A method and system for cooperative navigation of a drone swarm

Abstract

The application provides a kind of unmanned aerial vehicle cluster cooperative navigation method and system, by decoupling the local observation of unmanned aerial vehicle into three types of feature flow of task navigation, geometric obstacle avoidance and threat exposure response, and asymmetric residual adaptive fusion is carried out on the three types of features, so that the agent can actively adjust the navigation strategy according to the local danger exposure feedback in the case of not obtaining the global threat field. Experiments show that the application can actively bypass the high-risk area at the cost of moderately increasing the path length, maintain a higher final survival margin, and not simply pursue the shortest path or the lowest collision rate, but achieve a more robust dynamic compromise between task completion, geometric safety and danger exposure suppression.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) path planning, and in particular relates to a method and system for collaborative navigation of UAV swarms. Background Technology

[0002] Unmanned aerial vehicle (UAV) swarms sometimes encounter extreme "anti-access / area denial" (A2 / AD) environments during missions. These scenarios are characterized by significant information loss: Global Navigation Satellite System (GNSS) signals are frequently blocked or spoofed, and electromagnetic suppression across the entire frequency band causes frequent interruptions in communication links between UAVs and ground command stations, as well as within the swarm itself. Under these conditions of lacking external navigation references and stable information exchange, UAV swarms must rely on onboard sensing resources to achieve fully autonomous navigation and coordination, posing a severe challenge to the system's individual intelligence and swarm robustness.

[0003] Against this backdrop, the core challenge for UAV swarm navigation is no longer simply finding the shortest geometric path, but rather achieving a dynamic trade-off between mission advancement and survival avoidance under multi-source constraints. Specifically, hazards in strongly denied environments can be categorized into two types: explicit geometric hazards caused by static obstacles and neighboring UAVs, which can usually be directly estimated through local observations such as lidar, relative distance, and relative velocity; and implicit regional hazards caused by factors such as electromagnetic interference, fire coverage, radar illumination, or navigation denial, whose spatial boundaries, center positions, and intensity distributions are typically unknowable to UAVs in advance. For the latter, UAVs often only obtain feedback on local hazard exposure after entering the threat zone through interference intensity, communication quality degradation, decreased navigation reliability, or other onboard warning signals. Therefore, how to infer risk direction using local hazard exposure and its temporal trends in the absence of a global threat map is a key problem that autonomous navigation in strongly denied environments needs to solve.

[0004] For path planning problems in complex environments, existing literature mainly employs control methods based on optimization or artificial potential fields. These methods typically rely on the assumption of global prior information about the environment (such as a global map) or reliable communication topology, aiming to solve for the globally optimal trajectory. However, in strong denial-of-service scenarios, these assumptions are difficult to uphold. On the one hand, unknown terrain makes it impossible to construct an accurate global cost map, resulting in a high risk of collisions for pre-planned trajectories during actual execution. On the other hand, communication denial disrupts state sharing among teammates, causing centralized optimization algorithms to fail to converge due to information gaps. This over-reliance on globally deterministic information often renders traditional methods inadequately robust or even completely ineffective in highly dynamic, highly adversarial, and unstructured environments.

[0005] In recent years, deep reinforcement learning-based methods have been widely applied to navigation tasks under local observation due to their powerful end-to-end perception and decision-making capabilities. Although some multi-agent reinforcement learning (MARL) algorithms alleviate their dependence on global information by introducing local observation mechanisms, existing research still has significant limitations. Most algorithms mainly focus on task efficiency and explicit safety metrics such as path length, time, or energy consumption, while failing to adequately model the intensity of local exposure to implicit regional hazards and their cumulative impact. When the target direction coincides with a high-risk area, agents are prone to taking aggressive paths to obtain short-term task advancement rewards, leading to continuous exposure to high-threat areas. Secondly, conventional policy networks typically simply concatenate task state, radar observation, neighbor state, and risk feedback, lacking a structured decoupling mechanism for heterogeneous intentions such as task navigation, geometric obstacle avoidance, and threat avoidance. When task advancement and hazard avoidance conflict, a single feature fusion channel struggles to stably express the priority changes between different decision intentions. Summary of the Invention

[0006] The main problem addressed by this invention is how to improve the survivability of unmanned aerial vehicle (UAV) swarms in strong denial environments and enable them to successfully reach their destination. This invention provides a method and system for collaborative navigation of UAV swarms.

[0007] To solve the above technical problems, the technical solution adopted is: A method for cooperative navigation of unmanned aerial vehicle (UAV) swarms includes the following steps: Step 1: Treat each drone in the drone swarm as an intelligent agent, and each drone intelligent agent collects observation information at each decision step; Step 2: Each UAV agent preprocesses the collected observation information to obtain preprocessed features; Step 3: Each UAV agent decouples the preprocessed features according to their functions, and inputs each decoupled feature into the network model corresponding to the decoupled function for encoding and representation. The parameters of each network model are independent. Step 4: Perform asymmetric residual fusion on the encoded representations output by each network model to obtain fused features; Step 5: Input the fused features into a temporal recurrent network and output the hidden state at the current time step; Step 6: Input the hidden state at the current moment into the Actor network, and output the mean and standard deviation of the linear acceleration action component and the angular velocity action component respectively. Construct the probability distribution of the continuous stochastic control strategy based on the mean and standard deviation, and sample the linear acceleration command and angular velocity command of each UAV from the probability distribution. Step 7: Each UAV updates its position, velocity, and heading angle according to the obtained linear acceleration and angular velocity commands through the UAV's dynamic equations, and updates the UAV's survivability margin after calculating the local hazard exposure intensity at the current position. If the UAV's survivability margin is lower than the threshold, the mission is considered to have failed. The survivability margin is a state quantity that characterizes the remaining safety margin after the accumulation of hazard exposure. Step 8: Determine whether communication is open at the current moment; if so, exchange status information with neighboring drones within the drone's perception range and update the memory of neighboring drone status information; otherwise, continue to calculate the current position and speed of the neighboring drones based on their historical positions and speeds as the neighboring drone status information. Step 9: Return to Step 1 until the task is completed.

[0008] Furthermore, the observation information includes: mission navigation information, geometric obstacle avoidance information, and threat exposure response information. The mission navigation information includes speed, heading angle, target distance, target relative bearing, survival flag, and the aircraft's state information related to the UAV's own state at the previous moment. The geometric obstacle avoidance information includes UAV onboard radar observation information and neighboring aircraft relative state information with information freshness. The information freshness refers to the ratio of the time length from when the current UAV acquires the latest state of the neighboring aircraft to the current time to the preset longest tolerance interval. The neighboring aircraft relative state information includes the relative speed and relative position of the neighboring aircraft. The threat exposure response information includes survival margin, current risk exposure increment, and risk exposure increment at the previous moment.

[0009] Furthermore, the preprocessed features obtained by each UAV from the collected observation information refer to: The speed, heading angle, target distance and previous action in the mission navigation information are normalized, and the target relative bearing is converted into sine and cosine values ​​to obtain mission navigation features. The 16-dimensional ranging information from the UAV's airborne radar observation information is mapped using trigonometric functions according to the azimuth angles of each radar beam in the aircraft's coordinate system. Each dimension of ranging information is projected onto the horizontal and vertical components of the aircraft's coordinate system to obtain a 32-dimensional two-dimensional Cartesian feature vector. The relative state information of neighboring aircraft with information freshness is input into a relational encoding network, and invalid neighboring aircraft are masked by dynamic masking. After feature aggregation with masking, a 64-dimensional neighboring aircraft relational feature vector is obtained. The 32-dimensional two-dimensional Cartesian feature vector is concatenated with the 64-dimensional neighboring aircraft relational feature vector, and the concatenation result is used as a geometric obstacle avoidance feature. The trend of hazard exposure change is calculated based on the current hazard exposure increment and the hazard exposure increment at the previous moment in the threat exposure response information. The survival margin, the current hazard exposure increment, the hazard exposure increment at the previous moment, and the hazard exposure change trend are normalized and used as the threat exposure response features.

[0010] Furthermore, inputting the decoupled features into the network model corresponding to the decoupling function means: Three functionally independent and parameter-isolated network models are set up. The three network models are a task navigation sub-network for encoding and representing task navigation features, a geometric obstacle avoidance sub-network for encoding and representing geometric obstacle avoidance features, and a threat exposure response sub-network for encoding and representing threat exposure response features. The preprocessed task navigation features, geometric obstacle avoidance features, and threat exposure response features are input into the corresponding network models to obtain task navigation representations, geometric obstacle avoidance representations, and threat exposure response representations.

[0011] Furthermore, the method for obtaining fused features by asymmetric residual fusion of the encoded representations output by each network model is as follows: ; Indicates fusion characteristics, This represents the task navigation representation. This represents the geometric obstacle avoidance representation. This indicates the threat exposure response characterization. Represents the geometric obstacle avoidance gating weights. This indicates the threat exposure gating weight; as the geometric risk increases, Increase the geometric obstacle avoidance modifier. Enhancement occurs when the increment of hazardous exposure increases. Increase the number of threat exposure response modifications. Enhance.

