Multi-unmanned aerial vehicle formation dynamic obstacle avoidance method and device, and storage medium

Abstract

The invention discloses a multi-unmanned aerial vehicle formation dynamic obstacle avoidance method and device and a storage medium, and belongs to the field of unmanned aerial vehicles. According to the method, the dynamic obstacle speed relevance is modeled by adopting the space-time interaction graph, and the graph convolutional network and a noise enhancement mechanism are combined, so that high-precision trajectory prediction is realized. And designing a dual-channel generative adversarial network, synchronously optimizing the path smoothness and the space-time conflict risk, and evaluating the security of the multi-machine collaborative path in real time in combination with an attention network. The formation control layer breaks through the limitation of traditional local optimization, establishes a multi-agent game model, dynamically balances obstacle avoidance, formation keeping and energy consumption through a utility function, and realizes formation form stability in combination with a centralized training framework. And aiming at a severe environment perception deviation problem, a cross-domain feature alignment technology is adopted, so that the recognition robustness in scenes such as rain and fog is effectively improved. A complete technical closed loop from environmental perception, dynamic prediction to collaborative decision is constructed, and systematic theoretical support is provided for autonomous obstacle avoidance of an unmanned aerial vehicle cluster in a dense dynamic scene.

Description

Technical Field

[0001] The present invention belongs to the field of unmanned aerial vehicles (UAVs), and in particular relates to a dynamic obstacle avoidance method, device and storage medium for a multi-UAV formation. Background Art

[0002] With the widespread application of drone swarms in logistics and distribution, urban security, and emergency rescue, dynamic obstacle avoidance technology for multi-UAV formations has become a core component in ensuring mission safety and efficiency. In complex urban environments, drones must perceive static obstacles (such as buildings and overpasses) and dynamic obstacles (such as birds and other drones) in real time, and rapidly plan paths and adjust formations based on high-precision environmental models. However, traditional approaches face significant bottlenecks in the following areas: Multimodal sensor data suffers from insufficient environmental modeling accuracy (errors > 15 cm) due to hardware synchronization errors and cross-modal feature mismatches, making it difficult to eliminate ghosting from highly reflective surfaces; dynamic obstacle trajectory prediction relies on linear filtering or low-speed motion assumptions, making it difficult to effectively handle the sudden turns of high-speed targets (≥10 m / s), making it difficult to balance prediction error with real-time performance; and the path planning algorithm is decoupled from the formation control strategy, ignoring UAV dynamic constraints (such as maximum acceleration and curvature limits). This results in formation distortion rates exceeding 5% in actual flight, degrading path smoothness and energy efficiency. In addition, perception degradation in harsh environments such as rain, fog, and low light further restricts system reliability, and there is an urgent need to achieve breakthroughs in full-link performance through technological innovation.

[0003] Traditional multi-UAV formation obstacle avoidance methods suffer from the following technical flaws: 1) Inaccurate static environment modeling (error > 15%) makes it difficult to cope with dense urban obstacle scenarios; 2) Dynamic obstacle trajectory prediction is slow (response time > 500), making it incapable of handling high-speed moving targets such as flocks of birds; 3) Path planning algorithms fail to consider formation dynamics constraints, resulting in formation distortion rates exceeding 5%; 4) Perception failure is a significant issue in harsh environments, with recognition accuracy below 80% when visibility falls below 100°. Existing technologies struggle to meet the high safety and real-time requirements of complex urban environments. Summary of the Invention

[0004] The purpose of the present invention is to provide a multi-UAV formation dynamic obstacle avoidance method, device and storage medium to address the bottlenecks and challenges in the existing technology.

[0005] A multi-UAV formation dynamic obstacle avoidance method comprises the following steps:

[0006] Step 1: Acquire multi-source data through a multimodal sensing array deployed on each UAV, and pre-process the collected multi-source data to obtain pre-processed data;

[0007] Step 2: Based on the conditional generative adversarial network (cGAN), the preprocessed data is fused, a geometric consistency constraint loss function is introduced, and a static high-precision 3D environment model is constructed through the iterative closest point (ICP) algorithm.

[0008] Step 3: Based on the static high-precision 3D environment model, a spatiotemporal interaction graph is constructed to characterize the coupling relationship between dynamic obstacles. When the speed difference between any two dynamic obstacles exceeds the set threshold, a multi-layer spatiotemporal graph convolutional network (ST-GCN) is used to predict the trajectory.

