An unmanned cluster brain-like autonomous navigation method based on swarm intelligence

Abstract

An unmanned cluster brain-like autonomous navigation method based on swarm intelligence. Based on the topological relationship in the cognitive map and the cognitive cost constraint, a brain-like driven global path generation mechanism is constructed to realize the global feasible path search for task-oriented, energy-constrained and environment-constrained. For dynamic obstacles, perception noise and local disturbance, a local environment state representation and online obstacle avoidance control mechanism is established to form a local path planning method considering safety and real-time performance. The global path guidance and local obstacle avoidance output are fused to obtain the candidate navigation behavior of the unmanned cluster at the current time. The historical traffic experience, risk area information and local re-planning demand are dynamically coupled to complete the navigation instruction arbitration and path correction, and the cluster autonomous navigation decision result satisfying the spatiotemporal consistency constraint is output. The application can improve the path planning quality, local obstacle avoidance robustness and sustained navigation ability of the unmanned cluster in complex unknown environment.

Description

Technical Field

[0001] This invention belongs to the field of autonomous navigation and swarm intelligence collaborative control technology for unmanned systems. Specifically, it relates to a brain-like autonomous navigation method for unmanned swarms based on swarm intelligence, which is applicable to cognitive map utilization, path planning, local obstacle avoidance, memory linkage decision-making and collaborative navigation control of unmanned swarms in complex dynamic environments. Background Technology

[0002] With the development of autonomous and swarm technologies for unmanned systems, unmanned swarms have significant application value in tasks such as environmental exploration, emergency rescue, regional inspection, intelligent transportation, and collaborative reconnaissance. Existing unmanned navigation methods mostly employ a hierarchical framework combining perception mapping, path planning, and motion control, achieving certain results in static or structured environments. However, in complex dynamic environments, unmanned swarms face problems such as incomplete local perception, dynamic obstacle interference, frequent path conflicts, limited communication, and difficulties in multi-agent collaboration. Traditional path planning methods based on grid maps, geometric maps, or single cost functions struggle to simultaneously express environmental topological relationships, spatial metric relationships, and historical travel experience. Local obstacle avoidance methods often rely solely on current observations, lacking predictions of future states of dynamic obstacles and regression constraints on the global path, easily leading to path oscillations, local deadlocks, and frequent replanning. Therefore, there is an urgent need for an autonomous navigation method for unmanned swarms that integrates cognitive maps, memory mechanisms, dynamic obstacle avoidance, and group collaborative decision-making. Summary of the Invention

[0003] To address the shortcomings of existing unmanned swarm navigation methods, such as insufficient environmental representation capabilities, loose coupling between global path planning and local obstacle avoidance, weak dynamic obstacle prediction capabilities, difficulty in effectively resolving group path conflicts, and insufficient utilization of historical experience, this paper proposes a brain-like autonomous navigation method for unmanned swarms based on swarm intelligence. This method uses the topological cognitive map and cooperative localization results already constructed by the unmanned swarm as a foundation, unifying environmental nodes, connectivity relationships, spatial distances, and task constraints into a global navigation space. Based on this, global path construction is performed using brain-like impulse propagation, reward modulation, and path memory mechanisms, and conflict-free path allocation is achieved through group path intent synchronization and conflict suppression. Simultaneously, local path planning is completed by combining short-term memory, dynamic obstacle prediction, and a dynamic neural field model. Furthermore, the global path and local obstacle avoidance paths are fused through long-term and short-term memory linkage calibration. Finally, the final navigation decision output is completed using a group consensus decision-making and execution feedback correction mechanism, thereby improving the autonomous navigation capability, cooperative decision-making capability, and task execution stability of the unmanned swarm in complex dynamic environments.

