A multi-agent cooperative control method for sea-land-air-space cross-domain unmanned cluster
By adopting a hierarchical-frequency-domain decoupling model and a distributed reinforcement learning mechanism, the problem of cross-domain unmanned clusters coordinating between heterogeneous platforms is solved, achieving efficient and robust cross-domain collaborative control and improving the cluster's autonomous survivability and task success rate in complex environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI DONGHAI VOCATIONAL & TECH COLLEGE
- Filing Date
- 2026-04-03
- Publication Date
- 2026-06-26
AI Technical Summary
Existing technologies struggle to achieve efficient and robust collaboration among unmanned swarms spanning sea, land, air, and space in dynamic and uncertain cross-domain scenarios. In particular, autonomous collaboration and decision-making are difficult to achieve when communication between heterogeneous platforms is limited and information is incomplete.
A hierarchical-frequency-domain decoupling model is adopted, combined with a dynamically updated cross-domain collaborative situation map and an event-triggered distributed reinforcement learning mechanism. Global task allocation and collaborative decision-making are achieved through local perception and sparse communication, and multi-source sensor data are fused for environmental perception and action adjustment.
It effectively reduces system design complexity, achieves unified information cognition and efficient collaboration among heterogeneous platforms, possesses strong resilience and flexibility, and enhances the cluster's autonomous survivability and mission success rate in complex environments.
Smart Images
Figure CN122284460A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of cross-domain unmanned swarm collaborative control technology, and in particular to a multi-agent collaborative control method for cross-domain unmanned swarms. Background Technology
[0002] With the rapid development of unmanned systems technology, autonomous platforms such as drones, unmanned ground vehicles, and unmanned surface vessels have demonstrated enormous potential in fields such as reconnaissance, logistics, and emergency rescue. However, single types of unmanned systems are insufficient to cope with increasingly complex mission scenarios, such as large-scale disaster relief and wide-area environmental monitoring. These scenarios typically involve multi-dimensional spaces including sea, land, air, and even space, requiring coordinated execution of diverse tasks such as reconnaissance, transportation, and response. Therefore, cross-domain unmanned swarms composed of multiple heterogeneous unmanned platforms have become an important direction for current research and application. Cross-domain unmanned swarm collaborative control faces multiple challenges. Heterogeneous platforms have significant differences in dynamic characteristics, motion capabilities, communication bandwidth, and sensor configurations. Traditional centralized or homogeneous multi-agent control methods are difficult to apply directly. Complex task environments often have problems such as limited communication and incomplete information, requiring the swarm to have autonomous collaboration and decision-making capabilities under limited interaction. Existing methods mostly focus on collaboration within a single domain or rely on continuous and stable global communication, making it difficult to achieve efficient and robust collaboration in dynamic and uncertain cross-domain scenarios. Therefore, there is an urgent need for a cross-domain collaborative control method that can effectively handle heterogeneity, communication constraints and environmental uncertainties, and can integrate multi-source sensing information to dynamically adapt to complex environments, thereby improving the overall mission efficiency and adaptability of cross-domain unmanned swarms. Summary of the Invention
[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing a multi-agent collaborative control method for cross-domain unmanned swarms across land, sea, air, and space.
[0004] To achieve the above objectives, the present invention adopts the following technical solution: A multi-agent cooperative control method for cross-domain unmanned swarms across land, sea, air, and space includes the following steps: S1: Construct a decoupling model to separate the complexity of collaborative control along three dimensions: hierarchical, frequency-based, and domain-based. S2: Based on the decoupling model, a unified collaborative strategy is designed. The unified collaborative strategy generates personalized control objectives for agents of different dimensions according to a dynamically updated cross-domain collaborative situation map, which are adapted to their own dynamics and roles. S3: Introduce an event-triggered distributed reinforcement learning mechanism for each agent. The mechanism enables the agent to initiate communication and interaction with neighboring agents or upper-level nodes only when preset triggering conditions are met, so as to achieve intelligent collaboration under communication constraints. S4: Based on a distributed reinforcement learning mechanism, dynamic allocation and collaborative decision-making of global tasks are achieved through local perception and sparse event-driven communication of each agent. S5: Each intelligent agent integrates data from multiple sensors and inputs the integrated environmental information into the collaborative control algorithm in real time to dynamically adjust actions and task allocation.