[0009] Step 4: Combine the static high-precision 3D environment model with the multi-layer spatiotemporal graph convolutional network (ST-GCN) to predict the trajectory of dynamic obstacles and import it into the digital twin platform to rehearse the formation collaborative movement;

[0010] Step 5: Combining the predicted trajectory with the formation dynamics constraints, a safe path is generated through a dual-channel GAN path generator. The spatiotemporal conflict risk of the safe path is evaluated using a spatiotemporal discriminator to determine the feasible path.

[0011] Step 6: Optimize the formation through Nash equilibrium and the centralized training-distributed execution (CTDE) architecture. A utility function is used to balance obstacle avoidance, formation, and energy consumption, and output a UAV formation decision that can avoid obstacles and consume less energy.

[0012] Step 7: When the sensors on the drone detect an environmental anomaly during the decision-making process, the conditional variational autoencoder (CVAE) data augmentation module is activated to fuse the synthetic data generated by the physics engine with the real sensor input. By minimizing the domain adaptation loss and aligning the feature distribution, the enhanced data is input into the target detection network, driving the distributed controller to execute the optimized path and calibrate the formation shape in real time.

[0013] Furthermore, the preprocessed data includes millimeter-wave radar point cloud, visual images and IMU data.

[0014] Furthermore, in step 2, the pre-processed data is fused based on the conditional generative adversarial network (cGAN), a geometric consistency constraint loss function is introduced, and a static high-precision three-dimensional environment model is constructed by iterative closest point (ICP) algorithm. Specifically, the radar point cloud data is projected into the visual image coordinate system, cross-modal feature fusion is performed using the feature pyramid network (FPN) under the cGAN framework, a geometric consistency constraint loss function is introduced, and the point cloud registration error is calculated by iterative closest point (ICP) algorithm. When the joint loss function converges to a threshold, a static high-precision three-dimensional environment model is output;

[0015] The geometric consistency constraint loss function is L geo :

[0016]

[0017] Among them L cGAN (G, D) is the adversarial loss function of cGAN, which measures the game result between the generator G and the discriminator D. ICP(S i ,G(z i )) is the registration error calculated by the iterative closest point ICP algorithm, S i is the radar point cloud, is the point cloud output by the generator, G(z i ) is the generator’s response to the input latent variable z i The generated result is the generated point cloud coordinates, λ1 and λ2 are weight coefficients, balancing the adversarial loss and geometric consistency constraints, and N is the number of point cloud samples.

[0018] Furthermore, in step 3, a time-space interaction diagram is constructed to represent the dynamic obstacle coupling relationship, wherein the time-space interaction diagram for:

[0019]

[0020] Among them A t is the spatiotemporal adjacency matrix, encoding the interaction relationship between obstacles. When there is a certain speed difference, a connection edge is established. V t is the velocity vector of the dynamic obstacle, X [t-τ,t] is the historical trajectory data within the time window, It is the coupled operation between matrix and vector.

[0021] Furthermore, in step 3, a multi-layer spatiotemporal graph convolutional network (ST-GCN) is used for trajectory prediction. Specifically, the historical trajectory sequence is input, Gaussian noise is injected to enhance the generalization ability of the model, and the predicted trajectory of dynamic obstacles in the next few seconds is output, achieving minimal error between the predicted result and the actual trajectory, thereby enhancing robustness.

[0022] Furthermore, the GAN path generator loss function L in step 5 is G And the spatiotemporal discriminator output D(P) are:

[0023] The generator loss function L G for:

[0024]

[0025] Among them, k k is the curvature of path point k, p k is the path point position, o j is the obstacle position, α and β are weight coefficients, which represent the balance between path smoothness and collision risk;

[0026] The output D(P) of the spatiotemporal discriminator is:

[0027] D(P)=GAT({p i},{o j})

[0028] Among them, GAT is a graph attention network that calculates the path point p i With obstacles o j The probability of space-time conflict.

[0029] Furthermore, the utility function in step 6 is U i for:

[0030]

[0031] Among them, d i is the obstacle avoidance safety distance, θ i is the formation angle deviation, cosθ i Measure the formation maintenance, P motor is the motor power consumption, estimated by the ESC current feedback;

[0032] The centralized training-distributed execution CTDE architecture adopts the proximal policy optimization PPO algorithm and updates the policy network parameters through the policy gradient update model;

[0033] The policy gradient update model for:

[0034]

[0035] Among them, A GAE is the generalized advantage estimation function, π θ (a|s) is the action probability distribution output by the policy network.