[0004] A brain-like autonomous navigation method for unmanned swarms based on swarm intelligence, characterized by the following steps: Step 1: Using the topological cognitive map and collaborative localization results already constructed by the unmanned swarm as input, extract the node connectivity, spatial metric relationships, task target location and environmental constraint information in the environment to construct a global navigation space for autonomous navigation of the unmanned swarm. Step 2: In the global navigation space constructed in Step 1, a global path construction method driven by a cognitive map is used to search for paths, and the environmental topology relationship and cognitive cost information are fused to generate a globally feasible path that meets the requirements of the unmanned swarm mission. Step 3: During the navigation process of the unmanned swarm along the global path, acquire dynamic obstacle information, perceived noise information and local disturbance information in the local environment in real time, construct a local environment state representation model, and perform local path planning based on the model; Step 4: Combine the global path construction results obtained in Step 2 with the local path planning results obtained in Step 3 to form a candidate navigation decision set for the unmanned swarm at the current navigation moment; Step 5: Perform memory linkage calibration on the candidate navigation decision set formed in Step 4, and jointly process historical travel experience, risk area information and replanning strategy to complete the final navigation decision arbitration and output the autonomous navigation result of the unmanned swarm.

[0005] Compared with the prior art, the technical solution of this invention has the following advantages: This invention constructs a global navigation space through a topological cognitive map, which can uniformly express the connectivity relationships of environmental nodes, spatial metric relationships, and task constraint information, thereby improving spatial cognitive ability in complex environments. Global path construction is achieved through brain-like impulse propagation, reward modulation, and path memory mechanisms, enabling comprehensive optimization among path length, energy consumption, threat cost, and map matching degree, thus improving the reliability of the global path. By fusing dynamic obstacle prediction information with global target guidance information through short-term memory and a dynamic neural field model, the real-time performance and stability of local obstacle avoidance are enhanced, reducing path oscillations and local deadlocks. A long-term memory and short-term memory linkage calibration mechanism enables dynamic fusion of the global path and local obstacle avoidance paths, allowing the agent to promptly return to the global path after obstacle avoidance. A group consensus decision-making and execution feedback correction mechanism improves the path coordination and navigation robustness of unmanned swarms, reducing the risk of path conflicts and collisions among multiple agents. This invention effectively improves the autonomous navigation efficiency, safety, and collaborative stability of unmanned swarms in complex dynamic environments. Attached Figure Description

[0006] Figure 1 Overall flowchart of the unmanned swarm brain-like autonomous navigation method; Figure 2 Comparison chart of simulation results based on global path construction driven by cognitive map; Figure 3 Comparison of simulation trajectories between local dynamic obstacle avoidance and global path regression; Figure 4 Simulation results of memory linkage calibration and experience-enhanced memory intensity distribution. Detailed Implementation

[0007] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments: A brain-like autonomous navigation method for unmanned swarms based on swarm intelligence, the process of which is as follows: Figure 1 As shown, the specific steps include the following: Step 1: Using the topological cognitive map and collaborative localization results already constructed by the unmanned swarm as input, extract the node connectivity, spatial metric relationships, task target location and environmental constraint information in the environment to construct a global navigation space for autonomous navigation of the unmanned swarm.

[0008] The topological cognitive map is transformed into a neural topological network to simulate the way the hippocampus-entorhinal cortex represents spatial location, topological connectivity, and path memory. The neural topological network is represented as follows:

[0009] in, This represents the neural topology network corresponding to the cognitive map. This represents a set of neurons, each corresponding to a topological node in the cognitive map. Represents the set of synaptic connections. Represents neurons arrive The connection, This represents the set of synaptic weights, used to describe the spatial distance, topological similarity, and path reachability between nodes; Furthermore, in order to simultaneously consider distance metric and topological similarity in neural topological networks, a synaptic weight function is constructed, specifically expressed as:

[0010] in, Represents neurons With neurons Synaptic connection weights between Let be the distance between nodes i and j. Distance threshold For node topological feature similarity , These are the weighting coefficients, representing the synaptic connection weights. The larger the value, the lower the path cost between nodes, and the easier it is for the pulse signal to propagate.

[0011] Step 2: In the global navigation space constructed in Step 1, a global path construction method driven by cognitive map is used to search for paths, and the environmental topology relationship and cognitive cost information are fused to generate a globally feasible path that meets the requirements of unmanned swarm mission.

