-loop result hierarchical comprehensive evaluation mechanism. Use the analytic hierarchy process (AHP)-entropy weight method to perform subjective and objective weighting on the evaluation indicators, and comprehensively score the inference and decision-making results in the real environment and the virtual mapping environment to obtain the evaluation results required for model optimization. Step 7: Based on the evaluation results, construct a large-scale model collaborative evolution update mechanism. Through local case fine-tuning, sparse gradient sharing, and confidence-driven federated aggregation, the airborne mission planning model is adaptively updated to realize the continuous evolution of the unmanned system's swarm intelligence mission planning capability in a dynamic environment.

2. The task planning method for a large-scale swarm intelligence model of an unmanned system according to claim 1, characterized in that, The collected agent state information includes position vector, velocity vector, and resource vector. The environmental constraint information includes static obstacle information vector, dynamic threat information vector, and communication restriction condition vector. A unified state space representation is constructed using vectorization. Suppose that the unmanned system cluster includes The first intelligent agent, the... An intelligent agent at time The unified state vector is represented as: ； in, Indicates the first An intelligent agent at time The unified state vector, Represents a position vector. Represents the velocity vector. Represents a resource vector. Indicates a static obstacle. Indicates a dynamic threat. Indicates communication constraints. This represents the total number of intelligent agents. Indicates the agent index. Indicates the current moment; Position vector Represented as: ； in, Indicates the first An intelligent agent at time The position vector, , , Indicates the first An intelligent agent at time Three-dimensional spatial coordinates; velocity vector Represented as: ； in, Indicates the first An intelligent agent at time The velocity vector, , , They represent the first An intelligent agent at time The velocity components along the three coordinate axes; Resource Vector Represented as: ； in, Indicates the first An intelligent agent at time resource vector, Indicates the first An intelligent agent at time The remaining energy, Indicates the first An intelligent agent at time Remaining load capacity Indicates the first An intelligent agent at time The remaining task resources; Static obstacle information vector Represented as: ； in, Indicates the first The static obstacle information vector perceived by an agent. Indicates the number of static obstacles. Indicates the first The first agent and the second The distance between static obstacles Indicates the static obstacle index; Dynamic Threat Information Vector Represented as: ； in, Indicates the first An intelligent agent at time The perceived dynamic threat information vector, Indicates the number of dynamic threat sources. Indicates the first An intelligent agent at time With the The distance between dynamic threat sources Indicates the first The first dynamic threat source for the first An intelligent agent at time The intensity of the threat, Indicates a dynamic threat source index; Communication constraint vector Represented as: ； in, Indicates the first An intelligent agent at time The communication constraint vector, Indicates the first An intelligent agent at time Available communication bandwidth Indicates the first An intelligent agent at time Communication latency, Indicates the first An intelligent agent at time The stability coefficient of the communication link.

3. The task planning method for unmanned system swarm intelligence large model according to claim 1, characterized in that: In step two, Zhang quantitative representation includes multi-dimensional feature encoding of multi-agent perception information, compression encoding to reduce communication overhead through dimensionality reduction or feature filtering, and spatiotemporal registration to achieve consistency of multi-source information through time synchronization and spatial coordinate alignment. In step two, the multi-agent perception information is subjected to tensor quantization representation, compression encoding, and spatiotemporal registration processing: Let the first An intelligent agent at time The local perceptual feature tensor is: ； in, Indicates the first An intelligent agent at time The local perceptual feature tensor. This represents a multidimensional feature encoding function. Indicates the first An intelligent agent at time The unified state vector, Represents the third-order tensor space over the real number field. This represents the length of the first dimension of the feature tensor. This represents the length of the second dimension of the feature tensor. This represents the length of the third dimension of the feature tensor; To reduce the information transmission overhead between multiple agents, the local perceptual feature tensor is compressed and encoded to obtain a compressed feature tensor: ； in, Indicates the first An intelligent agent at time The compressed feature tensor, This represents a compression coding function, which is used to reduce the feature transmission dimension and communication overhead through dimensionality reduction or feature filtering. Furthermore, spatiotemporal registration is performed on the compressed feature tensor to obtain the registered feature tensor: ； in, Indicates the first An intelligent agent at time The registration feature tensor Represents the spatiotemporal registration function. Indicates the first The time offset of an agent relative to a unified reference time. Indicates the first The spatial transformation matrix from the local coordinate system of an agent to the global reference coordinate system; After completing the spatiotemporal registration, for all The registration feature tensors of each agent are fused to construct a globally shared feature tensor: ； in, Indicates time The globally shared feature tensor. This represents the feature fusion function of multiple agents. This represents the total number of intelligent agents.

