Unmanned aerial vehicle formation dynamic path planning method and device based on artificial intelligence, and storage medium

Abstract

The invention relates to the technical field of unmanned aerial vehicles, and discloses an unmanned aerial vehicle formation dynamic path planning method and device based on artificial intelligence, and a storage medium, the method adopts a distributed architecture, and each unmanned aerial vehicle independently and parallelly executes the following steps: obtaining environment information through an airborne sensor, exchanging a state with a neighbor unmanned aerial vehicle, and predicting a trajectory; calculating a neighbor attention weight by using a graph attention network; a local model predictive control optimization problem including navigation, safety, intelligent formation and elastic deformation cost items is constructed; closed-loop control is achieved through rolling optimization, the problems of communication bottleneck and single-point failure of traditional centralized control are effectively solved, the communication load is reduced through distributed collaborative decision, self-adaptive formation keeping and dynamic obstacle avoidance are achieved through the artificial intelligence technology, and the method is suitable for large-scale popularization and application. And the reliability, the safety and the cooperation efficiency of the unmanned aerial vehicle formation in a complex environment are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, specifically to an artificial intelligence-based method, apparatus, and storage medium for dynamic path planning of UAV formations. Background Technology

[0002] With the rapid development of drone technology, multi-drone swarms are increasingly being used in complex scenarios such as logistics delivery, disaster relief, and regional patrols. Path planning is one of the core technologies to ensure the efficient and safe execution of missions by drone swarms. Traditional drone swarm path planning methods mainly rely on centralized control architectures or simple distributed coordination strategies.

[0003] In a centralized control architecture, a ground control station or a designated lead drone typically acts as the central processor, collecting perception information from all wingmen, performing unified path planning, and then distributing instructions to each drone.

[0004] Although the above method has simple control logic, it has the following shortcomings: First, all communication between drones must go through the central node. When the formation size increases or the environmental complexity increases, it will generate huge communication bandwidth pressure and high latency, forming a communication bottleneck. Second, once the central node fails, the entire formation system will be paralyzed, posing a single point of failure risk and poor system reliability.

[0005] Therefore, we propose an AI-based method, device, and storage medium for dynamic path planning of drone formations to address the problems mentioned above. Summary of the Invention

[0006] This invention provides an AI-based dynamic path planning method for drone formations, which can solve the problem that in the prior art, all communication between drones in a drone formation must go through a central node. When the formation size increases or the environmental complexity increases, it will generate huge communication bandwidth pressure and high latency, forming a communication bottleneck.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: An AI-based dynamic path planning method for drone formations involves each drone in the formation independently and in parallel performing the following steps: S1. Obstacle information of the local environment is obtained through the onboard sensors of this UAV, and status information and predicted intention trajectory are exchanged with neighboring UAVs within the communication range through the communication module of this UAV; the status information includes at least position and speed, and the predicted intention trajectory is a short-term future trajectory predicted based on its own model. S2. Map the drone itself and all its neighboring drones together as nodes in a dynamic graph to construct a network representing the current formation topology. The drone uses a pre-trained graph attention network (GAT) model to calculate the attention weight between its own node and each neighboring drone node in the graph, so as to quantify the relative importance of different neighboring drones to the current drone's decision. S3. Based on the local environmental obstacle information obtained in step S1, the predicted intention trajectory obtained from the neighboring drone, and the attention weight calculated in step S2, a local model predictive control optimization problem for this drone is constructed. By solving the local model predictive control optimization problem, an optimal flight trajectory for this drone in a short future period is generated, namely the optimal control sequence. S4. Execute the immediate control instructions in the optimal control sequence described in step S3, and return to step S1 in the next sampling period to achieve closed-loop rolling optimization.

[0008] Preferably, the objective function of the local model predictive control optimization problem for this UAV described in step S3 includes the following cost terms: (a) Navigation cost term, used to penalize deviations from the global target point of this UAV; (b) A safety penalty item, used to penalize the potential collision risk with said environmental obstacles and neighboring drones' predicted intended trajectories; (c) Intelligent formation cost term, which is the deviation between the expected relative formation of all neighboring UAVs of this UAV and the attention weight obtained in step S2. (d) Elastic deformation cost term, used to allow temporary elastic deformation of the formation in the presence of collision risk.