[0036] Furthermore, the domain adaptation loss function L MMD for:

[0037] L MMD =||E p [φ(x sim )]-E q [φ(x real )]|| 2

[0038] Among them, φ(·) is the feature mapping function, which transforms the simulation data x sim With the real data x real Projected into the latent space, E p 、E q is the expectation of the distribution of simulation and real data;

[0039] The data enhancement model x~ of the conditional variational autoencoder CVAE data enhancement module is:

[0040]

[0041] where x real is the real sensor data, x sim Synthetic data generated for the physics engine.

[0042] A computer device / equipment / system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-mentioned multi-UAV formation dynamic obstacle avoidance method.

[0043] A computer-readable storage medium stores a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned multi-UAV formation dynamic obstacle avoidance method.

[0044] The beneficial effects of the present invention are:

[0045] The present invention fuses multi-source data through the conditional generative adversarial network (cGAN), introduces geometric constraints to optimize feature matching, and significantly improves the accuracy of environmental modeling.

[0046] To address the difficulty in predicting the motion of groups of dynamic obstacles, traditional single-target models are unable to effectively capture the interactions between obstacles. This solution innovatively uses a spatiotemporal interaction graph to model the velocity correlation of dynamic obstacles, and combines the graph convolutional network ST-GCN with a noise enhancement mechanism to achieve high-precision trajectory prediction.

[0047] In terms of path planning, traditional algorithms separate dynamic constraints from obstacle avoidance, leading to sudden path changes and formation instability. This system employs a dual-channel generative adversarial network to simultaneously optimize path smoothness and spatiotemporal conflict risk, incorporating an attention network to assess the safety of multi-machine collaborative paths in real time. Formation control overcomes the limitations of traditional local optimization by establishing a multi-agent game model. This model dynamically balances obstacle avoidance, formation maintenance, and energy consumption through a utility function, achieving formation stability through a centralized training framework.

[0048] Furthermore, to address perception bias in harsh environments, cross-domain feature alignment technology is employed to effectively improve recognition robustness in scenarios like rain and fog. The entire solution establishes a complete technical closed loop from environmental perception, dynamic prediction, to collaborative decision-making, providing systematic theoretical support for autonomous obstacle avoidance by drone swarms in dense, dynamic scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is the overall system architecture diagram of the invention;

[0050] Figure 2 It is the multimodal data fusion flow chart of the invention;

[0051] Figure 3It is a schematic diagram of the dynamic trajectory prediction and digital twin preview described in the invention;

[0052] Figure 4 This is a diagram of the dual-channel GAN path optimization process described in the invention;

[0053] Figure 5 It is the multi-agent collaborative decision-making framework diagram described in the invention;

[0054] Figure 6 This is a schematic diagram of the cross-domain perception enhancement described in the invention. DETAILED DESCRIPTION

[0055] The present invention is further described in detail below with reference to specific embodiments.

[0056] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solution in the embodiment of the present invention will be supplemented below with reference to the drawings in the embodiment of the present invention, so that the solution can be described more clearly and completely.

[0057] Traditional multi-UAV formation obstacle avoidance technology primarily relies on environmental modeling using a single sensor (such as lidar or vision), employing SLAM algorithms for static obstacle detection, combined with Kalman filtering or particle filtering to predict dynamic obstacle trajectories, and then planning paths using RRT or A algorithms. However, these methods have significant drawbacks: static modeling accuracy is low (error > 15cm), making it difficult to handle complex reflective scenes such as glass curtain walls; dynamic obstacle prediction is poorly performed in real time (response time > 500ms), making it incapable of handling the high-speed movement of flocks of birds; path planning ignores the dynamic constraints of the UAVs, resulting in sudden changes in path curvature and formation distortion rates exceeding 5%; and the recognition accuracy of traditional perception algorithms plummets to below 80% in harsh environments such as low light and haze, severely restricting their application in densely populated urban areas.

