Embodied intelligence based multi-agent on-demand deployment and self-organizing wireless communication method and system

Abstract

The application discloses a kind of multi-agent on-demand deployment and self-organizing wireless communication system based on embodied intelligence, for realizing end-to-end closed loop operation of agent autonomous exploration, on-demand dispatch, target access and service data backhaul in unknown environment.System is through including multiple intelligent agents with mobile and communication capabilities, target user and starting base station, wherein intelligent agent sets out from starting base station, independently decides according to local observation, forms decentralized control strategy;Multi-agent local perception and independent decision-making combine the dispatch control of base station strategy center, form the hybrid management of physical communication graph and logical tree topology, and adopt the Actor-Critic reinforcement learning architecture of communication perception combined with centralized training, decentralized execution strategy, so that agent can still make decision in favor of the overall network under partially observable environment, realize complete closed loop exploration, access and data backhaul.

Description

Technical Field

[0001] This invention relates to multi-agent systems, wireless self-organizing networks, and reinforcement learning techniques, and particularly to a method and system for on-demand deployment and self-organizing wireless communication of multi-agent systems based on embodied intelligence. Background Technology

[0002] In traditional wireless communication networks, coverage areas are typically planned and deployed based on fixed base stations or known target locations. This approach has significant limitations in scenarios such as emergency communications, underground spaces, or remote areas: firstly, fixed infrastructure cannot cover vast or complex environments, especially in post-disaster scenarios where base stations may be damaged, leading to wireless communication outages; secondly, existing multi-agent systems often rely on globally observable or centralized control strategies, making them difficult to adapt in partially observable environments and lacking dynamic deployment and self-organizing capabilities. Furthermore, traditional methods often involve "finding..."

[0003] The "User", "Connecting Users", and "Maintaining the Network" modules are handled as independent modules, lacking an end-to-end closed-loop mechanism. At the same time, communication indicators such as link capacity, latency and load are not fully considered, making it difficult to guarantee network performance and stability.

[0004] Therefore, there is an urgent need for a system and method that can support autonomous exploration, on-demand deployment, target access, and service backhaul of intelligent agents in partially observable and dynamically unknown environments, and can optimize the overall network performance by comprehensively considering link metrics. Summary of the Invention

[0005] To address the aforementioned issues, this invention discloses a multi-agent on-demand deployment and self-organizing wireless communication method and system based on embodied intelligence. This system enables agents to autonomously explore unknown environments, deploy on demand, access targets, and transmit service data back in an end-to-end closed-loop manner. The system is suitable for scenarios such as disaster relief emergency communication, wireless communication networking in underground spaces or remote areas, where a base station is located at the starting point, but the target coverage area may exceed the range of fixed infrastructure. The agents do not need to know the target location in advance and can dynamically adapt to environmental changes, achieving flexible network deployment and efficient maintenance.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method and system for on-demand deployment and self-organizing wireless communication using multi-agent systems based on embodied intelligence comprises the following modules: multiple agents with mobility and communication capabilities, target users, and a base station. Agents start from the base station and make independent decisions based on local observations, forming a decentralized control strategy. Each agent can only move up, down, left, or right or remain stationary at each time step, and can decide whether to initiate a deployment request based on local observations and neighbor information. During movement, agents maintain direct or indirect communication connections with the base station and recursively maintain a logical tree topology to ensure that deployment requests and service data can be transmitted back to the base station. If the Manhattan distance between agents, and between agents and target users or base stations, does not exceed a preset threshold... If so, it is considered that a valid communication link exists.

[0008] The system works as follows:

[0009] First, the initial intelligent agent starts from the starting base station and has limited perception capabilities within its perception range. The system can successfully acquire information about the environment and neighboring agents. Each agent possesses mobility, and at each step can only move up, down, left, or right, or remain stationary. Based on its own observations and neighbor information, it decides whether to initiate a deployment request. If a deployment request is initiated, it is transmitted back to the starting base station via a tree structure, which then decides whether to respond to the request. Agents can make deployment request decisions both during the exploration phase and after finding the target.