[0012] Furthermore, during the training phase, the Actor network employs an Actor-Critic network architecture to optimize the policy of the Actor network. In this architecture, the Critic network is used to estimate the value of the current state during the training phase and calculate the advantage function based on the value estimation results to assist the Actor network in updating its policy. During the training phase, the Critic network receives not only the current hidden state output by the temporal recurrent network, but also the geometric obstacle avoidance representation output by the geometric obstacle avoidance subnetwork. Threat exposure response characterization output by the threat exposure response subnetwork After the Actor-Critic network is trained, only the Actor network is used during the execution phase, and the Critic network does not participate in the generation of UAV actions. During the execution phase, the Actor network only receives the current hidden state obtained by processing the fused features through a time-series cyclic network, and generates linear acceleration and angular velocity commands for each UAV based on the current hidden state.

[0013] Furthermore, during training, the total reward function for the unmanned aerial vehicle agent includes four parts: mission navigation advancement reward, geometric obstacle avoidance penalty, hazard exposure response penalty, and maneuver energy consumption penalty. ; in, Indicates the first The drone in Rewards for each decision-making step This indicates the weight of the task navigation advancement reward. Indicates the first The drone in Rewards for progressing through task navigation at each decision-making step. Indicates the first The drone in Geometric obstacle avoidance reward for each decision step Indicates the first The drone in The current risk exposure increment at each decision-making step Indicates the first The drone in The energy consumption of each decision-making step. Indicates the geometric obstacle avoidance reward weight. Indicates the penalty weight for dangerous exposure. Indicates the weight of the penalty for vehicle energy consumption; Furthermore, the method for calculating the survival margin is as follows: ; Defined as the first The drone in The survival margin of each decision-making step. Indicates the increment of single-step hazard exposure; when Below the preset threshold In such cases, the drone will fail due to the exhaustion of its survival margin caused by continuous exposure to high-risk areas.

[0014] Furthermore, the method for calculating the trend of change in the hazardous exposure is as follows: ; Indicates the trend of changes in hazardous exposure; Indicates the current increase in risk exposure, This indicates an increase in historical risk exposure. when When the drone's current direction of motion is increasing its risk exposure, it prompts the strategy to proactively explore and change course; conversely, if... This indicates that the drone is moving away from a high-risk area.

[0015] The present invention also provides a collaborative navigation system for unmanned aerial vehicle (UAV) swarms, comprising the following modules: Data Acquisition Module: This module is used by the agents in each UAV to collect observation information at each decision step. The collected observation information includes mission navigation information, geometric obstacle avoidance information, and threat exposure response information. Feature processing module: used to preprocess the collected observation information to obtain preprocessed features, including task navigation features, geometric obstacle avoidance features, and threat exposure response features; Decoupling encoding module: used to decouple the preprocessed features according to their functions; Sub-network modules: These are used to input the decoupled features into the network model corresponding to the decoupling function for encoding and representation. The dual residual gating mechanism module includes an asymmetric residual fusion module for weighted fusion of geometric obstacle avoidance representation and threat exposure response representation in the encoded representation output by each sub-network module to obtain fused features; a geometric obstacle avoidance gating network for adjusting geometric obstacle avoidance gating weights according to static obstacle hazards and adjacent machine collision hazards; and a threat response gating network for adjusting threat exposure gating weights according to hazard exposure increment hazards. Temporal recurrent network: used to input the fused features into the temporal recurrent network and output the hidden state at the current time step; Actor-Critic Network: It is used to input the hidden state at the current moment into the Actor network, and output the mean and standard deviation of the linear acceleration action component and the angular velocity action component respectively. Based on the mean and standard deviation, it constructs the probability distribution of the continuous random control strategy, and samples the linear acceleration command and angular velocity command of each UAV from the probability distribution. Environment interaction module: Used to receive linear acceleration and angular velocity commands output by the Actor network, and apply the linear acceleration and angular velocity commands to the UAV dynamics model to update the UAV's position, velocity, heading, survivability margin, radar observation information and relative state information of neighboring UAVs, so as to obtain the environmental state and observation information after the action is executed; Reward module: It is used to receive the state before action execution, action instruction and state after action execution output by the environment interaction module, calculate the instant reward according to the state transition process, provide feedback to the agent and guide the agent to train.

[0016] By adopting the above technical solution, the present invention has the following beneficial effects: This invention provides a collaborative navigation method and system for unmanned aerial vehicle (UAV) swarms. By decoupling local UAV observations into three feature streams—mission navigation, geometric obstacle avoidance, and threat exposure response—and performing asymmetric residual adaptive fusion on these three features, the agent can proactively adjust its navigation strategy based on local hazard exposure feedback even when a global threat field is unavailable. Experiments show that this invention can proactively avoid high-risk areas at the cost of moderately increasing path length, maintaining a higher final survivability margin. It does not simply pursue the shortest path or the lowest collision rate, but rather achieves a more robust dynamic trade-off between mission completion, geometric safety, and hazard exposure suppression. Attached Figure Description

[0017] Figure 1 A schematic diagram of a strongly denied environment scenario; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a schematic diagram of the hazard perception gating architecture DAGA method of the present invention; Figure 4 This is a schematic diagram of the experimental scenario; Figure 5 This is a schematic diagram of a single-threat field detour scenario. Detailed Implementation

[0018] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Figures 1 to 5 A specific embodiment of the UAV swarm cooperative navigation method of the present invention is shown, such as... Figure 2 As shown, it includes the following steps: Step 1: Treat each drone in the drone swarm as an intelligent agent, and each drone intelligent agent collects observation information at each decision step.

[0020] In a strong denial-of-service (DOA) environment, drones cannot obtain complete and reliable global maps and threat field priors, and can only rely on onboard sensors to obtain local situational information.

[0021] In this embodiment, the observation information includes: mission navigation information, geometric obstacle avoidance information, and threat exposure response information. The mission navigation information includes speed, heading angle, target distance, target relative bearing, survival flag, and the aircraft's state information from the previous moment, all related to the UAV's own state. In this embodiment, the mission navigation information can be calculated from the UAV's dynamics model. (Assuming...) A swarm system composed of isomorphic unmanned aerial vehicles, the first Intelligent agents in The state vector at time t is determined by its position, heading angle, and linear velocity in a two-dimensional inertial coordinate system, denoted as . The evolution of the system state follows the following nonlinear differential equation: ; Among them, the control input vector These represent linear acceleration and angular velocity, respectively. To characterize the performance limitations of a real-world UAV physical system, the model constrains the UAV's velocity, acceleration, and angular acceleration within the following safety envelope: ; In a discrete decision-making framework, the system operates at a fixed time step. Perform state updates. Discretize using the semi-implicit Euler method, the first... The state transition for each decision step can be represented as: ; ; ; ; Since a drone cannot instantaneously change its motion vector direction, its state evolution is strictly limited by its current physical state and control input. According to the drone's dynamic equations, the minimum theoretical braking distance in the current state is... When a drone approaches an obstacle or nearby agent at high speed, even if the strategy has identified a local risk, a collision may be unavoidable due to limited physical deceleration capabilities. In a strong denial environment, such as... Figure 1 As shown, safe avoidance depends not only on geometric path selection, but also on the early perception and timely response to explicit geometric obstacle avoidance hazards and implicit area hazards.

[0022] The geometric obstacle avoidance information includes UAV-borne radar observation information and neighboring UAV relative state information with information freshness. Information freshness refers to the ratio of the time elapsed since the UAV acquired the latest state of a neighboring UAV to the current time to the preset maximum tolerance interval. The neighboring UAV relative state information includes the relative speed and relative position of the neighboring UAVs. In this embodiment, geometric obstacle avoidance hazards originate from static obstacles and neighboring UAVs. These hazards can be directly estimated by observations such as lidar, relative distance, and relative speed. For static obstacles, the UAV employs a local perception method based on airborne lidar, by emitting signals at equal intervals within the horizontal plane of the UAV. A ray is used to obtain distance information of obstacles within the sensing range, including the radial distance of obstacles within the range. and the relative azimuth angle of the velocity direction For neighboring drones, which are dynamic friendly drone agents, it is assumed that the drones can only operate within a limited sensing radius. It acquires the relative state information of neighboring nodes. Limited by sensing bandwidth and computing resources, each drone only extracts the information from the nearest neighboring node. Each neighbor is a dynamic observation object, and its observations include relative position vectors. and relative velocity vector However, due to electromagnetic suppression in a strong denial environment, the communication links within the drone swarm exhibit random discontinuities. The exchange of state information between nodes is no longer continuous, but only occurs at unknown random intervals. The brief opening and immediate disconnection of communication means that the actual movement and intentions of neighboring machines cannot be obtained during the communication interruption, further increasing the uncertainty risk in cluster decision-making. This embodiment introduces information freshness to determine whether the latest status of neighboring machines is timely based on the communication opening time. The longer the time, the greater the uncertainty risk regarding neighboring machines, and the increased risk of collision with neighboring machines.