[0012] In the constructed global navigation space, the neuron corresponding to the target node is used as the starting point of the pulse signal, and the pulse wave propagates backward along the neural topology network. The formula for updating the neuron membrane potential is as follows:

[0013] in, The membrane time constant of the SNN neuron. Represents neurons The membrane potential at time t For neurons The reverse neighborhood set of neurons, For neurons At any moment The generated pulse signal This is the delay in signal propagation between neurons. For neurons The received dopamine reward factor, transmitted from neighboring neurons, is updated using the following formula:

[0014] When the originating neuron The membrane potential reaches the threshold. When a pulse signal is generated, propagation stops, and the pulse propagation path from the endpoint to the starting point is traced back to obtain the path with optimal synaptic weights. This is the globally optimal path for a single entity. This indicates the neuron corresponding to the target node. This indicates the intermediate path nodes selected during the backtracking process; Furthermore, to meet the comprehensive constraints of energy consumption, threat cost, and map matching degree in complex tasks, multiple candidate paths generated by the pulse wave are written into the path memory unit, as shown in the following expression:

[0015] in, Let's define the topology of the k-th candidate path. For path energy consumption, For the cost of path threat, The degree of matching between the path and the cognitive map; Based on this, a multi-objective comprehensive evaluation is performed on each candidate path, and the evaluation function is as follows:

[0016] in, This represents the comprehensive evaluation value of the k-th candidate path. Indicates the path length cost. This indicates the cost of energy consumption. Indicates the cost of threats. Indicates the revenue from map matching. , , , These are weighting coefficients, dynamically adjusted by the agent's task priority; Furthermore, a reward prediction error mechanism is introduced to dynamically update the path evaluation weights, the expression of which is:

[0017] in, This represents the adjustment amount of the i-th weight parameter. The learning rate for the prefrontal cortex decision model. To reward prediction errors, For the actual path cost, To predict path cost; After evaluating individual candidate paths, a path intent confidence score is generated to describe the reliability of each agent's path selection. Its expression is:

[0018] in, , , The confidence level weighting coefficient is... For path matching degree, For path energy consumption, For the cost of path threat, , These are the energy and threat cost thresholds, respectively. When this happens, the intent for that path is highlighted and considered a valid intent; To achieve collaborative and conflict-free global path allocation in an unmanned swarm, the path intentions of each agent are mapped to the phase evolution process of a coupled oscillator, with the following dynamic equation:

[0019] Where N is the number of swarm agents. This indicates the natural frequency of the corresponding oscillator. Let be the path intent confidence of the j-th intelligent unit. Based on the fundamental coupling strength, for agent pairs (i,j) with severe conflicts, a phase repulsion factor is introduced to correct the coupling term, and its expression is:

[0020] in, The phase repulsion factor between agents i and j is used to characterize the intensity of path conflict, and k represents the index number.

[0021] Step 3: During the navigation process of the unmanned swarm along the global path, acquire dynamic obstacle information, perceived noise information and local disturbance information in the local environment in real time, construct a local environment state representation model, and perform local path planning based on the model.

[0022] A local environment state representation model is constructed and local path planning is performed. The velocity, equivalent size, and relative position of dynamic obstacles in the local environment are acquired in real time, and threat connection weights between sensor feature outputs and short-term memory units are constructed, with the expression being:

[0023] in, Indicates the obstacle threat connection weight. The speed of the obstacle's movement. The equivalent radius of the obstacle. , These are the maximum obstacle speed and the maximum equivalent radius that the agent can perceive. The larger the connection weight, the higher the threat level of the corresponding obstacle, and the more likely it is to be written into the local short-term memory. Furthermore, to stably store and represent the local dynamic obstacle states, a short-term memory model in the form of a continuous attractor neural network is constructed, with the connection weights between neurons represented as follows:

[0024] in, Represents neurons With neurons Connection weights between them Based on the fundamental connection strength, , These are the obstacle state vectors for the corresponding neurons. The width of the Gaussian kernel; Simultaneously, a short-term synaptic plasticity mechanism is introduced to dynamically update synaptic availability and resource consumption rates, specifically as follows:

[0025]

[0026] in, Synapse availability, This refers to the synaptic resource consumption rate. Based on availability, , These are the fast and slow time constants, respectively. The synaptic input current is regulated by the short-time synaptic plasticity mechanism (STP). To predict the future state of dynamic obstacles, a linear Kalman filter is used for trajectory estimation based on historical obstacle information in short-term memory. The state equation and observation equation are as follows:

[0027]

[0028] in, This represents the obstacle state vector at time k. Here is the state transition matrix. For the observation matrix, For process noise, To observe the noise, Represents the observation vector at time k; Furthermore, to unify global path guidance information and local obstacle avoidance information into a local direction selection space, a dynamic neural field model is constructed, whose dynamic equations are expressed as follows:

[0029] in, The time constant of the neural field, Indicates direction The activation value at time t, The weights represent the lateral connections between neurons. For the attractor input current, For the repulsion input current, For the external bias current, the lateral connection weight adopts a local excitation-long-range inhibition form, specifically expressed as follows:

[0030] in, , These are the excitation and inhibition weights, respectively. , These represent the Gaussian kernel widths for excitation and inhibition, respectively; Furthermore, the direction of the next signpost in the global path is mapped to the attractor input current, expressed as:

[0031] in, For the maximum attractor current, This represents the target orientation angle corresponding to the global LTM path marker. The scope of the attractor is defined by the fact that the attractor's existence ensures that the agent can return to the global path after obstacle avoidance. Simultaneously, the predicted dynamic obstacle direction is mapped to the repulsion input current, expressed as:

[0032] in, For the maximum repulsive current, For obstacle threat weight, The direction angle corresponding to the obstacle. The range of action of the repulsion unit; After the dynamic neural field converges, the direction of maximum activation is taken as the local optimal motion direction, and local control variables are generated. The linear velocity and angular velocity are expressed as follows:

[0033]

[0034] in, Indicates the desired linear velocity. The maximum linear velocity of the agent. This represents the neural field activation value at the optimal direction. This represents the maximum activation level of the neural field. Indicates the desired angular velocity. This is the angular velocity gain coefficient. This represents the current direction angle of the agent's movement. Furthermore, to maintain formation coordination during local obstacle avoidance in the unmanned swarm, local connections are established only with the k nearest neighboring agents, and the connection weights are represented as follows:

[0035] in, This represents the relative distance between agent i and agent j. The effective range for maintaining formation; Based on this, the local cooperative force is calculated using an artificial attractor network, and the expression is:

[0036] in, Let be the set of neighborhood nodes of agent i. , These are the attraction and repulsion gain coefficients, respectively. , These are the attraction potential function and the repulsion potential function, respectively. The attraction potential function is expressed as:

[0037] in, , Let i be the position vector of agent i and j. Let the relative position vectors of the two agents be given by the desired formation, and the repulsive potential function be expressed as:

[0038] in, This refers to the area of ​​exclusion.

[0039] Step 4: Integrate the global path construction results obtained in Step 2 with the local path planning results obtained in Step 3 to form a candidate navigation decision set for the unmanned swarm at the current navigation moment.

[0040] The global path construction results are fused with the local path planning results to form a candidate navigation decision set. To quantify the deviation and conflict degree between the global and local paths, a global-local path conflict evaluation function is constructed, the expression of which is:

[0041] in, , , For conflict assessment weighting coefficients, This represents the deviation distance between the local obstacle avoidance path and the global LTM path. The maximum allowable deviation distance, Prioritize short-term obstacle avoidance goals. Prioritize long-term navigation goals. A larger value indicates a more severe conflict, requiring priority for conflict resolution. Furthermore, a dopamine modulation signal is introduced to dynamically adjust the fusion weights of the global and local path results, and its expression is as follows:

[0042] in, Based on the intensity of dopamine signal, This represents the adjustment increment obtained from the execution feedback. Let represent the dopamine regulation signal at time t. Based on the dopamine regulation signal, the long-term memory decision weight corresponding to the global path is calculated, expressed as:

[0043] in, Represents the global path decision weights. The dopamine signal threshold; Furthermore, the short-term memory decision weights corresponding to the local paths are represented as follows:

[0044] in, Indicates the weight of the local path decision; After determining the global path weights and local path weights, the local paths are calibrated and corrected under global constraints to obtain candidate navigation paths, represented as follows:

[0045] in, This indicates the candidate navigation path after fusion calibration. This is the original STM obstacle avoidance path. For global LTM paths, This represents the path calibration coefficient.

[0046] Step 5: Perform memory linkage calibration on the candidate navigation decision set formed in Step 4, and jointly process historical travel experience, risk area information and replanning strategy to complete the final navigation decision arbitration and output the autonomous navigation result of the unmanned swarm.

[0047] The process involves memory-linked calibration and final arbitration output for the candidate navigation decision set, specifically including: The fused candidate navigation paths are converted into local decision vectors, represented as follows:

[0048] in, This represents the local navigation decision vector of the i-th agent. , To control the amount of obstacle avoidance, The direction of the target movement; Furthermore, the local decisions of neighboring agents are broadcast through a cluster communication mechanism, and a weighted average method is used to generate a group consensus decision, expressed as:

[0049] in, This represents the consensus decision obtained by agent i. Let i be the neighborhood set of agent i. Indicates the neighborhood connection weight. Assign weight to its own decision-making; To incorporate historical experience, risk area information, and current consensus results into the arbitration process, a reference decision is introduced to adaptively modify the group consensus decision. The expression is as follows:

[0050] in, This represents the final navigation decision of agent i. This refers to reference decisions generated from historical navigation experience, long-term memory paths, or prior knowledge of risk areas. Indicates the adaptive adjustment coefficient; Finally, to evaluate the effectiveness of the final navigation decision and to provide feedback and correction for subsequent navigation arbitration processes, an evaluation index for the effectiveness of the decision is constructed, with the following expression:

[0051] in, This represents the performance evaluation metric for agent i. Indicates positional deviation. Indicates heading deviation. Indicates the neighborhood cooperative distance error. , , To evaluate the weights, , , These represent the maximum permissible values ​​for the corresponding errors.

[0052] Figure 2 The simulation results of global path construction are shown. Figure 3 The simulation results of local dynamic obstacle avoidance and global path regression are shown. Figure 4 The simulation results of memory linkage calibration are shown.

[0053] Combination Figure 2 It can be seen that, in the comparison of global path construction, the method of this invention, while satisfying topological connectivity constraints, can comprehensively balance path length, energy consumption, threat cost, and map matching degree. Compared with methods that only pursue the geometric shortest path, although the method of this invention may lead to a slight increase in path length, it can effectively improve the safety margin and reduce the cumulative threat cost, thereby improving the executability and robustness of the global path. Figure 3 It can be seen that, in the presence of dynamic obstacle interference, this invention, through the linkage of short-term memory prediction and dynamic neural field local planning, enables the agent to return to the global reference path after completing local obstacle avoidance. The trajectory continuity is good, and a more reasonable safe distance is maintained from obstacles, demonstrating good dynamic obstacle avoidance stability and path return capability. Combined with... Figure 4 It is evident that the memory-linked calibration mechanism can collaboratively correct historical travel experience, risk area information, and current candidate decisions, reducing deadlock and collision risks, and minimizing invalid replanning. Simultaneously, the experience-enhanced memory strength gradually forms a stable distribution in high-frequency travel areas, providing effective prior knowledge for repetitive tasks and further improving the consistency and efficiency of subsequent navigation decisions.

[0054] It should be noted that the above description of the embodiments is only for the purpose of helping to understand the method and core idea of ​​this application. For those skilled in the art, several improvements and modifications can be made to this application without departing from the principle of this application, and these improvements and modifications are also within the protection scope of the claims of this application.