[0009] Preferably, the graph attention network (GAT) in step S2 is a multi-head graph attention network, and its calculation process includes: executing K independent attention mechanisms in parallel to obtain K different attention weights; and fusing the K sets of results to generate the final comprehensive attention weight.

[0010] Preferably, the calculation of the safety cost item in step S3 includes: predicting the future movement trajectory of the neighboring drone; calculating the minimum relative distance between the predicted intention trajectory of the current drone and the neighboring drone; and generating a penalty item when the distance is less than a safety threshold.

[0011] Preferably, the intelligent formation cost term is used to measure the degree of deviation between this UAV and all neighboring UAVs in maintaining the desired formation, and its core calculation formula can be expressed as: Formation cost = Sum of [(importance weight) × (deviation between actual and expected relative positions)²] of all neighboring drones; in: "All neighboring drones": refers to all drones within the communication range that require coordination; "Importance weight" ): refers to the attention weight for neighboring drone i calculated through the graph attention network in step S2; "Actual relative position" refers to the actual relative position vector (p-) between this UAV and its neighboring UAV i. ); "Expected relative position" () refers to the relative position vector that this UAV and its neighboring UAV i should maintain in an ideal formation; "Deviation" refers to the difference between the actual relative position and the expected relative position; "Squaring" refers to squaring the deviation, making the cost function more sensitive to larger deviations.

[0012] Preferably, the elastic deformation cost term in step S3 is a monotonically increasing function of the overall formation deformation degree δ, wherein the formation deformation degree δ is used to quantify the degree of distortion of the overall formation; it is measured by calculating the Frobenius norm of a specific matrix, which is composed of the actual relative positions and expected relative positions of all UAVs in the current formation.

[0013] Preferably, the predicted intention trajectory in step S1 is the set of predicted intention trajectory points for the next N steps obtained by the UAV in step S3 in the previous optimization cycle, where N is the prediction time domain length.

[0014] Preferably, the triggering condition for the next sampling period in step S4 is one or a combination of the following conditions: reaching a preset fixed time interval; the state change of this UAV exceeds a preset first dynamic sensitivity threshold; the state change of any neighboring UAV exceeds a preset second dynamic sensitivity threshold.

[0015] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the computer program. The electronic device also includes a sensor interface and a communication module for realizing local perception and collaborative information exchange functions of a UAV. The electronic device is deployed on each UAV in a formation, enabling each UAV to complete distributed collaborative decision-making in parallel and independently.

[0016] A computer-readable storage medium having a computer program stored thereon, the computer program specifically including instructions for performing the following functions: Environmental perception and processing instructions are used to acquire local environmental obstacle information through sensors; Cooperative communication commands are used to exchange status information and predict intended trajectories with neighboring drones via the communication module; The graph attention network computation instruction is used to calculate the attention weights between the node itself and each of its neighboring drone nodes. Distributed model predictive control optimization instructions are used to construct and solve local optimization problems involving multi-objective cost terms; Rolling optimization execution instructions are used to execute the optimal control sequence and achieve closed-loop control; The multi-objective cost term includes a navigation cost term, a safety cost term, an intelligent formation cost term, and an elastic deformation cost term.

[0017] Compared with the prior art, the beneficial effects achieved by the present invention are: The method of this invention avoids dependence on a central node by enabling each UAV to independently and in parallel perform local perception and decision-making, which greatly reduces communication load and latency. UAVs only need to exchange information with neighboring UAVs within the communication range, which enhances system scalability and effectively overcomes the communication bottleneck problem of large-scale formations. At the same time, the distributed architecture eliminates the risk of single point of failure, and the failure of any UAV does not affect the overall operation of the system, which significantly improves system reliability. Simultaneously, a graph attention network is introduced to dynamically calculate the attention weights of neighbors, realizing intelligent collaborative decision-making. By integrating costs such as navigation, safety, formation maintenance, and elastic deformation through a multi-objective framework, the UAV can autonomously achieve a balance between safe obstacle avoidance and formation maintenance in complex environments. Through the elastic deformation mechanism, the formation is allowed to temporarily deform when encountering obstacles, effectively solving the contradiction between safe obstacle avoidance and formation maintenance. This method employs a hybrid triggering mechanism, which ensures both regular optimization and rapid response to unforeseen circumstances, improving the system's adaptability in dynamic environments. Through the implementation of electronic devices and storage media, it ensures that each UAV possesses complete perception, computing, and communication capabilities, enabling the entire formation system to have a high degree of autonomy and coordination. This effectively solves the communication bottlenecks and single-point-of-failure problems of traditional centralized control, and realizes collaborative operation of UAV formations in complex environments. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the dynamic path planning method for UAV formations according to the present invention. Detailed Implementation

[0019] The specific embodiments of the present invention will be described in detail below, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.