[0010] The agent maintains a direct or indirect communication connection with the starting point throughout its movement. Between agents, and between an agent and the starting point or target user, if the Manhattan distance is less than or equal to a preset threshold... If a valid communication link exists, it is considered to exist. Next, when an agent detects a target user, it determines whether direct access is possible based on the target user's link capacity, latency, and load requirements. If the current network cannot meet the requirements, the agent waits for other cooperating agents or initiates a request for reinforcement. The base station policy center then decides whether to respond to the reinforcement and its location based on the global topology and communication indicators. The reinforced agent is added to the physical communication graph and forms a parent-node relationship in the logical tree topology to ensure the continuity of the backhaul path and the controllability of network management. Finally, the target user's services are reliably delivered to the base station through the backhaul path, achieving end-to-end closed-loop operation.

[0011] A method for on-demand deployment and self-organizing wireless communication of multi-agent systems based on embodied intelligence, specifically including the following steps:

[0012] Step 1: Agent Deployment and Initialization. The initial agent is deployed at the base station at the starting point, and the environment and parameters are initialized upon system startup. In this non-restrictive example, the environment boundary can be set as a discrete grid, and a preset Manhattan distance threshold is used to determine communication reachability. The system synchronously establishes two types of topologies: one is a physical communication graph reflecting physical connectivity relationships. The first function is used to determine the existence of a link based on the Manhattan distance threshold, and to carry link capacity, latency, and load metrics in real time; the second function is a logical tree topology used to manage request return and hierarchical scheduling. Using the base station as the root, the system ensures that each agent has a unique backhaul path at any given time. Each agent is assigned a unique identifier, an initial parent node reference (pointing to the uplink parent node), a neighbor list, and a local link state table. It also establishes local awareness of neighboring agents, environmental boundaries, and the origin through a built-in perception module. To avoid backhaul interruptions, the system enables heartbeat and link detection mechanisms during initialization and sets path integrity and connectivity thresholds.

[0013] Step Two: Autonomous Exploration and Local Decision-Making by the Agent. In this system, the agent's decision-making is based on a locally observed Actor-Critic reinforcement learning framework. The agent starts from the base station and moves across the grid with discrete step sizes; each step allows only one action: "up, down, left, right, or stay still." Before each action, the agent determines feasibility based on local observations and the states of its neighbors: if the expected action would disrupt the backhaul path between itself or its subtree and the base station, the action is rejected and replaced by an alternative action. Simultaneously, the agent determines whether to request additional deployments based on local link quality (such as the remaining capacity of neighbors, queuing latency, and current load) and spatial reachability. All decisions are made in a decentralized manner, meaning the agent relies only on local and neighbor information, not on global real-time information, thus ensuring the system's scalability and robustness.

[0014] Step 3: Topology Maintenance and Link Monitoring. During the exploration process, the system continuously monitors the physical communication graph. With logical tree topology Perform consistency maintenance. Physical communication diagram. According to the Manhattan distance threshold The rules recalculate adjacency relationships at each step and update the capacity, latency, and load metrics of each link; logical tree The hierarchical constraint of "single parent, multiple children" is maintained to avoid loops. Each agent immediately broadcasts the minimum necessary link state changes after moving, refreshing the neighbor list and parent node selection. Parent node selection is based on a comprehensive judgment prioritizing "maintaining connectivity, minimum hop count, sufficient remaining capacity, and low end-to-end latency." When parent node link degradation or disconnection is detected, the child node triggers a rapid parent node reselection process, reconnecting among reachable neighbors to ensure the continuity of the backhaul path. Topology and link states are recursively transmitted back to the base station from bottom to top in a tree structure, forming the global perspective required for deployment decisions and operational scheduling.