[0005] As a further aspect of the present invention: the decoupling model in S1 includes: Layered Dimensions: A three-layer control architecture is established, comprising a task planning layer, a collaborative decision-making layer, and a motion control layer; Frequency division dimension: Based on the response speed of each agent's physical dynamics, different control command update frequencies are set for its motion control layer; Domain-based dimension: Define the agent's exclusive task space and cross-domain collaboration rules based on the agent's mobility and scope of action.
[0006] As a further aspect of this invention: the cross-domain collaborative situational diagram includes the global task objective, the states of each agent, environmental information, and inter-domain collaborative constraints, specifically generating personalized sub-tasks for heterogeneous agents as follows: The collaborative situation map generates a regional search path grid for airspace platforms, generates target point reconnaissance or material receiving instructions for land platforms, and generates standby point coordinates or material transfer instructions for sea platforms.
[0007] As a further aspect of the present invention, the event-triggered distributed reinforcement learning mechanism specifically includes: Each agent's local agent module includes event triggers and a distributed policy network; An event trigger defines a set of triggering conditions, including a sudden change in its own state, discovery of a key target, a difference in situation assessment with a neighbor exceeding a threshold, or receipt of a collaboration request. Distributed policy networks take as input the interactive information received from local observations and event triggers, and output the action decisions of individuals; The mechanism adopts a hybrid centralized training and distributed execution paradigm.
[0008] As a further aspect of the present invention, the triggering condition of the event trigger is a learnable parameter that can adaptively adjust the triggering threshold according to the importance of the task or the state of the communication channel, so as to balance the collaborative performance and communication overhead.
[0009] As a further aspect of the present invention, the specific steps for implementing dynamic allocation of global tasks in the distributed consensus negotiation process include: Each agent calculates its competence or willingness to perform various available tasks based on local information and neighbor information obtained through event triggering using a local policy network. Through sparse event-driven communication, task competence or willingness information is exchanged among local neighbors. Each agent autonomously selects or adjusts tasks based on local negotiation results, enabling the cluster to emerge with an optimal task allocation pattern without continuous intervention from a central scheduler.
[0010] As a further aspect of the present invention: in the frequency division dimension, a high-frequency update is set for the attitude control loop of the air platform, a medium-frequency update is set for the path tracking loop of the ground platform, and a low-frequency update is set for the waypoint control loop of the water platform.
[0011] As a further aspect of the present invention: the fusion of multi-source sensor data specifically includes: The raw data collected by the sensors of each platform are preprocessed by denoising, filtering and calibration. Kalman filtering, extended Kalman filtering, or particle filtering algorithms are used to fuse information from different sensors such as lidar, camera, sonar, and inertial navigation unit to generate comprehensive environmental situation information. By inputting comprehensive environmental situation information into the collaborative control process, the intelligent agent can adjust its actions in real time according to changes in the environment.
[0012] As a further aspect of the present invention: the method is executed collaboratively by local intelligent agent modules deployed on each unmanned platform and at least one collaborative management node, wherein: The local intelligent agent module is responsible for performing local perception, event-triggered communication, distributed decision-making, and sensor data fusion. The collaborative management node is responsible for issuing global tasks, maintaining cross-domain collaborative situation maps, and making high-level strategy adjustments in the initial stage of a task.
[0013] As a further aspect of the present invention: In the application scenario, when the airspace platform detects a suspected target, it triggers a cross-domain collaborative event and sends a refined confirmation request to the land-based platform; the material delivery task, through a distributed consensus negotiation process, autonomously allocates a collaborative scheme based on the location, load, and power of each platform, with the airspace platform performing near-point delivery, the sea-based platform performing long-point transit, and the land-based platform performing final distribution.