4. The task planning method for a large-scale swarm intelligence model of unmanned systems based on virtual-real migration and feedback evolution as described in claim 1 is characterized in that... In step three, a large airborne mission planning model is constructed based on the unified state information and globally shared features of each agent. Mission planning reasoning is then performed on the airborne platform, and the mission planning results corresponding to each agent are output. Construct the first An intelligent agent at time Task planning results: ； in, The parameter is A large-scale model for airborne mission planning. Indicates the first An intelligent agent at time The unified state vector, Indicates the first An intelligent agent at time The target waypoint vector, Indicates the first An intelligent agent at time Task execution instructions, Indicates the first An intelligent agent at time Task priority coefficient, This represents the set of parameters for the large-scale airborne mission planning model. A lightweight deployment of the large-scale airborne mission planning model yields a lightweight planning model suitable for airborne platform operation: ； in, This represents the lightweight airborne mission planning model. This represents the original large-scale model of airborne mission planning. This represents the lightweight deployment mapping function of the model. Represents the set of model compression operations. Represents the set of model quantization operations. This represents the set of parameters for the lightweighted model; The output of the lightweight airborne mission planning model is as follows: An intelligent agent at time The task planning results are represented as follows: ； in, Indicates the first An intelligent agent at time The results of the task planning.

5. The task planning method for a large-scale swarm intelligence model of an unmanned system according to claim 1, characterized in that, In step four, a distributed collaborative decision-making and conflict resolution mechanism is constructed based on the task planning results. Safety distance constraints, resource allocation constraints, and timing coordination constraints are applied to the planning actions of each agent. The action results are then filtered and corrected to obtain collaborative decision results that meet the collaborative safety requirements. Based on the An intelligent agent at time Based on the task planning results, construct its collaborative action vector: ； in, Indicates the first An intelligent agent at time The cooperative action vector, Indicates the first An intelligent agent at time The target waypoint vector, Indicates the first An intelligent agent at time Task execution instructions, Indicates the first An intelligent agent at time Task priority coefficient; A uniform constraint is applied to the cooperative action vector, expressed as follows: ； in, Indicates the first An intelligent agent at time The position vector, Indicates the first An intelligent agent at time The position vector, Describing the Euclidean norm, This represents the minimum safe distance threshold between agents. Indicates the first An intelligent agent at time For the first The amount of resources used. Indicates the first An intelligent agent at time For the first The amount of resources used. Indicates the first The maximum total amount of such resources available. Indicates the first Each agent corresponds to the end time of the task. Indicates the first Each agent corresponds to the start time of the task. Indicates the first The first agent and the second The minimum time interval required for task coordination between agents; Based on the aforementioned unified constraints, the cooperative action vectors of each agent are filtered and modified to obtain the first... An intelligent agent at time Collaborative decision-making results: ； in, Indicates the first An intelligent agent at time The collaborative decision-making results after satisfying the constraints. This represents the action correction function. Indicates the first An intelligent agent at time The cooperative action vector.

6. The task planning method for a large-scale unmanned system swarm intelligence model based on virtual-real migration and feedback evolution as described in claim 1 is characterized in that... In step five, a virtual-real feature offset analysis mechanism is constructed based on the collaborative decision-making results to characterize the differences between the simulation expected features and the actual execution feedback features, thereby obtaining virtual-real migration difference information. Build Time The difference between real and virtual values: ； in, Indicates time The difference between the virtual and real quantities, Indicates time The simulation expected feature tensor generated based on the collaborative decision-making results Indicates time The real feedback feature tensor obtained during actual execution. The norm of a vector or tensor is used to measure the degree of deviation between the expected features of the simulation and the actual feedback features. Construct virtual-physical migration difference information composed of environment offset, perception offset, and execution offset: ； in, Indicates time Information on the difference between virtual and real migration Indicates the environmental feature offset. This represents the offset of the perceived result. Indicates the offset of the action execution; Furthermore, the virtual-to-real feature offset and the virtual-to-real migration difference information are combined as the virtual-to-real offset analysis result, expressed as: ； in, Indicates time The results of the virtual-real offset analysis, Indicates time The offset between real and virtual features.