[0020] Example 1: Please see Figure 1 The present invention provides a technical solution: A dynamic path planning method for drone formation based on artificial intelligence, wherein each drone in the formation independently and in parallel executes the following steps: S1. Obtain local environmental obstacle information through the onboard sensors of the drone, and exchange status information and predicted intention trajectory with neighboring drones within the communication range through the communication module of the drone; the status information includes at least position and speed, and the predicted intention trajectory is a short-term future trajectory predicted based on its own model; S2. Map the drone itself and all its neighboring drones together as nodes in a dynamic graph to construct a network representing the current formation topology. The drone uses a pre-trained graph attention network (GAT) model to calculate the attention weight between its own node and each neighboring drone node in the graph, so as to quantify the relative importance of different neighboring drones to the current drone's decision. S3. Based on the local environmental obstacle information obtained in step S1, the predicted intention trajectory obtained from neighboring drones, and the attention weights calculated in step S2, a local model predictive control optimization problem for this drone is constructed. The objective function of the optimization problem includes an intelligent formation cost term, which uses the attention weights as weighting coefficients to penalize the deviation between the relative position of the predicted intention trajectory of this drone and the predicted intention trajectories of all neighboring drones and the desired formation. By solving the local optimization problem, the optimal control sequence in the next prediction time domain is generated. S4. Execute the immediate control instructions in the optimal control sequence described in step S3, and return to step S1 in the next sampling period to achieve closed-loop rolling optimization.

[0021] The local model predictive control optimization problem for this UAV described in step S3 has an objective function that includes the following cost terms: (a) Navigation cost term, used to penalize deviations from the global target point of this UAV; (b) A safety penalty item, used to penalize the potential collision risk with said environmental obstacles and neighboring drones' predicted intended trajectories; (c) Intelligent formation cost term, which is the deviation between the expected relative formation of all neighboring UAVs of this UAV and the attention weight obtained in step S2. (d) Elastic deformation cost term, used to allow temporary elastic deformation of the formation in the presence of collision risk.

[0022] The above solution addresses the communication bottlenecks and single points of failure in traditional centralized control architectures by implementing distributed artificial intelligence decision-making and local collaborative optimization for drones.

[0023] This method is executed independently and in parallel by each UAV in the formation. First, it acquires local environmental obstacle information through onboard sensors, and then exchanges status information and short-term predicted trajectories with neighboring UAVs via limited-range communication, thus avoiding the need for global communication. Its key feature is the introduction of a Graph Attention Network (GAT) to dynamically model the formation topology. Each UAV maps itself and its neighbors as dynamic graph nodes, and uses a pre-trained GAT to calculate attention weights, adaptively quantifying the importance of different neighbors to the current decision, thereby achieving intelligent weight allocation in complex environments.

[0024] Based on the above scheme, each UAV constructs a Local Model Predictive Control (MPC) optimization problem, whose objective function incorporates multiple cost terms: navigation cost guides the UAV to move towards the global target point; safety cost avoids environmental obstacles and neighboring UAVs; intelligent formation cost uses attention weights to penalize deviations from the desired formation to ensure cooperative maintenance; and flexible deformation cost allows temporary formation deformation under collision risk to improve flexibility. By solving this optimization problem, the UAV generates the optimal control sequence in the future time domain and performs rolling optimization after executing immediate commands to form closed-loop control.

[0025] This method reduces communication load through distributed parallel computing, avoids central node dependence, and enhances system reliability and scalability. At the same time, GAT weighting and local MPC optimization realize adaptive formation maintenance and obstacle avoidance in dynamic environments, solving the problems of communication latency, single point of failure and coordination efficiency of large-scale formations in highly complex scenarios.

[0026] Furthermore, the graph attention network (GAT) in step S2 is a multi-head graph attention network, and its calculation process includes: executing K independent attention mechanisms in parallel to obtain K different attention weights; and fusing the K sets of results to generate the final comprehensive attention weight.