[0015] Step four, target discovery and access judgment. When the Manhattan distance between a certain agent and the target user is not greater than a threshold... When the target is detected, it is determined that "the target has been discovered." Subsequently, the agent assesses accessibility based on the target user's business requirements, including (but not limited to) the target user's minimum link capacity threshold, end-to-end latency limit, and the degree of matching between the estimated load and the remaining capacity of the current path. If the assessment is satisfactory, the agent establishes an access relationship with the target user and attaches the target as a leaf node to the logic tree. On the corresponding branch; if not satisfied, the agent enters one of two types of waiting logic: one is to wait for a cooperating agent to approach to enhance the local network capability; the other is to initiate a dispatch request to supplement a new agent near the target or at a key relay location to meet the service quality requirements of the target access.

[0016] Step 5: Reporting and Deployment Requests. Once a deployment request is generated, the information is processed along the logic tree. The request is transmitted from bottom to top to the base station policy center. The request message includes: the trigger node identifier and location, a summary of the node's parent-child relationship, surrounding physical link quality statistics (capacity, latency, load), candidate relay location suggestions, and their reachability verification. An expansion policy module is set up at the starting base station. Based on the received expansion request and its accompanying node information, it performs a comprehensive evaluation in the order of "coverage gain—connectivity constraints—link quality—resource overhead" to determine whether to respond to the expansion and the specific expansion location. The selected location must meet the following requirements: 1) The Manhattan distance to at least one node in the existing network does not exceed... 1) Ensures physical connectivity upon joining; 2) Significantly increases the target access capacity or reduces end-to-end latency after joining; 3) Does not disrupt the hierarchical stability of the logic tree. After the dispatch command is issued, the new agent joins immediately upon arriving at the designated location. and in The parent node relationship is established, and the routing table and return path are updated synchronously.

[0017] Step Six: Target User Data Backhaul and Path Selection. During system operation, the service load generated by the target user is uploaded layer by layer to the starting node along the backhaul path maintained by the intelligent agent, realizing data transmission from the user to the core network. Each intelligent agent, during the backhaul process, not only acts as a simple forwarding node but also monitors its neighbors' status and link quality in real time, dynamically evaluating parameters such as link load, latency, and packet loss rate, and adjusting the backhaul path as needed to optimize link load balancing. The entire backhaul process is based on a logical tree topology, where all nodes record their parent node information, allowing the backhaul path to be updated recursively, thus quickly switching to a backup path when link quality deteriorates or a node fails. Simultaneously, when a new intelligent agent joins the network, it can be seamlessly mounted in a suitable position within the topology tree structure, without affecting the stability of existing paths, while providing more redundancy and bandwidth resources for overall data backhaul.

[0018] The beneficial effects of this invention are specifically manifested in the following aspects:

[0019] 1. Significantly Enhanced Dynamic Adaptability: This invention introduces the concept of embodied intelligence, enabling multiple agents to instantly perceive and proactively respond to environmental changes when performing deployment and communication tasks. Compared to traditional solutions relying on static path planning or fixed communication topologies, the agents in this system can automatically adjust their positions and communication links when encountering complex scenarios such as rapid changes in user location, environmental obstruction, interference, or node failure, thereby ensuring the stability and continuous coverage of the overall system. This dynamic adaptability not only improves network robustness but also reduces the probability of service interruptions due to unforeseen events, significantly enhancing the system's reliability and application scope in uncertain environments.

[0020] 2. Improved On-Demand Coverage and Resource Utilization: This invention employs an on-demand deployment mechanism, combined with multi-agent collaborative decision-making, to achieve precise matching and dynamic allocation of resources even in situations of uneven or sudden changes in user demand. Through autonomous agent perception and distributed negotiation, the number of deployed agents and wireless communication links can be automatically adjusted according to user density and service demand intensity in different areas, thereby avoiding the problems of "over-coverage" or "under-coverage" common in traditional solutions. This design significantly improves resource utilization, reduces system energy consumption and hardware investment, and enables more extensive and efficient user services under limited resource conditions.