Claims

1. A brain-like autonomous navigation method for unmanned swarms based on swarm intelligence, characterized in that, Includes the following steps: Step 1: Using the topological cognitive map and collaborative localization results already constructed by the unmanned swarm as input, extract the node connectivity, spatial metric relationships, task target location and environmental constraint information in the environment to construct a global navigation space for autonomous navigation of the unmanned swarm. Step 2: In the global navigation space constructed in Step 1, a global path construction method driven by a cognitive map is used to search for paths, and the environmental topology relationship and cognitive cost information are fused to generate a globally feasible path that meets the requirements of the unmanned swarm mission. Step 3: During the navigation process of the unmanned swarm along the global path, acquire dynamic obstacle information, perceived noise information and local disturbance information in the local environment in real time, construct a local environment state representation model, and perform local path planning based on the model; Step 4: Combine the global path construction results obtained in Step 2 with the local path planning results obtained in Step 3 to form a candidate navigation decision set for the unmanned swarm at the current navigation moment; Step 5: Perform memory linkage calibration on the candidate navigation decision set formed in Step 4, and jointly process historical travel experience, risk area information and replanning strategy to complete the final navigation decision arbitration and output the autonomous navigation result of the unmanned swarm.

2. The unmanned cluster brain-like autonomous navigation method based on swarm intelligence according to claim 1, characterized in that, In step 1, a global path is constructed based on the topological cognitive map, which is then transformed into a neural topological network to simulate the hippocampus-entorhinal cortex's representation of spatial location, topological connectivity, and path memory. The neural topological network is represented as follows: ； in, This represents the neural topology network corresponding to the cognitive map. This represents a set of neurons, each corresponding to a topological node in the cognitive map. Represents the set of synaptic connections. Represents neurons arrive The connection, This represents the set of synaptic weights, used to describe the spatial distance, topological similarity, and path reachability between nodes; To simultaneously consider distance metric and topological similarity in neural topological networks, a synaptic weight function is constructed, specifically expressed as: ； in, Represents neurons With neurons Synaptic connection weights between Let be the distance between nodes i and j. Distance threshold For node topological feature similarity , These are the weighting coefficients, representing the synaptic connection weights. The larger the value, the lower the path cost between nodes, and the easier it is for the pulse signal to propagate.

3. The unmanned cluster brain-like autonomous navigation method based on swarm intelligence according to claim 2, characterized in that, Step 2 utilizes a cognitive map-driven global path construction method for path searching, specifically including: In the global navigation space constructed in step 1, the neuron corresponding to the target node is used as the starting point of the pulse signal, and the pulse wave propagates backward along the neural topology network. The neuron membrane potential update formula is as follows: ； in, The membrane time constant of the SNN neuron. Represents neurons The membrane potential at time t For neurons The reverse neighborhood set of neurons, For neurons At any moment The generated pulse signal This is the delay in signal propagation between neurons. For neurons The received dopamine reward factor, transmitted from neighboring neurons, is updated using the following formula: ； When the originating neuron The membrane potential reaches the threshold. When a pulse signal is generated, propagation stops, and the pulse propagation path from the endpoint to the starting point is traced back to obtain the path with optimal synaptic weights. This is the globally optimal path for a single entity. This indicates the neuron corresponding to the target node. This indicates the intermediate path nodes selected during the backtracking process; To meet the comprehensive constraints of energy consumption, threat cost, and map matching degree in complex tasks, multiple candidate paths generated by the pulse wave are written into the path memory unit, as shown in the following expression: ； in, Let's define the topology of the k-th candidate path. For path energy consumption, For the cost of path threat, The degree of matching between the path and the cognitive map; Based on this, a multi-objective comprehensive evaluation is performed on each candidate path, and the evaluation function is as follows: ； in, This represents the comprehensive evaluation value of the k-th candidate path. Indicates the path length cost. This indicates the cost of energy consumption. Indicates the cost of threats. Indicates the revenue from map matching. , , , These are weighting coefficients, dynamically adjusted by the agent's task priority; A reward prediction error mechanism is introduced to dynamically update the path evaluation weights. Its expression is: ； in, This represents the adjustment amount of the i-th weight parameter. The learning rate for the prefrontal cortex decision model. To reward prediction errors, For the actual path cost, To predict path cost; After evaluating individual candidate paths, a path intent confidence score is generated to describe the reliability of each agent's path selection. Its expression is: ； in, , , The confidence level weighting coefficient is... For path matching degree, For path energy consumption, For the cost of path threat, , These are the energy and threat cost thresholds, respectively. When this happens, the intent for that path is highlighted and considered a valid intent; To achieve collaborative and conflict-free global path allocation in an unmanned swarm, the path intentions of each agent are mapped to the phase evolution process of a coupled oscillator, with the following dynamic equation: ； Where N is the number of swarm agents. This indicates the natural frequency of the corresponding oscillator. Let be the path intent confidence of the j-th intelligent unit. Based on the fundamental coupling strength, for agent pairs (i,j) with severe conflicts, a phase repulsion factor is introduced to correct the coupling term, and its expression is: ； in, The phase repulsion factor between agents i and j is used to characterize the intensity of path conflict, and k represents the index number.