[0058] The present invention proposes a multi-UAV formation dynamic obstacle avoidance method, device and storage medium, which systematically solves the above technical bottlenecks through multimodal data fusion, spatiotemporal trajectory prediction and collaborative game optimization. Compared with traditional methods, its core innovation lies in: using conditional generative adversarial network cGAN to fuse millimeter wave radar point cloud, visual image and IMU data, introducing geometric consistency constraint loss function, and through iterative closest point ICP algorithm, achieving cross-modal feature alignment, constructing a static high-precision three-dimensional environment model, and improving the accuracy of static high-precision three-dimensional environment modeling to 5cm level (error reduced by 67%), and compressing the modeling time to 50ms, effectively eliminating the glass curtain wall ghost problem; for dynamic obstacle prediction, designing spatiotemporal graph convolutional network ST-GCN, establishing obstacle interaction relationship through speed difference threshold > 3m / s, and combining noise injection to enhance model robustness, so that the trajectory prediction error for the next 3 seconds is < 0.28m, which is 42% higher than the traditional Kalman filter accuracy. At the path planning level, an innovative dual-channel GAN architecture is designed, and the generator output complies with the UAV dynamic constraints (maximum curvature 0.15rad / m, acceleration 8m / s 2 ) smooth path, the discriminator evaluates spatiotemporal conflicts through the graph attention network GAT, screens a safe path within 200ms (2.3 times faster), improves path smoothness by 40%, and reduces energy consumption by 37%. In terms of formation control, a utility function model based on Nash equilibrium is proposed, combining obstacle avoidance benefits, formation retention and energy consumption indicators, and through the centralized training-distributed execution CTDE framework dynamic optimization strategy, the formation distortion rate is stabilized below 2.1% (a 57% reduction compared to the traditional pilot-follower method), and the reconstruction response time reaches sub-seconds. In addition, to address the problem of perception in harsh environments, the conditional variational autoencoder CVAE is used to align the simulation and real data distributions, maintaining a recognition accuracy of 93.5% when visibility is <100m (an increase of 16.5 percentage points compared to the traditional YOLO), with a false detection rate of <1.2%, systematically breaking through the perception bottleneck under rain and fog interference.

[0059] The technical advantages of this invention are concentrated in the full-process closed-loop collaborative capability: from multimodal perception, high-precision modeling, dynamic rehearsal to real-time decision-making, a millisecond-level response link is formed (latency <200ms), and the obstacle avoidance success rate is >95%. It supports 16 drones to work together in a dense area with a building spacing of 15m; the hardware level integrates 5G NR-U low-latency communication (latency <15ms) and edge computing units, and the algorithm layer deeply integrates advanced models such as GAN and ST-GCN to achieve collaborative optimization of software and hardware; it has strong compatibility and scalability, can adapt to new sensors such as LiDAR, and can be expanded to 50+ drone clusters. The interface complies with the ROS2 standard, providing highly reliable and low-latency technical support for scenarios such as urban logistics distribution and emergency rescue.

[0060] The following combination Figure 1 - Figure 6 The specific implementation steps are as follows:

[0061] First, a multimodal perception array consisting of a millimeter-wave radar (working frequency 77GHz, detection distance 150m), a visible light camera (2 million pixels, frame rate 30fps) and a ten-axis IMU (sampling rate 1kHz) is deployed on each drone. Each sensor uses a hardware synchronization module to ensure that the data timestamp alignment accuracy is less than 5ms. The collected raw data is pre-processed by the onboard NVIDIA Jetson AGX Xavier computing unit and then analyzed in real time using the GAN environment modeling technology of multimodal data fusion. Specifically, the radar point cloud data is projected into the visual image coordinate system, and the feature pyramid network FPN is used for cross-modal feature fusion under the cGAN framework. The geometric consistency constraint loss function is introduced, and the point cloud registration error is calculated by the iterative nearest point ICP algorithm. When the joint loss function L geo =0.6L cGAN +0.4L ICP When the algorithm converges to a threshold of 0.15, it outputs a static, high-precision 3D environment model with a resolution of 2 cm / pixel. This process completes within 50 milliseconds, achieving a threefold improvement in modeling accuracy compared to traditional SLAM methods and effectively eliminating ghosting caused by complex reflective surfaces such as building glass curtain walls.

[0062] In this implementation, the static high-precision 3D environment model constructed is imported into the digital twin platform and combined with the dynamic obstacle perception data for spatiotemporal evolution. Each drone shares its own perception data through the 5GNR-U communication module (latency <15ms). After the central server integrates the information of each node, it can collaboratively track dynamic targets such as flocks of birds. Extract the target motion velocity vector v i (accuracy ±0.2m / s) and azimuth angle θ i (error < 1°), construct the adjacency matrix A of the spatiotemporal interaction graph t , when the speed difference between any two targets Δv=|v j -v i A three-layer ST-GCN network is used for trajectory prediction. The network input is a 1.5s historical trajectory sequence. The generalization ability of the model is enhanced by injecting Gaussian noise with a standard deviation of σ = 0.05, and the predicted trajectory of dynamic obstacles in the next 3 seconds is output. The mean square error between the predicted results and the true trajectory is less than 0.28m.