[0027] In the above scheme, the network executes multiple independent attention mechanisms in parallel. Each mechanism learns and computes a set of attention weights from different representation subspaces, allowing the UAV to simultaneously pay attention to the differences of neighboring UAVs in different feature dimensions. Then, the multiple sets of different weight results are fused to generate the final comprehensive attention weight. By analyzing the information of neighboring UAVs from multiple perspectives, the one-sidedness or potential bias of a single attention mechanism is avoided, so that the final weighted fused attention value can more comprehensively and accurately reflect the true importance of different neighboring UAVs to its current decision, thereby improving the intelligence and reliability of the subsequent formation maintenance cost term calculation.

[0028] The calculation of the safety cost term in step S3 includes: predicting the future trajectory of neighboring drones; calculating the minimum relative distance between the predicted intention trajectories of the current drone and neighboring drones; and generating a penalty term when the distance is less than a safety threshold.

[0029] In the above scheme, the future safety status of the current drone is quantitatively assessed based on the received predicted trajectories of neighboring drones. First, the movement trajectories of neighboring drones over a future time period are predicted. Then, the minimum relative distance between the current drone's predicted trajectory and the predicted trajectories of each neighboring drone is calculated. This minimum distance is a key indicator for measuring potential collision risk. The decision logic is that if the calculated minimum distance is less than a preset safety threshold, a significant penalty term is generated in the safety cost term of the objective function. The value of this penalty term is usually related to the degree of violation, thus guiding the solver to select control commands that increase the spacing and avoid collisions when solving local optimization problems. This achieves a shift from passive static obstacle avoidance to active, predictive dynamic collision avoidance, greatly improving the ability of formations to fly safely in dense and dynamic environments.

[0030] Furthermore, the intelligent formation cost term is used to measure the degree of deviation between this UAV and all neighboring UAVs in maintaining the desired formation, and its core calculation formula can be expressed as: Formation cost = Sum of [(importance weight) × (deviation between actual and expected relative positions)²] of all neighboring drones; in: "All neighboring drones": refers to all drones within the communication range that require coordination; "Importance weight" ): refers to the attention weight for neighboring drone i calculated through the graph attention network in step S2; "Actual relative position" refers to the actual relative position vector (p-) between this UAV and its neighboring UAV i. ); "Expected relative position" () refers to the relative position vector that this UAV and its neighboring UAV i should maintain in an ideal formation; "Deviation" refers to the difference between the actual relative position and the expected relative position; "Squaring" refers to squaring the deviation, making the cost function more sensitive to larger deviations.

[0031] The calculation formula for the intelligent formation cost term involves a weighted sum of squares operation on all neighboring drones. Specifically, the importance weight of each neighboring drone is multiplied by the square of the deviation between its actual and desired relative positions. Then, these products of all neighboring drones are summed. The importance weight is dynamically calculated using a graph attention network, the actual relative position is calculated using real-time state data of the drone and its neighbors, and the desired relative position is derived from a preset ideal formation geometry. The squaring operation ensures that the cost function is more sensitive to larger deviations.

[0032] Intelligent formation maintenance is achieved through a dynamically weighted deviation penalty mechanism, addressing the rigidity of formation maintenance in traditional methods. This scheme utilizes importance weights calculated by graph attention networks, enabling UAVs to adjust their formation maintenance strategies based on the differences in the importance of their neighbors. This ensures accurate coordination among key neighbors while allowing for moderate flexibility among non-key neighbors. Simultaneously, by employing a squared weighted sum calculation method, larger formation deviations are preferentially corrected, thereby achieving adaptive formation maintenance in complex dynamic environments. This effectively solves the problem that fixed-weight strategies cannot adapt to dynamic topology changes and enhances the flexibility of UAV formation systems.

[0033] Example 2: Please see Figure 1 Furthermore, in conjunction with Embodiment 1, it is further found that the elastic deformation cost term in step S3 is a monotonically increasing function with respect to the overall formation deformation degree δ. The formation deformation degree δ is used to quantify the degree of distortion of the overall formation. It is measured by calculating the Frobenius norm of a specific matrix, which is composed of the actual relative positions and expected relative positions of all UAVs in the current formation.