[0021] 3. Enhanced Wireless Communication Link Stability and Self-Organization Capability: This system introduces a self-organizing wireless communication mechanism, enabling multiple agents to autonomously establish and maintain a stable wireless communication network even in the absence of fixed base stations or core control nodes. Agents can dynamically select neighboring nodes based on communication quality indicators and form a globally optimal network topology through local decision-making. This design overcomes the reliance on single-point control in centralized schemes and avoids network-wide paralysis caused by the failure of a central node. Simultaneously, the multi-agent network can quickly reconstruct its topology when nodes move, encounter environmental interference, or experience partial node failures, thus ensuring the continuity and reliability of wireless communication.

[0022] 4. Significantly Enhanced Learning and Decision-Making Capabilities: This invention introduces an embodied intelligence-driven Actor-Critic learning architecture, enabling agents to continuously accumulate experience during environmental exploration and task execution, and jointly optimize perception, action, and communication strategies. Unlike traditional rule-based or single-objective optimization schemes, the agents in this system can achieve dynamic trade-offs under multi-dimensional constraints through deep reinforcement learning, such as simultaneously considering coverage, communication quality, and energy consumption. This mechanism not only improves the decision-making level of individual agents but also achieves overall optimal results in multi-agent collaboration, greatly enhancing the system's long-term learning and evolution capabilities in complex environments.

[0023] 5. Scalability and Broad Application Prospects: The system architecture of this invention possesses excellent scalability, capable of adapting to the deployment needs of multiple agents in different scales and scenarios. Whether it's rapid networking in large-scale disaster relief, temporary communication enhancement in urban environments, or infrastructure-weak scenarios in future 6G / space wireless communication, this system can achieve rapid deployment and stable operation. Due to the adoption of a distributed collaborative decision-making mechanism and an on-demand coverage strategy, the system maintains high efficiency and stability even when expanded to dozens or even hundreds of agents. This scalability not only lowers the application threshold but also lays a solid foundation for future multi-scenario promotion and industrial applications. Attached Figure Description

[0024] Figure 1 This is a block diagram of a multi-agent on-demand deployment and self-organizing wireless communication system architecture based on embodied intelligence;

[0025] Figure 2 This is a schematic diagram of the Actor-Critic reinforcement learning framework driven by local observations of the agent.

[0026] Figure 3 This is a schematic diagram of the multi-agent addition request and deployment process;

[0027] Figure 4This is a schematic diagram of agent movement and coverage area (where (a) shows the initial deployment state (Step=0), b) shows the mid-stage state of exploration and target discovery (Step=75), (c) and (d) show the on-demand deployment and network reinforcement process (Step=140 to Step=244); (e) and (f) show the task completion state (Step=345 to Step=371)). Detailed Implementation

[0028] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, and the terms "inner" and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0029] Figure 1 The overall architecture of this system is demonstrated. In this system, the agent is configured as a mobile terminal with mobility and wireless communication awareness capabilities. After system startup, system initialization is performed first: the initial agent is deployed at the starting base station, and the environmental grid boundary and Manhattan distance communication threshold are set. and the threshold values ​​for each link performance metric (minimum link capacity required by the target user). Maximum delay The system simultaneously establishes a physical communication graph and a logical tree topology with the base station as the root node. Each agent loads perception, decision-making, and communication modules and enables a heartbeat mechanism to continuously monitor network connectivity. After initialization, the system enters the main loop. Agents begin the perception and exploration phase: each agent, based on its local observations (including its own location, neighbor states, link quality, and environmental information), utilizes the reinforcement learning-based Actor-Critic learning framework (see appendix for details). Figure 2 The system performs calculations, assesses the environmental state, and generates action strategies. These action strategies should include: agent movement strategies and decisions on whether to issue a reinforcement request. If an agent determines that the current network capacity cannot meet the needs of the discovered target (e.g., insufficient link capacity), or that the network needs to be expanded to optimize coverage, it triggers a decision to issue a reinforcement request.