[0014] Compared with existing technologies, this invention provides a multi-agent cooperative control method for cross-domain unmanned swarms across land, sea, air, and space, which has the following beneficial effects: 1. By using a triple decoupling model of hierarchical-frequency-domain separation, the extreme complexity of cross-domain collaboration is separated along three orthogonal dimensions: task level, control frequency, and physical space. Standard interfaces are defined to encapsulate the dynamics of heterogeneous platforms, which greatly reduces the complexity of system design and analysis and achieves optimized matching of computing resources and efficient and robust control of the system.
[0015] 2. Based on the dynamically updated cross-domain collaborative situation map, a unified understanding of information across the entire domain is achieved. Based on this, personalized sub-tasks are generated for different platforms to adapt to their roles and dynamic capabilities, effectively solving the problem of information silos between heterogeneous platforms and realizing intelligent conversion and efficient collaboration from global goals to individual executable actions.
[0016] 3. An event-triggered distributed reinforcement learning mechanism is introduced, which performs sparse communication only when there are key state changes or collaborative needs, greatly reducing communication overhead. Combined with hybrid training and distributed execution paradigms, the cluster can still emerge near-globally optimal adaptive collaborative behavior through local perception and limited interaction under communication constraints, which has strong resilience and flexibility.
[0017] 4. By using multi-sensor data fusion technology (such as Kalman filtering and feature fusion) to improve the quality of environmental perception, and feeding the fused high-dimensional environmental information into the collaborative decision-making closed loop in real time, each platform can dynamically respond to complex environmental changes (such as obstacle avoidance and detour), which significantly improves the cluster's autonomous survival ability and mission success rate in unknown and unstructured scenarios.
[0018] 5. The methodology has a clear three-layer architecture of platform-local intelligent agent-collaborative node, and verifies the closed-loop process from event triggering, distributed negotiation to cross-domain task execution through a complete search and delivery application scenario example. It demonstrates the connection from theoretical model to engineering implementation and has high practicality and scalability potential.
[0019] The parts of this device not covered herein are the same as or can be implemented using existing technologies. This invention has a simple structure and is easy to operate. Attached Figure Description
[0020] Figure 1 This is a flowchart illustrating the overall process of a multi-agent cooperative control method for cross-domain unmanned swarms proposed in this invention, applicable to land, sea, air, and space. Figure 2 This is a flowchart illustrating the global task dynamic allocation process for a multi-agent cooperative control method for cross-domain unmanned swarms proposed in this invention. Detailed Implementation
[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0022] Example 1: A multi-agent cooperative control method for cross-domain unmanned swarms across land, sea, air, and space, which achieves cooperative control based on the following devices: unmanned aerial vehicles (UAVs), unmanned ground vehicles (UAVs), and unmanned surface vessels (UAVs). The control method includes the following steps: S1: Utilize unmanned platforms in different domains such as sea, land, air, and space to construct a hierarchical-frequency-domain decoupling model, separating the complexity of collaborative control along three dimensions; The system is structured as follows: a task planning layer (managed by a cloud server or high-altitude drone, responsible for defining the monitoring area and determining the delivery point), a collaborative decision-making layer (local intelligent agents on each platform, responsible for negotiating specific search paths, target tracking, and material handover), and a motion control layer (lower-level controllers on each platform, responsible for tracking waypoints or trajectories given by the upper layer). Based on the time constant and response speed of the physical dynamics of each platform, different control command update frequencies are set for platforms of different domains or types; for example, a high-frequency attitude control loop is set for UAVs, a medium-frequency path tracking loop is set for unmanned vehicles, and a low-frequency waypoint control loop is set for unmanned surface vessels. Define the airspace as the rapid search and reconnaissance domain, the land domain as the detailed reconnaissance and land-based material receiving domain, and the sea domain as the water-based material transit domain; set different coordination rules for them, for example, after the airspace platform detects a suspected target, it sends a detailed confirmation request to the land-based platform, rather than directly intervening in land-based operations; Based on the platform's mobility (such as