[0023] The threat exposure response information includes survivability margin, current hazard exposure increment, and hazard exposure increment at the previous moment. In this embodiment, under a strong denial environment (a threat exposure hazardous environment caused by factors such as electromagnetic interference, fire coverage, radar illumination, or navigation denial), the possibility of the UAV obtaining the center, boundary, or global intensity distribution of the threat source is very small. Therefore, it is assumed that the UAV can only obtain local hazard exposure feedback at its current location through onboard sensors. This feedback can come from electromagnetic interference intensity, communication quality degradation, navigation reliability reduction, radar alarm intensity, or other local alarm signals. This embodiment introduces survivability margin to characterize the state quantity of the remaining safety margin after the accumulation of hazard exposure.

[0024] In this embodiment, the survival margin is calculated as follows: ; For the first The drone in The survival margin of each decision-making step. Indicates the first The drone in The incremental risk exposure per decision step; when Below the preset threshold In such cases, the drone will fail due to the exhaustion of its survival margin caused by continuous exposure to high-risk areas.

[0025] Step 2: Each UAV agent preprocesses the collected observation information to obtain preprocessed features.

[0026] In this embodiment, the preprocessed features obtained by each UAV from the collected local observation information refer to: The speed, heading angle, target distance and previous action in the mission navigation information are normalized, and the target relative bearing is converted into sine and cosine values ​​to obtain mission navigation features. The 16-dimensional ranging information from the UAV's airborne radar observations is mapped using trigonometric functions based on the direction angles of each radar beam in the aircraft's coordinate system. Each dimension of ranging information is projected onto the horizontal and vertical components of the aircraft's coordinate system, resulting in a 32-dimensional two-dimensional Cartesian feature vector. The relative state information of neighboring aircraft with information freshness is input into a relational encoding network, and invalid neighboring aircraft are masked using dynamic masking. After masked feature aggregation, a 64-dimensional neighboring aircraft relational feature vector is obtained. The 32-dimensional two-dimensional Cartesian feature vector is concatenated with the 64-dimensional neighboring aircraft relational feature vector, and the concatenation result is used as a geometric obstacle avoidance feature. In this embodiment, the 16-dimensional scalar ranging data acquired by the UAV's airborne lidar lacks intuitive directional information due to beam deflection. To enhance the network's geometric understanding of local safety gaps, the system transforms it into a local repulsive force vector field in the aircraft's two-dimensional Cartesian coordinate system. ,in and These are drones The relative distance in the x and y directions, and respectively with drones The relative distance and relative azimuth angle. To address the uncertainty in neighbor state observation caused by communication limitations due to strong denial environments, this paper introduces an uncertain neighbor relationship encoding mechanism based on information freshness labeling.

[0027] Considering the pulse-like interruptions in communication links of UAV swarms in strongly denied environments, the system no longer passively relies on transiently distorted neighbor observations, but instead uses first-order kinematics to infer the positions and velocities of neighboring UAVs. Let For drones Last successful reception of neighboring machine The timestamp of the status information defines the communication disconnection interval, i.e., the disconnection time. Within communication dead zones, the system no longer relies entirely on distorted historical locations, but instead employs a first-order kinematics inference strategy. This is based on neighboring machines. The last known velocity component Estimate its current expected coordinates : ; Subsequently, the predicted relative displacements were... With relative velocity Similarly, normalization is performed through rotational transformation of the machine's local coordinate system. To alleviate the confidence decay problem caused by the aforementioned position inference, this paper further introduces the Age of Information (AoI) feature for an uncertain neighbor relationship encoding mechanism. The disconnection time is then considered. Based on the preset maximum tolerance interval Normalization yields normalized information freshness. The normalized information freshness is concatenated with the normalized distance, normalized relative position component, and normalized relative velocity component of the neighboring machine in the local coordinate system, and then input into the relation encoding network. A 64-dimensional high-dimensional relation representation for each neighboring machine is obtained through fully connected mapping and nonlinear transformation. For neighboring machines whose predicted distance exceeds a preset inference distance threshold or whose information freshness reaches a preset invalidation threshold, they are marked as invalid inputs using a dynamic mask, and their corresponding high-dimensional relation representations are masked during subsequent feature aggregation. For the remaining valid neighboring machines, the attention score corresponding to their high-dimensional relation representation is calculated, and the attention score is normalized under the constraint of a dynamic mask to obtain the attention weight of each valid neighboring machine. The 64-dimensional high-dimensional relation representations of each valid neighboring machine are then weighted and summed according to the attention weights to obtain a fixed-dimensional 64-dimensional neighboring machine relation feature vector.

[0028] The threat exposure trend is calculated based on the current and previous threat exposure increments in the threat exposure response information. Survival margin, current threat exposure increment, previous threat exposure increment, and threat exposure trend are then normalized and used as threat exposure response features. For environmental threats posed by strongly denied environments, the agent cannot obtain the threat source center, boundary, or global intensity distribution; it can only assess its own situation based on local threat exposure feedback at its current location.

[0029] In this embodiment, the method for calculating the trend of change in hazardous exposure is as follows: ; Indicates the trend of changes in hazardous exposure; Indicates the current increase in risk exposure, This indicates an increase in historical risk exposure. when When the drone's current direction of motion is increasing its risk exposure, it prompts the strategy to proactively explore and change course; conversely, if... This indicates that the drone is moving away from the high-risk area. This embodiment introduces threat exposure response features to simulate the biological instinct to seek advantage and avoid harm. Through these features, a local risk trend judgment basis is provided for the final obstacle avoidance strategy, which does not rely on a global threat map.

[0030] Step 3: Each UAV agent decouples the preprocessed features according to their functions, and inputs each decoupled feature into the network model corresponding to the decoupled function for encoding and representation. The parameters of each network model are independent.

[0031] In this embodiment, inputting the decoupled features into the network model corresponding to the decoupling function means: Three functionally independent and parameter-isolated network models are set up. The three network models are a task navigation sub-network for encoding and representing task navigation features, a geometric obstacle avoidance sub-network for encoding and representing geometric obstacle avoidance features, and a threat exposure response sub-network for encoding and representing threat exposure response features. The preprocessed task navigation features, geometric obstacle avoidance features, and threat exposure response features are input into the corresponding network models to obtain task navigation representations, geometric obstacle avoidance representations, and threat exposure response representations.

[0032] By decoupling preprocessed features and inputting them into three functionally independent and parameter-isolated sub-network models, physical isolation of the decision-making dimension is achieved, reducing the optimization difficulty of fitting complex multi-objective strategies to a single network. First, the task navigation representation, as the backbone of task execution, mainly integrates kinematic states and target relative coordinates, aiming to extract potential representations to ensure smooth task progress. On the other hand, the geometric obstacle avoidance representation is defined as a low-level safety aid, utilizing only Cartesian radar features and the state relationship features of neighboring aircraft to identify local physical information and generate basic collision avoidance correction representations. Because this branch is designed as a weak auxiliary constraint, its weight is kept low during normal navigation, thus avoiding unnecessary interference with task progress caused by overly conservative obstacle avoidance behavior. The threat exposure response representation specifically handles survivability margin, current hazard exposure increment, historical hazard exposure increment, and hazard exposure change trends, used to generate threat exposure response representations when the threat of hazard exposure increases. Because geometric obstacle avoidance signals and threat exposure response signals are often masked by high-frequency fluctuating task coordinate signals in complex decision spaces, the agent's perception of hidden threats is slow. By separating threat exposure response features from general navigation features, a separate threat exposure sub-network learns how to generate threat exposure response representations based on threat exposure levels. This endows the agent with the ability to proactively avoid threats in exposed environments.

[0033] Step 4: Perform asymmetric residual fusion on the encoded representations output by each network model to obtain fused features.

[0034] Before feature fusion, a geometric obstacle avoidance gating network and a threat response gating network are constructed to adaptively adjust the influence of the corresponding encoded representation on the fused features based on the current geometric obstacle avoidance state and threat exposure state of the UAV.