[0030] Subsequently, the agent makes action decisions: based on the action probability distribution output by the Actor network, it selects a movement action (up, down, left, right, or stay still). Before moving, a connectivity check is performed; if the action may cause the backhaul path with the base station to be interrupted, the action is rejected and a new action is selected.

[0031] After the decision is executed, the system continuously determines whether the target user has been detected: by calculating the Manhattan distance between the agent and the target user, if it is less than or equal to the communication threshold... If a target is detected, it is considered a target discovery. If not detected, the sensing and exploration process continues in a loop; if a target is detected, the target access process is initiated. After a target is detected, the agent first assesses whether the current network status meets the target's service requirements (such as capacity and latency). If it does, a connection is established with the target, and the target user is added as a leaf node to the logic tree. The system then enters the data backhaul link maintenance phase to ensure that service data is transmitted back along the optimized path. At the same time, the system continuously maintains and updates the topology: it refreshes the physical communication graph and logic tree structure in real time, dynamically adjusts the parent node selection according to the link status, and triggers rerouting when the link deteriorates. Regardless of whether the target access is successful or not, the system will check whether all target users have been detected. If not all have been detected, the exploration continues; if all have been detected and access is completed, the task ends.

[0032] During the exploration and access process, if an agent determines that the current network capacity cannot meet the needs of the discovered target (e.g., insufficient link capacity), or that the network scale needs to be expanded to optimize coverage, it triggers a deployment request. The request information is reported to the base station along the logic tree. After receiving the deployment request, the base station decides whether to respond to the deployment based on the global network view (e.g., topology, link load, target location). If rejected, the process returns to continue exploration; if approved, the new agent is deployed: the new agent starts from the base station, moves to the designated location, accesses the network, selects the best neighbor node as the parent node, and updates local and global routing information. Subsequently, the system re-enters the topology maintenance and update phase to ensure network stability after the new node is integrated. The system runs in a loop according to the above process until all target users are discovered and successfully accessed, and service data is reliably transmitted back to the base station, achieving end-to-end closed-loop operation.

[0033] Figure 2 This paper demonstrates the reinforcement learning framework upon which the agent dispatch mechanism in this invention relies. This framework constructs an environment that encapsulates the global state of the multi-agent system (e.g., overall network topology, locations of all target users), the local state of each agent (e.g., its own coordinates, neighbor information, link quality), and the control state (e.g., the dispatch queue of the base station policy center). The framework includes a centralized Critic (evaluator) and multiple distributed Actors (executors). The Critic employs a GNN structure, and at each time step... For the set of active agents Each intelligent agent Embedded vectors are generated by aggregating information from their neighbors. This is used to capture the interactions and topological relationships between agents. Subsequently, a learnable weight matrix is ​​used... and bias vector The activation function processes each embedding, and the global Q-value is calculated by mean pooling over all active agents. :

[0034]

[0035] in and These are the global state and joint actions, respectively. This design ensures that the Q-value estimation is independent of the agent's order of arrangement and stable under changes in team size.

[0036] Critic is trained by minimizing the temporal difference (TD) error, and the training objective is... The calculation method is as follows:

[0037]

[0038] in As a reward, As a discount factor, For the next set of states, The policy action corresponding to the next state. Target network parameters. Critic's loss function. for:

[0039] ,

[0040] Parameter updates are performed using gradient descent: ,in This is the learning rate.

[0041] Each agent maintains a distributed Actor with a random policy. Local state Mapped to composite actions Including discrete directions of movement Binary increment request decision The Actor also integrates a GNN, aggregating neighbor information through a graph attention layer, and outputs the following action: Conversion of one-hot codes:

[0042]

[0043] The Actor updates gradients using a deterministic policy:

[0044] ,

[0045] The parameters are updated as follows: ,in For Actor learning rate, This represents the batch size.

[0046] The reward function employs a hierarchical design. Global reward. Used for Critic training, balancing task performance and exploration costs:

[0047] ,

[0048] in To score the quality of communication, The number of target users already identified. To explore efficiency, Apply excessive penalties to intelligent agents. For crossing the boundary, These are the weight parameters.