[0063] Furthermore, in this implementation, distributed trajectory optimization is initiated based on the dynamic preview results. After each UAV local controller receives the global environment information sent by the central server, the dual-channel GAN path generator analyzes the current airspace state S in real time. t, combined with the UAV dynamic constraints (maximum acceleration 8m / s2, angular velocity limit 1.2rad / s), three candidate paths P1, P2, and P3 are generated. The path curvature is strictly limited to k max =0.15rad / m, to avoid instability of the formation structure due to sharp turns. The spatiotemporal discriminator uses the graph attention network GAT to construct a spatiotemporal conflict heat map and calculate the collision risk index of each path in the next 5 time steps (interval 0.4s). When the path safety score The process completes the decision within 200ms, which is faster than traditional RRT. * The algorithm is 2.3 times faster and the path smoothness is improved by 40%.

[0064] Furthermore, in this embodiment, the adaptive adjustment of the formation morphology is achieved through multi-agent game. When each UAV performs motion planning according to the optimized path, the local controller calculates the utility function including obstacle avoidance benefit, formation maintenance degree and energy consumption in real time. The formation angle deviation θ i The motor power consumption P is measured by fusing visual odometry (VO) with UWB positioning data. motor The central server uses a centralized training architecture, optimizes the PPO algorithm through a proximal policy, and updates the policy network parameters using a policy gradient update model. The GAE advantage function coefficient λ is set to 0.95, and the policy update step size ε is set to 0.2, which stabilizes the morphological distortion rate during formation reconstruction below 2.1%.

[0065] Furthermore, in this embodiment, the reliability in harsh environments is guaranteed by adaptive perception enhancement technology. When the light sensor detects that the ambient illumination is <50lux or the laser radar echo intensity attenuation is >30%, the conditional variational autoencoder CVAE data enhancement module is activated. This module aligns the feature distribution of the synthetic harsh environment data (including special effects such as haze, rain and snow) generated by the Unity physics engine with the real sensor input in the latent space, and minimizes the MMD distance loss L MMD =||E p [φ(x sim )]-E q [φ(x real )]|| 2 Achieving cross-domain adaptation. The enhanced data feeds into the improved YOLO-v11 detection network, increasing obstacle recognition accuracy from 77% to 93.5% in 100m visibility conditions, with a false detection rate below 1.2%. Furthermore, each drone shares enhanced perception results to build a collaborative cognitive map, keeping the formation's obstacle avoidance response time under 250ms in dense fog.

[0066] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A multi-UAV formation dynamic obstacle avoidance method, characterized in that: The following steps are involved: Step 1: Acquire multi-source data through a multimodal sensing array deployed on each UAV, and pre-process the collected multi-source data to obtain pre-processed data; Step 2: Based on the conditional generative adversarial network (cGAN), the preprocessed data is fused, a geometric consistency constraint loss function is introduced, and a static high-precision 3D environment model is constructed through the iterative closest point (ICP) algorithm. Step 3: Based on the static high-precision 3D environment model, a spatiotemporal interaction graph is constructed to characterize the coupling relationship between dynamic obstacles. When the speed difference between any two dynamic obstacles exceeds the set threshold, a multi-layer spatiotemporal graph convolutional network (ST-GCN) is used to predict the trajectory. Step 4: Combine the static high-precision 3D environment model with the multi-layer spatiotemporal graph convolutional network (ST-GCN) to predict the trajectory of dynamic obstacles and import it into the digital twin platform to rehearse the formation collaborative movement; Step 5: Combining the predicted trajectory with the formation dynamics constraints, a safe path is generated through a dual-channel GAN path generator. The spatiotemporal conflict risk of the safe path is evaluated using a spatiotemporal discriminator to determine the feasible path. Step 6: Optimize the formation through Nash equilibrium and the centralized training-distributed execution (CTDE) architecture. A utility function is used to balance obstacle avoidance, formation, and energy consumption, and output a UAV formation decision that can avoid obstacles and consume less energy. Step 7: When the sensors on the drone detect an environmental anomaly during the decision-making process, the conditional variational autoencoder (CVAE) data augmentation module is activated to fuse the synthetic data generated by the physics engine with the real sensor input. By minimizing the domain adaptation loss and aligning the feature distribution, the enhanced data is input into the target detection network, driving the distributed controller to execute the optimized path and calibrate the formation shape in real time.