air, ground, water, and orbit) and range of action, define its own task space and coordination rules to avoid ineffective cross-domain coupling calculations; The dynamic decoupling model encapsulates the unique dynamics of each agent using a standard interface, exposing only its reachable velocity set, maximum acceleration, turning radius, and other constraint parameters to the outside world for use by upper-layer collaborative strategies. S2: The collaborative management node maintains a cross-domain collaborative situation map, which marks the disaster area, material distribution points, searched areas, and the location and status of each platform (such as battery level and payload). Based on this map, the unified strategy generates a sub-task for the UAV to "search the grid in a zigzag pattern", a sub-task for the unmanned ground vehicle to "go to coordinates (X, Y) to collect samples", and a sub-task for the unmanned surface vessel to "stand by at coordinates (P, Q) to receive airdropped supplies". Based on the decoupling model, a unified collaborative strategy generation mechanism is designed. The core of this strategy is to define a cross-domain collaborative situational awareness graph, which is a dynamically updated data structure containing: Global / local task objectives; The state of each agent (position, velocity, resources, health); Environmental information (obstacles, threats, communication quality); Inter-domain collaboration interfaces and constraints (e.g., airspace platforms provide guidance to ground platforms, and maritime platforms provide resupply point coordinates to airspace platforms). The unified strategy generates personalized control objectives or sub-tasks that are adapted to the dynamics and roles of agents at different levels, frequencies, and domains based on the collaborative situation map, rather than the original control quantities. S3: Each agent is equipped with a local agent module, which includes: Local sensing and feature extraction unit: processes its own sensor data to generate local observations; Event trigger: Define a set of triggering conditions, and initiate communication and interaction with neighboring agents or upper-level nodes only when the conditions are met (e.g., a drastic change in its own state, discovery of a key target, a difference in situation assessment with neighboring agents exceeding a threshold, or receipt of a specific cooperation request). Distributed policy network: A trained neural network whose inputs are local observations and the most recently received neighbor / global summary information triggered by an event, and whose outputs are individual action decisions (such as the next direction of movement, task selection, and communication requests). Local value / comment network: used to evaluate actions and, when an event triggers an interaction, to exchange small-scale, sparse gradient or value information with surrounding agents to achieve co-evolution and optimization of the strategy; The distributed reinforcement learning mechanism adopts a hybrid centralized training and distributed execution paradigm: in the offline or simulated training phase, global information is used to guide the training of policy networks of multiple agents; in the online execution phase, each agent makes independent decisions relying only on its local policy network and event-triggered communication. S4: Global tasks (such as searching area A, monitoring target B, and transporting supplies C) are published to the cluster; the cluster dynamically allocates these tasks through the following process: Each agent, based on its local perception and limited non-global information obtained through event-triggered communication, uses its own distributed policy network to calculate its own competence or willingness value for each available task. Through sparse, event-driven communication, these task evaluation information are exchanged among local neighbors, forming a distributed consensus negotiation process that spreads throughout the cluster. Each agent autonomously selects or adjusts its tasks based on local negotiation results. Without the need for a central scheduler, the entire cluster emerges with a globally near-optimal or satisfactory task allocation pattern, which can be dynamically adjusted according to the environment. The conditions of the event trigger are designed as learnable parameters, which enable it to adaptively adjust the trigger threshold according to the importance of the task and the state of the communication channel, thus balancing collaborative performance and communication overhead. The distributed reinforcement learning mechanism adopts a hybrid centralized training and distributed execution paradigm: in the offline or simulated training phase, global information is used to guide the training of policy networks of multiple agents; in the online execution phase, each agent makes independent decisions relying only on its local policy network and event-triggered communication. S5: The UAV is equipped with sensors such as thermal imaging, lidar, cameras, and inertial measurement units to collect information on terrain, obstacles, and weather conditions; the unmanned surface vessel is equipped with sonar, water temperature sensors, and salinity sensors to collect marine environmental information.