[0049] Local rewards are calculated based on the agent's position in the tree structure. Agents at the bottom-level leaf nodes are incentivized to explore new regions and discover targets.

[0050] .

[0051] in, For intelligent agents At time step The number of newly explored grid cells; This is an indicator function; if the agent discovers a new target user within the current time step, the value is 1, otherwise it is 0. The density of the intelligent agent's neighbors is used to prevent excessive aggregation of intelligent agents from wasting resources; The penalty items are accumulated for various rule violations (such as moving out of bounds, position conflicts, connection interruptions, etc.) caused by the agent performing actions. These are the weight parameters.

[0052] Agents at the upper levels of the tree structure are encouraged to maintain connectivity and load balancing.

[0053] .

[0054] in, This is an indicator function; it takes the value 1 if the agent is currently providing wireless communication relay service to a discovered target user, and 0 otherwise. To address the deviation between the agent's current load and the average load level, load balancing is encouraged to prevent a single node from becoming a bottleneck. As a penalty item, Here are the weighting parameters. The agent's final reward is the weighted sum of the local and global rewards.

[0055] The on-demand deployment module is tightly integrated with reinforcement learning. The explorer agent can initiate deployment requests, which traverse the logic tree. Uploaded to the base station, along with aggregated statistical information. The base station processes the deployment request based on local exploration efficiency and neighbor density, and selects a suitable parent node for the deployment of the new agent, thereby achieving load balancing and adaptive network expansion.

[0056] Figure 3 This diagram illustrates the multi-agent deployment request and deployment process in this invention. When any agent in the network (such as agent X in the diagram) determines that deployment needs to be initiated based on its local state (e.g., a target user is found but the link capacity is insufficient, or the network scale needs to be expanded to optimize coverage), it issues a deployment request. This request message includes the identifier and location of the triggering node, a summary of the current parent-child relationship, statistics on the quality of surrounding physical links (such as capacity, latency, and load), and suggested candidate relay locations. This request is not sent directly to the base station, but is first passed along with its state to its current parent node. The request information is then passed hop-by-hop from bottom to top along the established logical tree topology (with the base station as the root node and agents Y, Z, etc., as intermediate nodes). In each hop, the intermediate node can aggregate its own and its subtree's state information (such as load status and exploration efficiency), continuously enriching the request information during the reporting process and providing a more comprehensive context for the base station's decision-making.

[0057] The deployment request eventually reaches the policy center of the base station (root node). The base station evaluates the request based on a global network view (including topology, link load, resource utilization, and the location of discovered targets) and decides whether to respond. The decision is based on: 1) the reasonableness of the request (e.g., whether the target access requirement truly exists and cannot be met by the current network); 2) the coverage gain and connectivity improvement that the deployment may bring; and 3) the current system resources and deployment cost. If the evaluation result is negative, the request is rejected and the process terminates; if the evaluation result is positive, the base station generates a deployment instruction.

[0058] The base station assigns additional locations based on the request information and the global status. The selected location must meet the following requirements: 1) Its Manhattan distance to at least one node in the existing network does not exceed the communication threshold. 1) Ensure physical connectivity; 2) Significantly improve the target access capacity or reduce end-to-end latency after addition; 3) Meet the hierarchical stability constraints of the logic tree to avoid forming loops.

[0059] After the dispatch command is issued, the new agent departs from the base station and moves to the designated location. Upon arrival, the new agent performs initial deployment and state synchronization: First, it negotiates with existing neighbors at the designated location, selecting the optimal neighbor as its parent node based on priorities such as "maintaining connectivity, minimum hop count, sufficient remaining capacity, and low end-to-end latency," thus establishing a parent-child relationship. Subsequently, the new agent synchronizes and updates its routing table, local link state, and logical tree topology, and broadcasts a join message to notify neighboring nodes to update their neighbor lists. The parent node adds it as a child node, thus expanding the logical tree topology.