[0035] The geometric obstacle avoidance gating network uses the distance between the drone and the target. Current flight speed and normalized nearest range information obtained from airborne radar observations As input, its input vector is represented as: ; in, For the ranging information detected by the k-th radar, This represents the radar's maximum observation range. The input vector... The geometric obstacle avoidance gating network is input, and after processing with nonlinear mapping and the Sigmoid function, the geometric obstacle avoidance gating coefficients are obtained. ; The threat response gating network is based on the drone's current survivability margin. Current increase in risk exposure and trends in risk exposure As input, its input vector is represented as: ; The input vector is processed by a threat response gating network and a sigmoid function to obtain the threat response gating coefficients. The input vector The input threat response gating network is processed by nonlinear mapping and the sigmoid function to obtain the magnitude of the threat response gating coefficient. ; In this embodiment, the method for obtaining fused features by asymmetric residual fusion of the encoded representations output by each network model is as follows: ; Indicates fusion characteristics, This represents the task navigation representation. This represents the geometric obstacle avoidance representation. This indicates the threat exposure response characterization. Represents the geometric obstacle avoidance gating weights. This indicates the threat exposure gating weight; when the geometric obstacle avoidance hazard increases, Increase the geometric obstacle avoidance modifier. Enhancement occurs when the increment of hazardous exposure increases. Increase the number of threat exposure response modifications. Enhance.

[0036] When the agent detects a changing trend in risk exposure At that time, the dual residual gating mechanism will quickly activate the threat exposure gating. to give Higher intervention priority. This mechanism does not rely on a global threat map, but dynamically adjusts the fusion ratio between mission navigation representation, geometric obstacle avoidance representation, and threat exposure response representation based on local geometric hazards, survivability margins, and hazard exposure trends. When lidar or neighbor status indicates an increased collision risk, geometric obstacle avoidance gating increases the geometric obstacle avoidance gating weight, enhancing the intervention strength of the geometric obstacle avoidance residual, causing the strategy to tend to avoid obstacles or other UAVs. When the hazard exposure increment increases, indicating that the UAV is entering a higher threat area, the threat exposure gating weight increases, enhancing the intervention strength of the threat exposure response residual, causing the strategy to generate maneuver corrections to deviate from high-risk areas while retaining target mission navigation information.

[0037] This embodiment employs a danger perception gating architecture (DAGA) that combines geometric obstacle avoidance gating and threat exposure gating. This architecture uses a decentralized execution model, decoupling local observations into three feature streams: task navigation, geometric obstacle avoidance, and threat exposure response, and achieving adaptive fusion through a dual residual gating mechanism. Unlike directly concatenating all observation information features into a single policy network, this approach... Figure 3 As shown, the Danger Awareness Gated Architecture (DAGA) explicitly distinguishes between directly observable static obstacles and dynamic neighboring aircraft states, as well as threat exposure trends caused by environmental threats. To achieve adaptive feature fusion under different risk states, a geometric obstacle avoidance gating network and a threat response gating network are set up. The geometric obstacle avoidance gating network takes the target distance, current flight speed, and maximum obstacle proximity obtained based on radar ranging information as inputs to generate geometric obstacle avoidance gating weights, which are used to adjust the contribution of geometric obstacle avoidance representations to the fused features. The threat response gating network takes the current survival margin, current hazard exposure increment, and hazard exposure change trend as inputs to generate threat exposure gating weights, which are used to adjust the contribution of threat exposure response representations to the fused features. Through a reinforcement learning training process, the geometric obstacle avoidance gating network and the threat response gating network can learn the mapping relationship between risk states and gating weights, thereby adaptively adjusting the corresponding weights to respond to changes in collision risk or threat exposure risk. This enables the agent to maintain mission navigation propulsion during normal flight, enhance geometric avoidance when obstacles or neighboring aircraft approach, and enhance threat avoidance response when hazard exposure increases.

[0038] Step 5: Input the fused features into a temporal recurrent network (GRU) and output the hidden state at the current moment. The fused features are input into the GRU. Since the spatial distribution of threats in a strongly denied environment cannot be directly observed, trend judgment can only be made through the temporal difference of local hazard exposures. The GRU is used to model the temporal dependency between hazard exposure changes and motion decisions, and is responsible for transforming instantaneous geometric obstacle avoidance features and threat exposure response features into a temporally continuous state representation. This enables cooperative navigation decisions that balance task efficiency, geometric obstacle avoidance safety, and survivability margin maintenance in partially observable strongly denied environments.

[0039] Step 6: Input the hidden state at the current moment into the Actor network, and output the mean and standard deviation of the linear acceleration action component and the angular velocity action component respectively. Construct the probability distribution of the continuous stochastic control strategy based on the mean and standard deviation, and sample the linear acceleration command and angular velocity command of each UAV from the probability distribution.

[0040] In this embodiment, the Actor network adopts an Actor-Critic network architecture to optimize the policy of the Actor network during the training phase. The Critic network is used to estimate the value of the current state and calculate the advantage function based on the value estimation result to assist the Actor network in updating the policy. During the training phase, the Critic network receives not only the current hidden state output by the temporal recurrent network, but also the geometric obstacle avoidance representation output by the geometric obstacle avoidance subnetwork. Threat exposure response characterization output by the threat exposure response subnetwork This improves the ability of the value function to estimate hazards; after the Actor-Critic network is trained, only the Actor network is used in the execution phase, and the Critic network does not participate in the generation of UAV actions.

[0041] During the execution phase, the Actor network only receives the current hidden state obtained by processing the fused features through a time-series cyclic network, and generates linear acceleration and angular velocity commands for each UAV based on the current hidden state to control the UAV.

[0042] Step 7: Each UAV updates its position, velocity, and heading angle according to the obtained linear acceleration and angular velocity commands through the UAV's dynamic equations, and updates the UAV's survivability margin after calculating the local hazard exposure intensity at the current position. If the UAV's survivability margin is lower than the threshold, the mission is considered to have failed. The survivability margin is a state quantity that characterizes the remaining safety margin after the accumulation of hazard exposure. In this embodiment, after updating the drone's position, speed, and heading angle, each drone also needs to calculate its survivability margin in order to determine its current state.

[0043] Step 8: Determine whether communication is open at the current moment; if so, exchange status information with neighboring drones within the drone's perception range and update the memory of neighboring drone status information; otherwise, continue to calculate the current position and speed of the neighboring drones based on their historical positions and speeds as the neighboring drone status information.

[0044] Step 9: Return to Step 1 until the task is completed.

[0045] In this embodiment, the total reward function of the UAV agent during training includes four parts: task progress reward, geometric obstacle avoidance penalty, hazard exposure response penalty, and maneuver energy consumption penalty. ; in, Indicates the first The drone in Rewards for each decision-making step Indicates the weight of the task progress reward. Indicates the first The drone in Rewards for each step of the decision-making process. Indicates the first The drone in Geometric obstacle avoidance reward for each decision step Indicates the first The drone in The current risk exposure increment at each decision-making step Indicates the first The drone in The energy consumption of each decision-making step. Indicates the geometric obstacle avoidance reward weight. Indicates the penalty weight for dangerous exposure. Indicates the weight of the penalty for vehicle energy consumption; The weighting of the task progress reward is as follows: ; in, Indicates the first The drone in The distance between each decision step and the target point This represents the task reward bonus coefficient. Indicates the smoothing factor; The task progress reward This includes rewards for reducing the target distance, rewards for aligning the heading angle, rewards for reaching the target area, penalties for crossing the target area, and time penalties. ; Among them, the first item The reward for reducing the distance to the target represents the amount of time the drone reduces its distance to the target within adjacent time steps. Let be the relative distance between the UAV and the target at decision step t. This is the distance reward coefficient, used to encourage continuous progress towards the target; The second item is Heading angle alignment bonus, among which Let be the angle difference between the heading and the target direction at the current decision step t. The relative angle reward coefficient is used to guide the drone to maintain a target-oriented motion attitude. Third item This represents a fixed, one-time reward given when the drone enters the target threshold area; if the drone does not reach the target, this item is 0. Fourth item This indicates that when the drone is approaching the target However, there is a reverse heading. The fixed boundary violation penalty generated at that time is used to avoid high-speed passage through the target area and oscillation near the target. When the drone does not meet the above two conditions, this item is 0; Fifth item This represents a negative reward that increases over time, used to suppress ineffective retention behavior; The geometric obstacle avoidance reward includes penalties for adjacent aircraft collision risk and penalties for radar close-range collision warnings: ; The penalty for collision risk with adjacent aircraft is as follows: ; in The relative distance between the drone and its nearest neighboring drone. The estimated collision time is obtained by dividing the relative distance by the relative velocity. and . represents the weighting coefficient. To be a decimal that avoids having a denominator of zero; This represents the penalty for near-range collision warning based on the square of the lidar ranging information. This is the penalty coefficient; The penalty for hazardous exposure is calculated based on the current single-step hazardous exposure increment. Weight of risk exposure Multiplying them together yields a method used to suppress the continuous exposure of drones to high-risk areas; The penalty for motor energy consumption The smoothness used to constrain the control behavior of the UAV is composed of the current control input amplitude and its rate of change over time, and is expressed as: ; in . indicates that the drone was in the . The acceleration of the decision step This is a weighting coefficient used to balance the influence of control amplitude and control rate of change.