[0023] The acquired sensor data is processed through operations including noise reduction, filtering, and calibration to improve data quality and reliability. For example, LiDAR data is filtered to remove noise points, and camera images are calibrated to eliminate distortion. By employing data fusion algorithms, such as Kalman filtering, extended Kalman filtering, and particle filtering, data from different sensors are fused. For example, distance information from lidar and image information from cameras are fused to more accurately identify the location and shape of obstacles. Furthermore, sonar data is fused with data from water temperature and salinity sensors to provide a more comprehensive understanding of the marine environment. The fused environmental information is input into the collaborative control algorithm, enabling each unmanned platform to adjust its actions and task allocation according to the environmental information, thereby improving its adaptability to complex environments. When the UAV detects an obstacle ahead, it can automatically adjust its flight trajectory to avoid the obstacle; when the unmanned surface vessel detects a change in the marine environment, it can adjust its navigation route to continue completing the mission.
[0024] Example 2: A multi-agent cooperative control method for cross-domain unmanned swarms across land, sea, air, and space. The multi-agent cooperative control method is based on a control system, which includes: Multiple heterogeneous unmanned platforms: belonging to different physical domains such as sea, land, air, and space, and equipped with necessary sensors, actuators, and local computing units; Local intelligent agent module: Deployed on each unmanned platform, used to perform the local perception, event triggering, and distributed decision-making functions in steps S3 and S4; Collaborative Management Node: Used to issue global tasks, receive key situation reports, and perform high-level adjustments to cross-domain collaborative situation map maintenance and unified strategies in steps S1 and S2 during the initial stage of a mission or when necessary. This node can be located in the cloud, on a satellite, or on an unmanned platform.
[0025] Example 3: A multi-agent cooperative control method for cross-domain unmanned swarms across land, sea, air, and space. To facilitate the demonstration of the control flow of the cooperative control method in a practical application scenario, this example introduces an event-triggered example, with the following operation process: Search the entire area and deliver supplies; each platform begins to act based on local observations and initialization information; when a drone senses a heat source signal through its sensors, it uploads this information to other platforms within its communication range based on the sensors; the unmanned ground vehicle receives this message, and its local policy network, combined with its own observations, decides on a new action to go to the heat source for confirmation. Regarding the delivery of supplies, each platform conducts distributed negotiation based on its own location, power level, and distance from the delivery point through sparse, event-driven communication (such as exchanging task load information when two platforms are close enough). Final result: The distribution scheme is as follows: multiple drones are responsible for near-point delivery, unmanned surface vessels use their large payload to handle long-point delivery and transport to the shore, and then unmanned ground vehicles are used for final distribution.
[0026] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A multi-agent cooperative control method for cross-domain unmanned swarms across land, sea, air, and space, characterized in that, Includes the following steps: S1: Construct a decoupling model to separate the complexity of collaborative control along three dimensions: hierarchical, frequency-based, and domain-based. S2: Based on the decoupling model, a unified collaborative strategy is designed. The unified collaborative strategy generates personalized control objectives for agents of different dimensions according to a dynamically updated cross-domain collaborative situation map, which are adapted to their own dynamics and roles. S3: Introduce an event-triggered distributed reinforcement learning mechanism for each agent. The mechanism enables the agent to initiate communication and interaction with neighboring agents or upper-level nodes only when preset triggering conditions are met, so as to achieve intelligent collaboration under communication constraints. S4: Based on a distributed reinforcement learning mechanism, dynamic allocation and collaborative decision-making of global tasks are achieved through local perception and sparse event-driven communication of each agent. S5: Each intelligent agent integrates data from multiple sensors and inputs the integrated environmental information into the collaborative control algorithm in real time to dynamically adjust actions and task allocation.
2. A multi-agent cooperative control method for cross-domain unmanned swarms based on land, sea, air, and space, as described in claim 1, is characterized in that... The decoupling model in S1 includes: Layered Dimensions: A three-layer control architecture is established, comprising a task planning layer, a collaborative decision-making layer, and a motion control layer; Frequency division dimension: Based on the response speed of each agent's physical dynamics, different control command update frequencies are set for its motion control layer; Domain-based dimension: Define the agent's exclusive task space and cross-domain collaboration rules based on the agent's mobility and scope of action.