[0034] By defining a monotonically increasing function of overall formation deformation as a cost term, the formation deformation is quantified by calculating the Frobenius norm of a specific matrix, which consists of the actual and expected relative positions of all UAVs in the current formation. The larger the norm value, the more severe the formation distortion. When encountering obstacles or sudden situations requiring avoidance, this cost term allows the formation to undergo moderate elastic deformation rather than being forced to maintain a rigid formation. This provides necessary maneuver space for individual UAVs while maintaining overall cooperation, effectively solving the contradiction between safe avoidance and formation maintenance in traditional methods.

[0035] Furthermore, the predicted intention trajectory in step S1 is the set of predicted intention trajectory points for the next N steps obtained by this UAV in the previous optimization cycle through step S3, where N is the prediction time domain length.

[0036] The above scheme stipulates that each UAV obtains the set of predicted trajectory points for the next N steps by solving the local model predictive control problem in the previous optimization cycle, which serves as the predicted trajectory for the current cycle. This design ensures that the trajectory prediction comes from the latest optimization calculation results, maintaining the timeliness and accuracy of the predicted trajectory, and enabling all UAVs to use the same prediction generation logic, thereby enhancing the consistency of behavior prediction in the distributed system and providing a reliable basis for collaborative decision-making among neighboring UAVs.

[0037] Furthermore, the triggering condition for the next sampling period in step S4 is one or a combination of the following conditions: reaching a preset fixed time interval; the state change of this UAV exceeds a preset first dynamic sensitivity threshold; the state change of any neighboring UAV exceeds a preset second dynamic sensitivity threshold.

[0038] The above scheme adopts a multi-mode triggering mechanism, specifically including three scenarios: reaching a preset fixed time interval; the state change of the UAV exceeding a preset first dynamic sensitivity threshold; or the state change of any neighboring UAV exceeding a preset second dynamic sensitivity threshold. These conditions can be used individually or in combination as triggering conditions. This design ensures that the system can maintain regular optimization and respond immediately when its own or neighboring states change significantly, greatly improving the system's response speed and environmental adaptability to emergencies, and effectively solving the response lag problem of fixed-period sampling in dynamic environments.

[0039] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the computer program. The electronic device also includes a sensor interface and a communication module for realizing local perception and collaborative information exchange functions of a UAV. The electronic device is deployed on each UAV in a formation, enabling each UAV to complete distributed collaborative decision-making in parallel and independently.

[0040] A computer-readable storage medium having a computer program stored thereon, the computer program specifically including instructions for performing the following functions: Environmental perception and processing instructions are used to acquire local environmental obstacle information through sensors; Cooperative communication commands are used to exchange status information and predict intended trajectories with neighboring drones via the communication module; The graph attention network computation instruction is used to calculate the attention weights between the node itself and each of its neighboring drone nodes. Distributed model predictive control optimization instructions are used to construct and solve local optimization problems involving multi-objective cost terms; Rolling optimization execution instructions are used to execute the optimal control sequence and achieve closed-loop control; The multi-objective cost term includes a navigation cost term, a safety cost term, an intelligent formation cost term, and an elastic deformation cost term.

[0041] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A dynamic path planning method for unmanned aerial vehicle (UAV) formations based on artificial intelligence, characterized in that, Each drone in the formation performs the following steps independently and in parallel: S1. Obstacle information of the local environment is obtained through the onboard sensors of this UAV, and status information and predicted intention trajectory are exchanged with neighboring UAVs within the communication range through the communication module of this UAV; the status information includes at least position and speed, and the predicted intention trajectory is a short-term future trajectory predicted based on its own model. S2. Map the drone itself and all its neighboring drones together as nodes in a dynamic graph to construct a network representing the current formation topology. The drone uses a pre-trained graph attention network model to calculate the attention weight between its own node and each neighboring drone node in the graph, so as to quantify the relative importance of different neighboring drones to the current drone's decision. S3. Based on the local environmental obstacle information obtained in step S1, the predicted intention trajectory obtained from neighboring drones, and the attention weights calculated in step S2, a local model predictive control optimization problem for this drone is constructed. The objective function of the optimization problem includes an intelligent formation cost term, which uses the attention weights as weighting coefficients to penalize the deviation between the relative position of the predicted intention trajectory of this drone and the predicted intention trajectories of all neighboring drones and the desired formation. By solving the local optimization problem, the optimal control sequence in the next prediction time domain is generated. S4. Execute the immediate control instructions in the optimal control sequence described in step S3, and return to step S1 in the next sampling period to achieve closed-loop rolling optimization.