[0046] In this embodiment, in addition to applying a quadratically increasing collision warning penalty to near-range threats detected by radar, the instantaneous survival margin loss and maneuver energy consumption from the previous time step are directly deducted from the total reward function. Since geometric obstacle avoidance information and threat exposure information are used not only as inputs to the three sub-networks for extracting risk response representations but also as inputs to the gating network for generating corresponding gating weights, the risk state can simultaneously influence feature representation and feature fusion processes. This direct coupling, occurring simultaneously at the perception input and reward feedback ends, enables the network to learn geometric obstacle avoidance and threat avoidance strategies more quickly and improves the convergence stability of the training process.

[0047] To evaluate the cooperative navigation performance of the proposed Hazard Awareness Gated Architecture (DAGA) in a strongly denied environment, a two-dimensional continuous multi-UAV simulation platform was built using Python and PyTorch. The simulation environment includes nonholonomic dynamic constraints, local geometric obstacles, multi-agent dynamic interactions, intermittent communication, and an unobservable regional threat field, simulating the multi-source uncertainties faced by UAV swarms performing autonomous navigation tasks under GNSS denial and communication-restricted conditions.

[0048] In each task round, the cluster consists of The system consists of four isomorphic drones. Each drone starts from a given initial area, with the goal of reaching a designated target point within a limited time step, while avoiding collisions with static obstacles and other drones, and minimizing the cumulative exposure to dangers caused by the regional threat field to maintain a high survivability margin. The drone in The control input for the UAV at each decision step is a two-dimensional continuous action: ; in and These represent the normalized linear acceleration command and the normalized angular velocity command, respectively. The environment maps these to actual control quantities based on the maximum acceleration and maximum steering angular velocity, and then performs the mapping according to a fixed time step. Update the drone's position, speed, and heading. To simulate the maneuver limitations of a real drone, the maximum flight speed is set to [value missing] in the experiment. The maximum acceleration is The maximum steering angular velocity is .

[0049] To verify the effectiveness of DAGA, this invention selects three representative deep reinforcement learning algorithms as control methods: PPO (Proximal Policy Optimization): A classic on-policy stochastic policy optimization method used to evaluate the basic performance of the standard Actor-Critic framework in local navigation tasks.

[0050] SAC (Soft Actor-Critic): An off-policy continuous control method based on maximum entropy reinforcement learning, used to compare the performance of stochastic policies with strong exploratory capabilities in threat avoidance tasks.

[0051] MATD3 (Multi-Agent Twin Delayed Deep Deterministic Policy Gradient): A deterministic actor-critic approach for multi-agent continuous control, used to compare the performance of multi-agent deterministic policies in dynamic interaction and collision avoidance tasks.

[0052] DAGA: This invention proposes a hazard perception gating strategy. Under the same observation input conditions, this method decouples the local state from the basic navigation flow, geometric obstacle avoidance flow, and survival stress flow, and performs residual dynamic fusion through geometric obstacle avoidance gating and hazard exposure response gating.

[0053] To eliminate the unfair impact of differences in observational information, all control algorithms use the exact same environmental observation inputs and reward functions. That is, PPO, SAC, MATD3, and DAGA can all receive survival margins, current hazard exposure increments, hazard exposure trends, processed lidar observations, and relative states of neighboring machines labeled with information freshness AoI. The main difference between the methods lies not in the amount of accessible information, but in how the policy network organizes, integrates, and utilizes this heterogeneous information. This setting allows for a more rigorous test of DAGA's core assumption: that, under conditions of equal local information, the decoupling of task navigation—geometric obstacle avoidance—threat exposure response features and the adaptive gating mechanism for hazard perception can more effectively handle the conflict between task advancement and survival avoidance.

[0054] All models were evaluated offline after training, and no parameter updates were performed during the evaluation phase. Multi-agent test results were statistically analyzed on a per-drone basis; that is, if a test included... Launch drones and repeat If there are episodes, then the total number of statistical samples is The experimental parameters are set as shown in Table 1.

[0055] Table 1. Basic Experimental Parameter Settings

[0056] To ensure fairness in the comparison, all methods were trained under the same environment generation rules, observation inputs, reward functions, and training strategies. The differences between the methods mainly lie in the policy optimization algorithms and the way the network structure organizes heterogeneous observation information. DAGA is trained using a PPO-based stochastic policy optimization framework. Its policy network consists of a task navigation subnetwork, a geometric obstacle avoidance subnetwork, a threat exposure response subnetwork, and a dual residual gating module. The dual residual gating module mentioned here refers to the structure formed by asymmetric residual fusion, a temporal recurrent network GRU, and an Actor-Critic network. PPO, SAC, and MATD3, on the other hand, use their respective standard Actor-Critic structures.

[0057] Because strong-denial cooperative navigation tasks involve multiple learning challenges, such as multi-machine collision avoidance, static obstacle avoidance, and strong-denial environmental threat avoidance, training directly from scratch in a complete and complex scenario can easily lead to large-scale collisions or exhaustion of survival margins in the early stages, resulting in sparse effective task rewards. To alleviate these problems, this experiment adopts a three-stage learning strategy, gradually transitioning from simple tasks to complete strong-denial scenarios.

[0058] Phase 1 does not introduce static obstacles or threat fields, only retaining the multi-UAV origin-to-end navigation task. The training objective is to enable the agent to master the basic ability to advance towards the target under constraints of velocity, acceleration, angular acceleration, and dynamics, and to initially learn to avoid dynamic collisions during multi-UAV parallel movement. This phase provides the foundation for subsequent task execution capabilities in complex environments. Phase 2 introduces randomly distributed static circular obstacles. This phase requires the UAVs to learn geometric obstacle avoidance behavior based on local LiDAR observations while maintaining target advancement. Through training in this phase, the policy can form basic local safety response capabilities, avoiding over-coupling of the learning target due to the simultaneous appearance of obstacles and threat fields in Phase 3. Phase 3 further introduces a strong denial regional threat field. This phase simultaneously includes multi-UAV interaction, static obstacles, and implicit regional hazards, requiring the agent to make comprehensive trade-offs between task navigation target advancement, geometric obstacle avoidance, and hazard exposure suppression. By training in stages, the agent first trains its navigation capabilities in a simple environment, then gradually introduces obstacles and threat fields, and allows model parameters to be inherited stage by stage, thereby reducing the difficulty of learning complex strong denial tasks from scratch.

[0059] The main parameters for the course are shown in Table 2. All three phases use a 200m×200m training map and 8 drones. Phase 1 has no obstacles or threat fields; Phase 2 has 8 static obstacles; Phase 3 has both 8 static obstacles and 3 strong denial environment threat fields, with a threat radius of 25m and a threat intensity coefficient of 10.0. Training for each phase... Each stage involves an interactive environment step, and the next stage inherits the model parameters obtained from the previous stage for further training.

[0060] Table 2 Training parameters at different stages

[0061] To evaluate the comprehensive navigation capabilities of different algorithms in the presence of multiple sources of risk, this paper first constructs a structured strong denial test scenario. This scenario aims to simulate the complex situation where a drone swarm encounters mission target attraction, threat blockade in a strong denial environment, local geometric obstacles, and dynamic interactions among multiple drones while performing a traversal mission. Compared to a completely random scenario, this scenario has a more explicit mission conflict structure, enabling a focused examination of whether the strategy possesses the ability to dynamically trade off between the "shortest path" and the "safe survival path."

[0062] The experimental map is a 200m × 200m two-dimensional continuous space. The scene includes 5 drones, 3 static circular obstacles, and 3 threat fields that cannot be directly observed. The initial positions of all drones are concentrated in the following areas: Figure 4 The map shows the lower left area, with target points concentrated in the upper right area. The threat field is deployed diagonally from the lower left to the upper right, creating an approximate blockade of straight paths. Therefore, if the agent only adopts the shortest path strategy according to the target direction, it will inevitably enter the threat area and suffer continuous health loss. Meanwhile, static obstacles are randomly placed in the central area of ​​the map, and multiple drones with similar starting points and target directions will also pose a potential dynamic collision risk. This scenario simultaneously requires the strategy to achieve target advancement, threat avoidance, static obstacle avoidance, and multi-drone collision avoidance, making it the primary test scenario for the comprehensive capabilities of the proposed method.

[0063] In the experiment, the regional threat radius was set to 30m, the threat intensity coefficient to 2.0, the local perception range to 30m, and the maximum number of test steps to 300. Each method was tested independently for 100 episodes under the same settings, and the statistical results were calculated on a per-drone basis. Since each episode contained 5 drones, a total of 500 drone samples were collected for each method. The termination status of each sample was categorized into five types: successful arrival, failure due to insufficient survival margin, static obstacle collision, dynamic collision (collision with other drones), and timeout.