[0060] Finally, the deployment was completed. The new intelligent agent (as shown in the diagram, a newly added node) successfully integrated into the network, becoming part of the physical communication graph and logical tree topology, enhancing the network capabilities of the local area, thereby meeting the access needs of target users or optimizing coverage.

[0061] Figure 4 A series of time-series snapshots demonstrate the complete working process of the system of the present invention in a simulated environment, from initial deployment to final task completion.

[0062] Figure 4 (a) shows the initial deployment state (Step=0). The base station is located at the starting point, and the initial agent has been deployed. At this time, all target users have not been discovered, the agent's coverage is limited, and an effective data return path has not yet been formed. As the system runs, the agent begins to explore the environment autonomously.

[0063] Figure 4 (b) This section shows the mid-stage of exploration and target discovery (Step=75). Agents have spread from the base station to the surrounding area, significantly expanding the coverage. Some agents have discovered target users (e.g., User 1, User 2) through movement and established initial connections with them. Data backhaul paths are beginning to form, relaying target user data towards the base station. At this point, some target users (e.g., User 4, User 5) may still remain undiscovered, and the overall network capacity is insufficient to meet the service needs of all users.

[0064] Figure 4 (c) and (d) illustrate the on-demand deployment and network reinforcement process (Step=140 to Step=244). The system triggers the on-demand deployment mechanism because discovered target users generate service requests (Data), or to find remaining undiscovered targets. New agents (such as the Agent added in scenario (d)) are deployed from base stations to critical relay locations or near target users to enhance local network capacity, reduce latency, or expand coverage. The network topology becomes more complex and dense, forming multiple reliable data backhaul paths to ensure stable data transmission from already connected users.

[0065] Figure 4(e) and (f) illustrate the task completion status (Step=345 to Step=371). Through continuous exploration and multiple effective on-demand deployments, all target users (User 1 to User 5) have been discovered and successfully connected to the network. Each target user is connected to the base station through one or more multi-hop paths composed of agents, forming a stable and reliable self-organizing wireless communication network. Service data can be reliably transmitted back to the base station along the optimized path, achieving end-to-end closed-loop operation, and the system task is completed.

[0066] The accompanying drawings and embodiments demonstrate that the system of the present invention can gradually expand its coverage from an initial state through autonomous exploration, collaborative decision-making, on-demand deployment, and topology self-organization of intelligent agents, discover all targets, and dynamically construct a reliable wireless communication network that can meet business needs, fully demonstrating the system's dynamic adaptability and end-to-end closed-loop control capabilities.

[0067] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features.

Claims

1. A multi-agent on-demand deployment and self-organizing wireless communication system based on embodied intelligence, characterized in that: The system comprises multiple mobile and communication-capable agents, a target user, and a starting base station. Agents, starting from the starting base station, make independent decisions based on local observations, forming a decentralized control strategy. Decisions made by agents based on current local perception information include action decisions and deployment decisions. If a deployment request is initiated, it is transmitted back to the starting base station via a tree structure, which then decides whether to respond to the request. Agents can make deployment request decisions during both the exploration phase and the phase after the target has been found. The system maintains a physical communication graph and a logical tree topology to facilitate communication connections between agents, data transmission, and deployment responses. The process forms an end-to-end closed-loop control. When an agent discovers a target user, it determines whether direct access is possible based on the target user's link capacity, latency, and load requirements. If the current network cannot meet the requirements, the agent waits for other cooperating agents or initiates a request to add more agents. If a request to add more agents is initiated, it is transmitted back to the originating base station via a tree structure. The originating base station then decides whether to respond to the request. The base station's policy center decides whether to respond to the addition and the addition location based on the global topology and communication indicators. The added agent is added to the physical communication graph and forms a parent node relationship in the logical tree topology to ensure the continuity of the backhaul path and the controllability of network management.