4. The unmanned cluster brain-like autonomous navigation method based on swarm intelligence according to claim 3, characterized in that, Step 3 involves constructing a local environment state representation model and performing local path planning based on this model, specifically including: The system acquires the velocity, equivalent size, and relative position of dynamic obstacles in the local environment in real time, and constructs the threat connection weights between the sensor feature outputs and the short-term memory unit. The expression for this weighting is: ； in, Indicates the obstacle threat connection weight. The speed of the obstacle's movement. The equivalent radius of the obstacle. , These are the maximum obstacle speed and the maximum equivalent radius that the agent can perceive. The larger the connection weight, the higher the threat level of the corresponding obstacle, and the more likely it is to be written into the local short-term memory. To stably store and represent the local dynamic obstacle state, a short-term memory model in the form of a continuous attractor neural network is constructed, and the connection weights between its neurons are represented as follows: ； in, Represents neurons With neurons Connection weights between them Based on the fundamental connection strength, , These are the obstacle state vectors for the corresponding neurons. The width of the Gaussian kernel; Simultaneously, a short-term synaptic plasticity mechanism is introduced to dynamically update synaptic availability and resource consumption rates, specifically as follows: ； ； in, Synapse availability, This refers to the synaptic resource consumption rate. Based on availability, , These are the fast and slow time constants, respectively. The synaptic input current is regulated by the short-time synaptic plasticity mechanism (STP). To predict the future state of dynamic obstacles, a linear Kalman filter is used for trajectory estimation based on historical obstacle information in short-term memory. The state equation and observation equation are as follows: ； ； in, This represents the obstacle state vector at time k. Here is the state transition matrix. For the observation matrix, For process noise, To observe the noise, Represents the observation vector at time k; To unify global path guidance information and local obstacle avoidance information into a local direction selection space, a dynamic neural field model is constructed, whose dynamic equations are expressed as follows: ； in, The time constant of the neural field, Indicates direction The activation value at time t, The weights represent the lateral connections between neurons. For the attractor input current, For the repulsion input current, For the external bias current, the lateral connection weight adopts a local excitation-long-range inhibition form, specifically expressed as follows: ； in, , These are the excitation and inhibition weights, respectively. , These represent the Gaussian kernel widths for excitation and inhibition, respectively; Mapping the direction of the next signpost in the global path to the attractor input current is expressed as: ； in, For the maximum attractor current, This represents the target orientation angle corresponding to the global LTM path marker. The scope of the attractor is defined by the fact that the attractor's existence ensures that the agent can return to the global path after obstacle avoidance. Simultaneously, the predicted dynamic obstacle direction is mapped to the repulsion input current, expressed as: ； in, For the maximum repulsive current, For obstacle threat weight, The direction angle corresponding to the obstacle. The range of action of the repulsion unit; After the dynamic neural field converges, the direction of maximum activation is taken as the local optimal motion direction, and local control variables are generated. The linear velocity and angular velocity are expressed as follows: ； ； in, Indicates the desired linear velocity. The maximum linear velocity of the agent. This represents the neural field activation value at the optimal direction. This represents the maximum activation level of the neural field. Indicates the desired angular velocity. This is the angular velocity gain coefficient. This represents the current direction angle of the agent's movement. To maintain formation coordination during local obstacle avoidance in the unmanned swarm, local connections are established only with the k nearest neighboring agents, and the connection weights are represented as follows: ； in, This represents the relative distance between agent i and agent j. The effective range for maintaining formation; Based on this, the local cooperative force is calculated using an artificial attractor network, and the expression is: ； in, Let be the set of neighborhood nodes of agent i. , These are the attraction and repulsion gain coefficients, respectively. , These are the attraction potential function and the repulsion potential function, respectively. The attraction potential function is expressed as: ； in, , Let i be the position vector of agent i and j. Let the relative position vectors of the two agents be given by the desired formation, and the repulsive potential function be expressed as: ； in, This refers to the area of ​​exclusion.