[0064] Table 3 Performance Comparison of Different Methods in Structured Strong Denial Scenarios

[0065] In terms of task completion capability, the DAGA method of this invention achieved the highest task success rate, reaching 83.2%. Compared with PPO, SAC, and MATD3, DAGA's success rate improved by 32.2, 25.6, and 5.4 percentage points, respectively. PPO's success rate in this scenario was only 51.0%, indicating that standard on-policy strategies struggle to stably learn effective safety compromise strategies when simultaneously facing task objectives, threat areas, and multi-machine interactions. SAC's success rate was 57.6%, slightly higher than PPO, but still significantly lower than DAGA. This suggests that while the maximum entropy exploration mechanism can improve continuous control adaptability to some extent, when the threat field has spatial blockade characteristics, simply relying on flattened feature fusion is still insufficient to form stable survival avoidance behavior. MATD3 achieved a success rate of 77.8%, the strongest among the three baseline methods, indicating that the deterministic multi-agent Actor-Critic structure has strong capabilities in continuous control and dynamic collision avoidance; however, its success rate is still lower than DAGA, indicating that explicit health stress regulation still has additional advantages in scenarios with implicit threat fields and significant task progress conflicts.

[0066] In terms of survival margin maintenance, DAGA's advantage is more significant. DAGA's survival margin depletion rate is 0.0%, while PPO, SAC, and MATD3's survival margin depletion rates are 7.8%, 10.4%, and 10.2%, respectively. This indicates that DAGA effectively avoids margin depletion failure due to continuous exposure to danger in this scenario. Meanwhile, DAGA's average survival margin for successful samples reaches 0.938, and the average survival margin for the entire sample reaches 0.946, both the highest values ​​among the four methods. In particular, the Avg Overall Survival Margin metric, because it simultaneously counts successful and failed samples, reflects the algorithm's overall survival quality throughout the entire task better than AvgSuccess Survival Margin.

[0067] This result is consistent with the structural design of DAGA. All control methods receive the same survival margin, current hazard exposure increment, and hazard exposure trend inputs; therefore, DAGA's advantage does not stem from additional observational information, but rather from the structured utilization of heterogeneous information. PPO, SAC, and MATD3 primarily fuse mission, radar, neighbor, and hazard exposure features directly through a unified policy network, easily leading to representational competition between target attraction rewards and threat avoidance requirements. In contrast, DAGA allocates mission navigation, geometric obstacle avoidance, and threat exposure response to different expert streams and dynamically adjusts them through a dual residual gating mechanism. This allows the policy to strengthen the intervention of the threat exposure response branch when the hazard exposure trend increases, thereby proactively deviating from high-risk paths.

[0068] In terms of average reward, DAGA also achieved the highest cumulative reward. Since the reward function includes mission progress, arrival reward, collision penalty, hazard exposure penalty, and energy consumption penalty, the increase in average reward indicates that DAGA achieved a better balance in terms of overall objectives. It is worth noting that MATD3's average reward is close to DAGA's, but its survival margin depletion rate still reaches 10.2%. This indicates that while MATD3 can complete the mission in some samples through strong continuous control, it is still prone to taking paths with high survival margin depletion when traversing threat-blocked areas. In contrast, DAGA significantly reduced the risk of survival margin depletion while maintaining high mission rewards.

[0069] To further analyze the differences in failure modes among different algorithms, Table 4 reclassifies the failure types of each method into three categories: survival margin exhaustion, physical collision failure, and task timeout failure. Among them, physical collision failure is the sum of static collision rate and dynamic collision rate.

[0070] Table 4. Failure types of different methods in structured strong rejection scenarios.

[0071] The failure type decomposition reveals that DAGA has the lowest overall failure rate at 16.8%, with all failures stemming from physical collisions; no health failures or timeouts occurred. This indicates that DAGA demonstrates strong stability in mission progression and threat avoidance, avoiding large-scale timeouts due to over-conservatism and preventing survivability depletion due to neglecting the threat field. In contrast, MATD3 has the lowest physical collision rate at only 9.8%, indicating strong performance in explicit static geometric obstacle avoidance and dynamic collision avoidance; however, MATD3's survivability depletion rate is still 10.2%, suggesting deficiencies in avoiding threats that are not directly observable. PPO and SAC have physical collision failure rates of 34.8% and 32.0%, respectively, indicating that ordinary strategy structures struggle to maintain stable local safety in multi-machine interactions and the coexistence of obstacles.

[0072] It's important to note that DAGA is not a purely geometric obstacle avoidance algorithm, and its core objective is not to achieve the lowest collision rate in all scenarios. Static obstacles and neighbor states can be directly perceived through radar and local neighbor observations. Therefore, methods like the MATD3 strongly continuous control can also learn better reactive obstacle avoidance strategies based on explicit geometric information. Experimental results show that, compared to MATD3, which has the lowest collision rate, DAGA, while not absolutely optimal in terms of physical collision rate, further reduces the overall failure rate by lowering the survival margin depletion rate from 10.2% to 0.0%, achieving the highest task success rate and the highest average survival margin value. Therefore, DAGA's main advantage lies in a better trade-off between task and survival, rather than single-dimensional geometric obstacle avoidance optimization.

[0073] In summary, experimental results in a structured strong denial scenario demonstrate that DAGA can effectively handle the conflict between mission advancement and covert threat avoidance in multi-UAV cooperative navigation. Compared to standard PPO, SAC, and MATD3, DAGA outperforms in success rate, average reward, average survivability margin, and survivability margin depletion rate. These results validate the effectiveness of risk perception and dual-residual gating structures in strong denial environments; that is, by explicitly modeling the increment of hazard exposure and its rate of change, agents can formulate more survivably robust navigation strategies in threat fields where they cannot be directly observed.

[0074] Mechanism verification in a single-threat field bypass scenario: To further verify whether DAGA has truly learned proactive avoidance behavior driven by changing hazard exposure trends, this paper constructs a simplified single-threat field avoidance scenario in addition to the main experiment. Unlike the previously mentioned structured multi-drone strong denial scenario, this experiment removes static obstacles and multi-drone interactions, retaining only a single drone, a single target point, and a circular threat field located at the center of the map. This setting maximizes the isolation of geometric obstacle avoidance and multi-drone collision avoidance factors, allowing the experiment to focus on analyzing the survival decision-making capabilities of different strategies when facing threats in invisible areas.

[0075] Specifically, the experimental map is a two-dimensional continuous space of 100m × 100m. For example... Figure 5 As shown, the drone's initial position is in the lower left corner, the target point is in the upper right corner, and a fixed circular threat field is set in the center of the map. Since the starting point, the center of the threat field, and the target point are approximately on the same diagonal, the geometric shortest path will directly pass through the high-risk area. Therefore, this scenario constructs a typical task-survival conflict: if the strategy excessively pursues the shortest path and rapid arrival, it will result in significant accumulation of danger exposure within the threat field, leading to a rapid decline in survival margin; if the strategy has the ability to perceive danger trends, it should actively deviate from the straight path, exchanging a longer trajectory for a higher survival margin.

[0076] This experiment conducted 10 independent tests on four methods: DAGA, PPO, SAC, and MATD3, and statistically analyzed the final survival margin, average number of steps, mission success rate, and threat zone exposure time ratio. The threat zone exposure time ratio was defined as the percentage of time steps within the threat zone radius of the drone's trajectory. The experiment recorded the drone's position, survival margin, whether it was within the threat zone, action commands, and rewards at each time step, and calculated the average statistical results for the 10 tests based on these data.

[0077] Table 5 Performance Comparison of Different Methods in Single Threat Field Bypass Scenarios

[0078] In terms of task completion, DAGA and MATD3 both achieved a 100% success rate, while PPO and SAC only achieved 30.0% and 0.0% respectively. This indicates that ordinary PPO and SAC struggle to reliably handle the conflict between target advancement and threat avoidance. Although SAC had the shortest average number of steps, its success rate was 0%, suggesting that its short runtime was mainly due to the strategy entering high-risk areas prematurely and failing, rather than higher task efficiency. While PPO was able to reach the target in some tests, its final survival margin was only 0.142, indicating that it still tends to directly traverse threat zones and lacks stable detour capabilities.

[0079] MATD3 also achieved a 100% success rate, but its average final survival margin was only 0.426. This indicates that while MATD3 could complete the mission, its strategy was more aggressive, resulting in significant survival margin loss when traversing threat areas. In contrast, DAGA significantly improved its final survival margin while maintaining mission success, suggesting that it could proactively choose detours with lower survival margin losses.

[0080] DAGA's average number of steps is 133.0, higher than MATD3's 115.0, indicating that it does not simply pursue the shortest path, but rather moderately increases the path length to reduce the risk of exhausting survivability margin when a threat field exists. This is consistent with the needs of UAV missions in strongly denied environments, where safer detour trajectories are often more beneficial to improving the overall survivability of the mission than the shortest path.

[0081] Although the difference between DAGA and MATD3 in terms of exposure time in the threat zone is limited, their final survival margins differ significantly. This indicates that DAGA's advantage lies not only in reducing the time spent in the threat zone, but also in avoiding deep penetration into the core areas of high-intensity threats, thereby reducing cumulative risk exposure.