3. A multi-agent cooperative control method for cross-domain unmanned swarms based on land, sea, air, and space, as described in claim 1, is characterized in that... The cross-domain collaborative situational awareness map includes the global task objective, the states of each agent, environmental information, and inter-domain collaborative constraints. Specifically, it generates personalized sub-tasks for heterogeneous agents as follows: The collaborative situation map generates a regional search path grid for airspace platforms, generates target point reconnaissance or material receiving instructions for land platforms, and generates standby point coordinates or material transfer instructions for sea platforms.
4. A multi-agent cooperative control method for cross-domain unmanned swarms based on land, sea, air, and space, as described in claim 1, is characterized in that... Event-triggered distributed reinforcement learning mechanisms specifically include: Each agent's local agent module includes event triggers and a distributed policy network; An event trigger defines a set of triggering conditions, including a sudden change in its own state, discovery of a key target, a difference in situation assessment with a neighbor exceeding a threshold, or receipt of a collaboration request. Distributed policy networks take as input the interactive information received from local observations and event triggers, and output the action decisions of individuals; The mechanism adopts a hybrid centralized training and distributed execution paradigm.
5. A multi-agent cooperative control method for cross-domain unmanned swarms based on land, sea, air, and space, as described in claim 4, is characterized in that... The triggering conditions of the event trigger are learnable parameters that can adaptively adjust the triggering threshold according to the importance of the task or the state of the communication channel to balance collaborative performance and communication overhead.
6. A multi-agent cooperative control method for cross-domain unmanned swarms based on land, sea, air, and space, as described in claim 1, is characterized in that... The specific steps involved in the distributed consensus negotiation process to achieve dynamic allocation of global tasks include: Each agent calculates its competence or willingness to perform various available tasks based on local information and neighbor information obtained through event triggering using a local policy network. Through sparse event-driven communication, task competence or willingness information is exchanged among local neighbors. Each agent autonomously selects or adjusts tasks based on local negotiation results, enabling the cluster to emerge with an optimal task allocation pattern without continuous intervention from a central scheduler.
7. A multi-agent cooperative control method for cross-domain unmanned swarms based on land, sea, air, and space, as described in claim 2, is characterized in that... In the frequency division dimension, a high-frequency update is set for the attitude control loop of the air platform, a medium-frequency update is set for the path tracking loop of the ground platform, and a low-frequency update is set for the waypoint control loop of the surface platform.
8. A multi-agent cooperative control method for cross-domain unmanned swarms based on land, sea, air, and space, as described in claim 1, is characterized in that... The fusion of multi-source sensor data specifically includes: The raw data collected by the sensors of each platform are preprocessed by denoising, filtering and calibration. Kalman filtering, extended Kalman filtering, or particle filtering algorithms are used to fuse information from different sensors such as lidar, camera, sonar, and inertial navigation unit to generate comprehensive environmental situation information. By inputting comprehensive environmental situation information into the collaborative control process, the intelligent agent can adjust its actions in real time according to changes in the environment.
9. A multi-agent cooperative control method for cross-domain unmanned swarms based on land, sea, air, and space, as described in claim 1, is characterized in that... The method is executed collaboratively by local intelligent agent modules deployed on each unmanned platform and at least one collaborative management node, wherein: The local intelligent agent module is responsible for performing local perception, event-triggered communication, distributed decision-making, and sensor data fusion. The collaborative management node is responsible for issuing global tasks, maintaining cross-domain collaborative situation maps, and making high-level strategy adjustments in the initial stage of a task.
10. A multi-agent cooperative control method for cross-domain unmanned swarms based on land, sea, air, and space, as described in claim 1, is characterized in that... In application scenarios, when the airspace platform detects a suspected target, it triggers a cross-domain collaborative event and sends a detailed confirmation request to the land-based platform. The material delivery task, through a distributed consensus negotiation process, autonomously allocates a collaborative scheme based on the location, load, and power of each platform, with the airspace platform performing near-point delivery, the sea-based platform performing long-point transit, and the land-based platform performing final distribution.