7. The task planning method for a large-scale swarm intelligence model of unmanned systems based on virtual-real migration and feedback evolution as described in claim 1 is characterized in that... In step six, a closed-loop result hierarchical comprehensive evaluation mechanism is constructed based on the virtual-real migration difference information and task execution results. The evaluation indicators are weighted by subjective and objective integration through the analytic hierarchy process-entropy weight method. Real environment data and virtual mapping environment data are fed into the current version of the airborne mission planning big model in parallel to obtain the corresponding real comprehensive score and virtual comprehensive score, so as to provide a basis for the subsequent collaborative evolution and update of the big model. Construct the first The combined weights of the evaluation indicators: ； in, Indicates the first The combined weights of the evaluation indicators Indicates the first Subjective weighting of each evaluation indicator Indicates the first The objective weight of each evaluation indicator This indicates the total number of evaluation indicators. Indicates the index of evaluation indicators; By feeding real-world environmental data and virtual-mapped environmental data in parallel into the current version of the large-scale airborne mission planning model, two sets of corresponding inference decision outputs are obtained: ； in, Indicates time Real-world scenario-based decision-making output Indicates time The virtual mapping environment inference decision output, Indicates time Input data from the real-world environment Indicates time Input data to the virtual mapping environment; Based on the combined weights, the decision outputs from the real-world environment simulation and the virtual-mapped environment simulation are comprehensively scored to obtain the real-world comprehensive score and the virtual comprehensive score: ； in, Indicates time The actual overall score, Indicates time The virtual composite score, Indicates the first The combined weights of the evaluation indicators Indicates the first The standardization operator for each evaluation indicator is used to standardize the corresponding inference decision output; Construct a closed-loop evaluation difference: ； in, Indicates time The difference between the virtual and real comprehensive scores, Indicates time The virtual composite score, Indicates time The actual overall score; Based on the system's allowed normal disturbance tolerance threshold Construct closed-loop result evaluation criteria: ； in, This represents the system's permissible normal disturbance tolerance threshold. Indicates time The difference between the real and virtual scores.

8. The task planning method for a large-scale swarm intelligence model of an unmanned system based on virtual-real migration and feedback evolution as described in claim 1 is characterized in that... In step seven, a large-scale model co-evolution update mechanism is constructed based on the evaluation results. When the closed-loop result hierarchical comprehensive evaluation determines that the comprehensive score of the airborne mission planning large model in the virtual environment is significantly higher than the comprehensive score in the real environment, the large-scale model evolution update is triggered. The large-scale model co-evolution update mechanism includes local case fine-tuning, sparse gradient sharing, and confidence-driven federated aggregation. When the An intelligent agent at time When an evolutionary update is triggered, the low-scoring edge scenario data encountered in the real environment is stored in a local case memory, and the airborne mission planning model is fine-tuned locally based on the local case memory, as shown below: ； in, Indicates the first An intelligent agent at time The local model parameter set, Indicates the first An intelligent agent at time The set of model parameters after fine-tuning for local cases. This indicates local fine-tuning of the learning rate. Indicates the first Each agent uses a case fine-tuning loss function built based on its local case memory. This represents the gradient of the fine-tuned loss function relative to the local model parameters in the given case; Based on the locally fine-tuned model gradient, extract the top gradient values ​​with the largest gradient magnitudes. The components form a sparse shared gradient, represented as: ； in, Indicates the first An intelligent agent at time sparse shared gradient, This indicates that the gradient with the largest magnitude is retained. Sparsity operators for each component, This indicates the number of gradient components retained. Indicates the first Fine-tuning the gradient of the loss function with respect to the model parameters for each agent in local cases; The actual comprehensive score obtained from the hierarchical comprehensive evaluation of the closed-loop results is converted into the evolutionary confidence weight of the nodes, expressed as: ； in, Indicates the first An intelligent agent at time The evolutionary confidence weights, Indicates the first An intelligent agent at time The actual overall score, Indicates the smoothing temperature coefficient. Indicates the first The set of neighboring nodes of an agent in an ad hoc network Indicates the neighbor node index; Based on the evolutionary confidence weights, a decentralized federated aggregation of the sparse gradients shared by neighboring nodes is performed to update the [number]th [node]. The local model parameters of each agent are represented as follows: ； in, Indicates the first An intelligent agent at time The set of model parameters after decentralized federated aggregation Indicates the first An intelligent agent at time The set of model parameters after fine-tuning for local cases. Indicates the federated learning rate. Indicates the first The neighboring nodes at time... The evolutionary confidence weights, Indicates the first The neighboring nodes at time... sparse shared gradient, Indicates the first The set of neighboring nodes of an agent.

Images