[0082] This experiment further validates the role of the survival stress branch in DAGA. DAGA perceives threat changes through survival margin, current hazard exposure increments, and their trends, and dynamically adjusts the navigation representation using a residual gating mechanism, causing the UAV to deviate from the straight path and engage in detour behavior when hazard exposure increases. In contrast, although PPO, SAC, and MATD3 receive the same hazard exposure observations, they lack an explicit task-seeking graph decoupling mechanism, making it difficult to stably transform the survival margin depletion trend into proactive path replanning.

[0083] In summary, the single-threat field detour experiment demonstrates that when the shortest path conflicts with the safe path, DAGA can learn a detour strategy that better conforms to survival constraints. This result indicates that DAGA's performance improvement stems not only from the danger exposure penalty in the reward function, but also from the dynamic adjustment of policy representation by danger perception and residual gating mechanisms, validating its proactive survival avoidance capability in invisible threat fields at the micro-behavioral level.

[0084] The experimental results above demonstrate that DAGA's main advantage lies not in single-dimensional path efficiency or geometric obstacle avoidance performance, but in its comprehensive balance between task success rate and survival margin maintenance in strong denial environments. Compared to PPO, SAC, and MATD3, DAGA achieved the highest task success rate, the highest average survival margin across all samples, the highest average reward, and the lowest survival margin depletion rate in structured strong denial scenarios. This indicates that, under conditions where the regional threat field is not directly observable and agents can only perceive risks through local danger exposure feedback, simply using danger exposure-related variables as ordinary input features is insufficient to guarantee the formation of stable threat avoidance behavior. By decoupling task navigation, geometric danger avoidance, and threat exposure response representation, and employing a dual residual gating mechanism to dynamically fuse different features, DAGA enables the strategy to proactively enhance the intervention of the threat exposure response branch when the danger exposure trend increases, thereby avoiding blindly traversing high-risk areas under the inducement of high task rewards.

[0085] The failure type decomposition results show that DAGA does not achieve optimal performance across all physical collision metrics. In particular, compared to MATD3, DAGA's static and dynamic collision rates are not always the lowest. This phenomenon does not weaken the core conclusions of this paper; rather, it reveals the strengths and weaknesses of different algorithms. Static obstacles and neighboring UAVs represent explicit observable risks, which can be directly perceived through LiDAR and relative states of neighbors. Therefore, a strong continuous control baseline can learn a better reactive obstacle avoidance strategy through dense safety penalties. The geometric obstacle avoidance gating in DAGA is more suitable as a residual correction module for local physical safety. Its role is to maintain acceptable geometric safety boundaries during mission progression and threat avoidance, rather than as an independent obstacle avoider specifically pursuing the lowest collision rate. In contrast, threat exposure response gating addresses the implicit threat field, which cannot be directly observed. Its core role is to infer risk trends based on survivability margin, instantaneous attrition rate, and attrition change rate, and to dynamically correct the basic navigation intent with a survival-oriented approach. Therefore, DAGA's contribution should be understood as improving the mission-survivability tradeoff, rather than replacing traditional pure geometric obstacle avoidance algorithms.

[0086] The single-threat field detour experiment further explains the aforementioned performance differences from a behavioral perspective. In this scenario, while MATD3 also achieved a 100% target reach rate, its final survivability margin was significantly lower than DAGA, indicating that it tends to choose paths with more aggressive mission progression and higher survivability margin loss. DAGA, on the other hand, trades a longer average number of steps for a higher final survivability margin, demonstrating that its strategy can proactively choose a detour trajectory that better meets survivability constraints when faced with conflicts between the shortest and safest paths. This phenomenon suggests that DAGA learns not a simple target-tracking strategy, but a risk-avoidance strategy triggered by the trend of survivability margin loss. Since threat field damage increases non-linearly with the approach to the center, simply reducing the time spent in the threat zone is insufficient to explain the survivability margin advantage; more importantly, DAGA can avoid penetrating deep into the core area of ​​high-intensity threats, thereby reducing the cumulative survivability margin loss.

[0087] In summary, DAGA's experimental performance demonstrates a clear asymmetric functional division: the basic navigation flow maintains mission progression, the obstacle avoidance residual handles local explicit geometric risks, and the survival stress residual addresses latent threats revealed by survivability margin depletion feedback. While the dual-gating structure formally regulates both the obstacle avoidance expert and the survival expert, they are not entirely equivalent at the mission level. The obstacle avoidance gating primarily handles local safety corrections, while the survival gating is the core mechanism for improving survivability in strongly denied environments. This design enables the UAV to dynamically adjust its strategy bias based on hazard trends while maintaining mission objective information, thus achieving a more reasonable trade-off between mission completion rate, survivability margin maintenance, and risk avoidance.

[0088] This paper addresses the conflict between mission advancement, geometrical safety, and regional threat avoidance in collaborative navigation of UAV swarms under strong denial environments, proposing a Danger Awareness Dual Residual Gated Architecture (DAGA). This method does not interpret the UAV state as a health diagnostic variable in the reliability domain. Instead, it defines survivability margin based on local hazard exposure feedback and uses instantaneous hazard exposure increments and their trends to characterize the risk changes of the UAV in an unknown threat field. DAGA decouples the UAV's local observations into three feature streams: mission navigation, geometrical obstacle avoidance, and threat exposure response. It then performs residual adaptive fusion through geometrical obstacle avoidance gating and threat exposure response gating, enabling the agent to proactively adjust its navigation strategy based on local hazard exposure feedback even when the global threat field is unavailable.

[0089] Experimental results show that in structured strong denial scenarios, DAGA achieves higher task success rate, average survivability margin, and average reward compared to PPO, SAC, and MATD3, and significantly reduces the survivability margin depletion rate. In single-threat field detour experiments, DAGA can proactively avoid high-risk areas at the cost of moderately increasing path length, maintaining a higher final survivability margin, verifying the effective induction of proactive threat avoidance behavior by the hazard exposure trend perception mechanism. These results indicate that DAGA's core advantage is not simply pursuing the shortest path or the lowest collision rate, but rather achieving a more robust dynamic trade-off between task completion, geometric safety, and hazard exposure suppression.

[0090] This invention also provides a collaborative navigation system for unmanned aerial vehicle (UAV) swarms, such as... Figure 3 As shown, it includes the following modules: Data Acquisition Module: This module is used by the agents in each UAV to collect observation information at each decision step. The collected observation information includes mission navigation information, geometric obstacle avoidance information, and threat exposure response information. Feature processing module: used to preprocess the collected observation information to obtain preprocessed features, including task navigation features, geometric obstacle avoidance features, and threat exposure response features; Decoupling encoding module: used to decouple the preprocessed features according to their functions; Sub-network modules: These are used to input the decoupled features into the network model corresponding to the decoupling function for encoding and representation. The dual residual gating mechanism module includes an asymmetric residual fusion module for weighted fusion of geometric obstacle avoidance representation and threat exposure response representation in the encoded representation output by each sub-network module to obtain fused features; a geometric obstacle avoidance gating network for adjusting geometric obstacle avoidance gating weights according to static obstacle hazards and adjacent machine collision hazards; and a threat response gating network for adjusting threat exposure gating weights according to hazard exposure increment hazards. Temporal recurrent network: used to input the fused features into the temporal recurrent network and output the hidden state at the current time step; Actor-Critic Network: It is used to input the hidden state at the current moment into the Actor network, and output the mean and standard deviation of the linear acceleration action component and the angular velocity action component respectively. Based on the mean and standard deviation, it constructs the probability distribution of the continuous random control strategy, and samples the linear acceleration command and angular velocity command of each UAV from the probability distribution. Environment interaction module: Used to receive linear acceleration and angular velocity commands output by the Actor network, and apply the linear acceleration and angular velocity commands to the UAV dynamics model to update the UAV's position, velocity, heading, survivability margin, radar observation information and relative state information of neighboring UAVs, so as to obtain the environmental state and observation information after the action is executed; The reward module receives the pre-action state, action command, and post-action state output by the environmental interaction module, calculates the immediate reward based on the state transition process, provides feedback to the agent, and guides the agent's training. In this embodiment, the immediate reward is jointly determined by task navigation information, geometric obstacle avoidance information, and threat exposure response information. The first two types of information correspond to the task advancement reward and geometric obstacle avoidance reward in the reward function, while the state degradation information corresponds to the hazard exposure penalty and maneuver energy consumption penalty in the reward function.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for cooperative navigation of unmanned aerial vehicle (UAV) swarms, characterized in that, Includes the following steps: Step 1: Treat each drone in the drone swarm as an intelligent agent, and each drone intelligent agent collects observation information at each decision step; Step 2: Each UAV agent preprocesses the collected observation information to obtain preprocessed features; Step 3: Each UAV agent decouples the preprocessed features according to their functions, and inputs each decoupled feature into the network model corresponding to the decoupled function for encoding and representation. The parameters of each network model are independent. Step 4: Perform asymmetric residual fusion on the encoded representations output by each network model to obtain fused features; Step 5: Input the fused features into a temporal recurrent network and output the hidden state at the current time step; Step 6: Input the hidden state at the current moment into the Actor network, and output the mean and standard deviation of the linear acceleration action component and the angular velocity action component respectively. Construct the probability distribution of the continuous stochastic control strategy based on the mean and standard deviation, and sample the linear acceleration command and angular velocity command of each UAV from the probability distribution. Step 7: Each UAV updates its position, velocity, and heading angle according to the obtained linear acceleration and angular velocity commands through the UAV's dynamic equations, and updates the UAV's survivability margin after calculating the local hazard exposure intensity at the current position. If the UAV's survivability margin is lower than the threshold, the mission is considered to have failed. The survivability margin is a state quantity that characterizes the remaining safety margin after the accumulation of hazard exposure. Step 8: Determine whether communication is open at the current moment; if so, exchange status information with neighboring drones within the drone's perception range and update the memory of neighboring drone status information; otherwise, continue to calculate the current position and speed of the neighboring drones based on their historical positions and speeds as the neighboring drone status information. Step 9: Return to Step 1 until the task is completed.