2. The AI-based UAV formation dynamic path planning method according to claim 1, characterized in that: The local model predictive control optimization problem for this UAV described in step S3 has an objective function that includes the following cost terms: (a) Navigation cost term, used to penalize deviations from the global target point of this UAV; (b) A safety penalty term, which penalizes the potential collision risk with the environmental obstacles and the predicted trajectory of neighboring drones; (c) Intelligent formation cost term, which is the deviation between the expected relative formation of this UAV and all neighboring UAVs calculated based on the attention weight obtained in step S2; (d) Elastic deformation cost term, used to allow temporary elastic deformation of the formation in the presence of collision risk.

3. The AI-based UAV formation dynamic path planning method according to claim 1, characterized in that, The graph attention network in step S2 is a multi-head graph attention network, and its calculation process includes: executing K independent attention mechanisms in parallel to obtain K different attention weights; and fusing the K sets of results to generate the final comprehensive attention weight.

4. The AI-based UAV formation dynamic path planning method according to claim 1, characterized in that: The calculation of the safety cost term in step S3 includes: predicting the future trajectory of neighboring drones; calculating the minimum relative distance between the predicted intention trajectories of the current drone and neighboring drones; and generating a penalty term when the distance is less than a safety threshold.

5. The AI-based UAV formation dynamic path planning method according to claim 2 or 3, characterized in that, The intelligent formation cost term measures the degree of deviation between this drone and all neighboring drones in maintaining the desired formation. Its core calculation formula can be expressed as: Formation cost = Sum of [(importance weight) × (deviation between actual and expected relative positions)²] of all neighboring drones; in: "All neighboring drones": refers to all drones within the communication range that require coordination; "Importance weight" ): refers to the attention weight for neighboring drone i calculated through the graph attention network in step S2; "Actual relative position" refers to the actual relative position vector (p-) between this UAV and its neighboring UAV i. ); "Expected relative position" () refers to the relative position vector that this UAV and its neighboring UAV i should maintain in an ideal formation; "Deviation" refers to the difference between the actual relative position and the expected relative position; "Squaring" refers to squaring the deviation, making the cost function more sensitive to larger deviations.

6. The AI-based UAV formation dynamic path planning method according to claim 2, characterized in that: In step S3, the elastic deformation cost term is a monotonically increasing function of the overall formation deformation degree δ, whereby the formation deformation degree δ is used to quantify the degree of distortion of the overall formation. It is measured by calculating the Frobenius norm of a specific matrix, which consists of the actual and expected relative positions of all UAVs in the current formation.

7. The AI-based UAV formation dynamic path planning method according to claim 1, characterized in that: The predicted intention trajectory mentioned in step S1 is the set of predicted intention trajectory points for the next N steps obtained by this UAV in the previous optimization cycle through step S3, where N is the prediction time domain length.

8. The AI-based dynamic path planning method for UAV formations according to claim 1, characterized in that, The triggering condition for the next sampling period in step S4 is one or a combination of the following conditions: reaching a preset fixed time interval; the state change of this UAV exceeds a preset first dynamic sensitivity threshold; the state change of any neighboring UAV exceeds a preset second dynamic sensitivity threshold.

9. An electronic device for implementing the AI-based UAV formation dynamic path planning method as described in any one of claims 1 to 8, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor is used to execute the computer program. The electronic device also includes a sensor interface and a communication module for realizing the local perception and collaborative information exchange functions of the UAV. The electronic device is deployed on each UAV in the formation, enabling each UAV to complete distributed collaborative decision-making in parallel and independently.

10. A computer-readable storage medium for implementing the AI-based UAV formation dynamic path planning method as described in any one of claims 1 to 8, wherein a computer program is stored thereon, characterized in that, The computer program specifically includes instructions for implementing the following functions: Environmental perception and processing instructions are used to acquire local environmental obstacle information through sensors; Cooperative communication commands are used to exchange status information and predict intended trajectories with neighboring drones via the communication module; The graph attention network computation instruction is used to calculate the attention weights between the node itself and each of its neighboring drone nodes. Distributed model predictive control optimization instructions are used to construct and solve local optimization problems involving multi-objective cost terms; Rolling optimization execution instructions are used to execute the optimal control sequence and achieve closed-loop control; The multi-objective cost term includes a navigation cost term, a safety cost term, an intelligent formation cost term, and an elastic deformation cost term.

Images