5. The unmanned cluster brain-like autonomous navigation method based on swarm intelligence according to claim 4, characterized in that, Step 4 integrates the global path construction results with the local path planning results to form a candidate navigation decision set, specifically including: To quantify the deviation and conflict degree between global and local paths, a global-local path conflict evaluation function is constructed, the expression of which is: ； in, , , For conflict assessment weighting coefficients, This represents the deviation distance between the local obstacle avoidance path and the global LTM path. The maximum allowable deviation distance, Prioritize short-term obstacle avoidance goals. Prioritize long-term navigation goals. A larger value indicates a more severe conflict, requiring priority for conflict resolution. A dopamine modulation signal is introduced to dynamically adjust the fusion weights of the global and local path results, and its expression is as follows: ； in, Based on the intensity of dopamine signal, This represents the adjustment increment obtained from the execution feedback. Let represent the dopamine regulation signal at time t. Based on the dopamine regulation signal, the long-term memory decision weight corresponding to the global path is calculated, expressed as: ； in, Represents the global path decision weights. The dopamine signal threshold; The short-term memory decision weights corresponding to local paths are represented as follows: ； in, Indicates the weight of the local path decision; After determining the global path weights and local path weights, the local paths are calibrated and corrected under global constraints to obtain candidate navigation paths, represented as follows: ； in, This indicates the candidate navigation path after fusion calibration. This is the original STM obstacle avoidance path. For global LTM paths, This represents the path calibration coefficient.

6. The unmanned cluster brain-like autonomous navigation method based on swarm intelligence according to claim 1, characterized in that, Step 5, which involves memory-linked calibration and final arbitration output of the candidate navigation decision set, specifically includes: The fused candidate navigation paths are converted into local decision vectors, represented as follows: ； in, This represents the local navigation decision vector of the i-th agent. , To control the amount of obstacle avoidance, The direction of the target movement; The local decisions of neighboring agents are broadcast through a cluster communication mechanism, and a weighted average method is used to generate a group consensus decision, which is represented as: ； in, This represents the consensus decision obtained by agent i. Let i be the neighborhood set of agent i. Indicates the neighborhood connection weight. Assign weight to its own decision-making; To incorporate historical experience, risk area information, and current consensus results into the arbitration process, a reference decision is introduced to adaptively modify the group consensus decision. The expression is as follows: ； in, This represents the final navigation decision of agent i. This refers to reference decisions generated from historical navigation experience, long-term memory paths, or prior knowledge of risk areas. Indicates the adaptive adjustment coefficient; Finally, to evaluate the effectiveness of the final navigation decision and to provide feedback and correction for subsequent navigation arbitration processes, an evaluation index for the effectiveness of the decision is constructed, with the following expression: ； in, This represents the performance evaluation metric for agent i. Indicates positional deviation. Indicates heading deviation. Indicates the neighborhood cooperative distance error. , , To evaluate the weights, , , These represent the maximum permissible values ​​for the corresponding errors.

Images