2. The UAV swarm cooperative navigation method according to claim 1, characterized in that, The observation information includes: mission navigation information, geometric obstacle avoidance information, and threat exposure response information. The mission navigation information includes speed, heading angle, target distance, target relative bearing, survival flag, and the aircraft's state information related to the UAV's own state at the previous moment. The geometric obstacle avoidance information includes UAV onboard radar observation information and neighboring aircraft relative state information with information freshness. The information freshness refers to the ratio of the time length from when the current UAV acquires the latest state of the neighboring aircraft to the current time to the preset longest tolerance interval. The neighboring aircraft relative state information includes the relative speed and relative position of the neighboring aircraft. The threat exposure response information includes survival margin, current risk exposure increment, and risk exposure increment at the previous moment.

3. The UAV swarm cooperative navigation method according to claim 2, characterized in that, The preprocessed features obtained by each UAV from the collected observation information refer to: The speed, heading angle, target distance and previous action in the mission navigation information are normalized, and the target relative bearing is converted into sine and cosine values ​​to obtain mission navigation features. The 16-dimensional ranging information from the UAV's airborne radar observation information is mapped using trigonometric functions according to the azimuth angles of each radar beam in the aircraft's coordinate system. Each dimension of ranging information is projected onto the horizontal and vertical components of the aircraft's coordinate system to obtain a 32-dimensional two-dimensional Cartesian feature vector. The relative state information of neighboring aircraft with information freshness is input into a relational encoding network, and invalid neighboring aircraft are masked by dynamic masking. After feature aggregation with masking, a 64-dimensional neighboring aircraft relational feature vector is obtained. The 32-dimensional two-dimensional Cartesian feature vector is concatenated with the 64-dimensional neighboring aircraft relational feature vector, and the concatenation result is used as a geometric obstacle avoidance feature. The trend of hazard exposure change is calculated based on the current hazard exposure increment and the hazard exposure increment at the previous moment in the threat exposure response information. The survival margin, the current hazard exposure increment, the hazard exposure increment at the previous moment, and the hazard exposure change trend are normalized and used as the threat exposure response features.

4. The UAV swarm cooperative navigation method according to claim 3, characterized in that, Inputting the decoupled features into the network model corresponding to the decoupling function means: Three functionally independent and parameter-isolated network models are set up. The three network models are a task navigation sub-network for encoding and representing task navigation features, a geometric obstacle avoidance sub-network for encoding and representing geometric obstacle avoidance features, and a threat exposure response sub-network for encoding and representing threat exposure response features. The preprocessed task navigation features, geometric obstacle avoidance features, and threat exposure response features are input into the corresponding network models to obtain task navigation representations, geometric obstacle avoidance representations, and threat exposure response representations.

5. The UAV swarm cooperative navigation method according to claim 4, characterized in that, The method for obtaining fused features by asymmetric residual fusion of the encoded representations output by each network model is as follows: ； Indicates fusion characteristics, This represents the task navigation representation. This represents the geometric obstacle avoidance representation. This indicates the threat exposure response characterization. Represents the geometric obstacle avoidance gating weights. This indicates the threat exposure gating weight; when the geometric obstacle avoidance hazard increases, Increase the geometric obstacle avoidance modifier. Enhancement occurs when the increment of hazardous exposure increases. Increase the number of threat exposure response modifications. Enhance.

6. The UAV swarm cooperative navigation method according to claim 5, characterized in that, The Actor network employs an Actor-Critic network architecture for policy optimization during the training phase. The Critic network is used to estimate the value of the current state and calculate the advantage function based on the value estimation results to assist the Actor network in policy updates. During the training phase, the Critic network receives not only the current hidden state output by the temporal recurrent network, but also the geometric obstacle avoidance representation output by the geometric obstacle avoidance subnetwork. Threat exposure response characterization output by the threat exposure response subnetwork After the Actor-Critic network is trained, only the Actor network is used during the execution phase, and the Critic network does not participate in the generation of UAV actions. During the execution phase, the Actor network only receives the current hidden state obtained by processing the fused features through a time-series cyclic network, and generates linear acceleration and angular velocity commands for each UAV based on the current hidden state.

7. A method for cooperative navigation of unmanned aerial vehicle (UAV) swarms according to claim 6, characterized in that, During training, the total reward function of the unmanned aerial vehicle agent includes four parts: mission navigation advancement reward, geometric obstacle avoidance penalty, hazard exposure response penalty, and maneuver energy consumption penalty. ； in, Indicates the first The drone in Rewards for each decision-making step This indicates the weight of the task navigation advancement reward. Indicates the first The drone in Rewards for progressing through task navigation at each decision-making step. Indicates the first The drone in Geometric obstacle avoidance reward for each decision step Indicates the first The drone in The current risk exposure increment at each decision step Indicates the first The drone in The energy consumption of each decision-making step. Indicates the geometric obstacle avoidance reward weight. Indicates the penalty weight for dangerous exposure. This indicates the weight of the penalty for motor energy consumption.

8. The UAV swarm cooperative navigation method according to claim 7, characterized in that, The method for calculating the survival margin is as follows: Defined as the first The drone in The survival margin of each decision-making step. This represents the incremental risk exposure of the i-th drone at the i-th decision step; when Below the preset threshold In such cases, the drone will fail due to the exhaustion of its survival margin caused by continuous exposure to high-risk areas.

9. A method for cooperative navigation of unmanned aerial vehicle (UAV) swarms according to claim 8, characterized in that, The method for calculating the trend of change in the hazardous exposure is as follows: ； Indicates the trend of changes in hazardous exposure; Indicates the current increase in risk exposure, This indicates an increase in historical risk exposure. when When the drone's current direction of motion is increasing its risk exposure, it prompts the strategy to proactively explore and change course; conversely, if... This indicates that the drone is moving away from a high-risk area.

10. A collaborative navigation system for unmanned aerial vehicle (UAV) swarms, characterized in that: Includes the following modules: Data Acquisition Module: This module is used by the agents in each UAV to collect observation information at each decision step. The collected observation information includes mission navigation information, geometric obstacle avoidance information, and threat exposure response information. Feature processing module: used to preprocess the collected observation information to obtain preprocessed features, including task navigation features, geometric obstacle avoidance features, and threat exposure response features; Decoupling encoding module: used to decouple the preprocessed features according to their functions; Sub-network modules: These are used to input the decoupled features into the network model corresponding to the decoupling function for encoding and representation. The dual residual gating mechanism module includes an asymmetric residual fusion module for weighted fusion of geometric obstacle avoidance representation and threat exposure response representation in the encoded representation output by each sub-network module to obtain fused features; a geometric obstacle avoidance gating network for adjusting geometric obstacle avoidance gating weights according to static obstacle hazards and adjacent machine collision hazards; and a threat response gating network for adjusting threat exposure gating weights according to hazard exposure increment hazards. Temporal recurrent network: used to input the fused features into the temporal recurrent network and output the hidden state at the current time step; Actor-Critic Network: It is used to input the hidden state at the current moment into the Actor network, and output the mean and standard deviation of the linear acceleration action component and the angular velocity action component respectively. Based on the mean and standard deviation, it constructs the probability distribution of the continuous random control strategy, and samples the linear acceleration command and angular velocity command of each UAV from the probability distribution. Environment interaction module: Used to receive linear acceleration and angular velocity commands output by the Actor network, and apply the linear acceleration and angular velocity commands to the UAV dynamics model to update the UAV's position, velocity, heading, survivability margin, radar observation information and relative state information of neighboring UAVs, so as to obtain the environmental state and observation information after the action is executed; Reward module: It is used to receive the state before action execution, action instruction and state after action execution output by the environment interaction module, calculate the instant reward according to the state transition process, provide feedback to the agent and guide the agent to train.

Images