2. The multi-UAV formation dynamic obstacle avoidance method according to claim 1, characterized in that: The pre-processed data includes millimeter wave radar point cloud, visual image and IMU data.

3. The multi-UAV formation dynamic obstacle avoidance method according to claim 2, characterized in that: In step 2, the pre-processed data is fused based on the conditional generative adversarial network (cGAN), a geometric consistency constraint loss function is introduced, and a static high-precision three-dimensional environment model is constructed by iterative closest point (ICP) algorithm. Specifically, the radar point cloud data is projected into the visual image coordinate system, cross-modal feature fusion is performed using the feature pyramid network (FPN) under the cGAN framework, a geometric consistency constraint loss function is introduced, and the point cloud registration error is calculated by iterative closest point (ICP) algorithm. When the joint loss function converges to a threshold, a static high-precision three-dimensional environment model is output; The geometric consistency constraint loss function is L geo : Among them L cGAN (G, D) is the adversarial loss function of cGAN, which measures the game result between the generator G and the discriminator D. ICP(S i ,G(z i )) is the registration error calculated by the iterative closest point ICP algorithm, S i is the radar point cloud, is the point cloud output by the generator, G(z i ) is the generator’s response to the input latent variable z i The generated result is the generated point cloud coordinates, λ1 and λ2 are weight coefficients, balancing the adversarial loss and geometric consistency constraints, and N is the number of point cloud samples.

4. The multi-UAV formation dynamic obstacle avoidance method according to claim 1, characterized in that: In step 3, a spatiotemporal interaction graph is constructed to represent the dynamic obstacle coupling relationship, wherein the spatiotemporal interaction graph for: Among them A t is the spatiotemporal adjacency matrix, encoding the interaction relationship between obstacles. When there is a certain speed difference, a connection edge is established. V t is the velocity vector of the dynamic obstacle, X [t-τ,t] is the historical trajectory data within the time window, It is the coupled operation between matrix and vector.

5. The multi-UAV formation dynamic obstacle avoidance method according to claim 1, characterized in that: In step 3, a multi-layer spatiotemporal graph convolutional network (ST-GCN) is used for trajectory prediction. Specifically, the historical trajectory sequence is input, Gaussian noise is injected to enhance the generalization ability of the model, and the predicted trajectory of dynamic obstacles in the next few seconds is output. This minimizes the error between the predicted result and the actual trajectory, thereby enhancing robustness.

6. The multi-UAV formation dynamic obstacle avoidance method according to claim 1, characterized in that: The GAN path generator loss function L in step 5 G And the spatiotemporal discriminator output D(P) are: The generator loss function L G for: Among them, k k is the curvature of path point k, p k is the path point position, o j is the obstacle position, α and β are weight coefficients, which represent the balance between path smoothness and collision risk; The output D(P) of the spatiotemporal discriminator is: D(P)=GAT({p i },{either j }) Among them, GAT is a graph attention network that calculates the path point p i With obstacles o j The probability of space-time conflict.

7. The multi-UAV formation dynamic obstacle avoidance method according to claim 1, characterized in that: The utility function in step 6 is U i for: Among them, d i is the obstacle avoidance safety distance, θ i is the formation angle deviation, cosθ i Measure the formation retention, P motor is the motor power consumption, estimated by the ESC current feedback; The centralized training-distributed execution CTDE architecture adopts the proximal policy optimization PPO algorithm and updates the policy network parameters through the policy gradient update model; The policy gradient update model for: Among them, A GAE is the generalized advantage estimation function, π θ (a|s) is the action probability distribution output by the policy network.

8. The multi-UAV formation dynamic obstacle avoidance method according to claim 1, characterized in that: The domain adaptation loss function L MMD for: L MMD [||E p [ϕ(x sim )]-AND q [ϕ(x real )]|| 2 Among them, φ(·) is the feature mapping function, which transforms the simulation data x sim With the real data x real Projected into the latent space, E p 、E q is the expectation of the distribution of simulation and real data; The data enhancement model of the conditional variational autoencoder CVAE data enhancement module for: where x real is the real sensor data, x sim Synthetic data generated for the physics engine.

9. A computer device / apparatus / system comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 8.

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

Images