2. The multi-agent on-demand deployment and self-organizing wireless communication system based on embodied intelligence according to claim 1, characterized in that: Each agent can only move up, down, left, or right or remain stationary at each time step. During movement, the agent maintains a direct or indirect communication connection with the base station and decides whether to initiate a dispatch request based on local observations and neighbor information. It also recursively maintains the topology to ensure that dispatch requests and service data can be transmitted back to the starting base station.

3. The multi-agent on-demand deployment and self-organizing wireless communication system based on embodied intelligence according to claim 1, characterized in that: Between intelligent agents, and between an intelligent agent and the starting or target user, if the Manhattan distance is less than or equal to a preset threshold. If so, it is considered that a valid communication link exists.

4. A method for on-demand deployment and self-organizing wireless communication of multiple agents based on embodied intelligence, based on the on-demand deployment and self-organizing wireless communication system of multiple agents based on embodied intelligence as described in any one of claims 1-3, characterized in that: The specific steps are as follows: Step 1, Agent Deployment and Initialization; Deploy the initial agent at the base station at the starting point and complete the environment and parameter initialization when the system starts; Establish the physical communication graph and logical tree topology; Step 2, autonomous exploration and local decision-making by the agent; the agent's decision-making is based on a local observation-driven Actor-Critic reinforcement learning architecture; Step 3, Topology Maintenance and Link Monitoring: During the exploration process, maintain the physical communication graph and logical tree topology, and dynamically update link status and parent node selection. Step 4, Target Discovery and Access Judgment: Determine whether the current network meets its service requirements. If it does, access is granted; otherwise, wait for cooperation or initiate an additional dispatch request. If an additional dispatch request is issued, it is transmitted back to the starting base station via the tree structure, and the starting base station decides whether to respond to the additional dispatch request. Step 5: Reporting additional requests and making deployment decisions; The request to increase the number of dispatchers is reported up the logic tree to the starting base station, which then decides whether to increase the number of dispatchers and the location of the dispatch. Step 6: Target user data backhaul and path selection; the new intelligent agent is deployed and connected to the network, and the topology and routing information are updated; the target user service data is backhauled to the starting base station along the logic tree to achieve end-to-end data transmission.

5. The method for on-demand deployment and self-organizing wireless communication of multi-agent systems based on embodied intelligence according to claim 4, characterized in that: Step 1 specifically includes the following steps: Step 11: In the non-restricted example, the environmental boundary can be set as a discrete grid, and a preset Manhattan distance threshold is used to determine communication reachability; Step 12: Establish two types of topologies: one is a physical communication graph that reflects physical connectivity. The first is used to determine the existence of a link based on the Manhattan distance threshold, and to carry link capacity, latency, and load metrics in real time; the second is a logical tree topology used to manage request return and hierarchical scheduling. With the starting base station as the root, it ensures that each intelligent agent has a unique backhaul path at any time; Step 13: Each agent is assigned a unique identifier, an initial parent node reference, a neighbor list, and a local link state table. The built-in perception module establishes local perception capabilities for neighboring agents, environmental boundaries, and the starting point. To avoid backhaul interruptions, the system enables heartbeat and link detection mechanisms during the initialization phase and sets path integrity and connectivity thresholds.

6. The multi-agent on-demand deployment and self-organizing wireless communication method based on embodied intelligence according to claim 4, characterized in that: Step 2 specifically includes: Step 21: The agent starts from the starting base station and moves on the grid with a discrete step size; each step is only allowed to perform one of the actions "up, down, left, right or stay still"; before each action, the agent makes a feasibility judgment based on local observation and neighbor state: if the expected action will cause the backhaul path between itself or its subtree and the base station to be interrupted, the action is rejected and replaced by an alternative action. Step 22: The agent determines whether to make an additional dispatch request based on the local link quality and spatial reachability; all decisions are made under the premise of decentralization.

7. The method for on-demand deployment and self-organizing wireless communication of multi-agent systems based on embodied intelligence according to claim 4, characterized in that: Step 3 specifically includes: Step 31: Continuously analyze the physical communication graph With logical tree topology Perform consistency maintenance; physical communication graph According to the Manhattan distance threshold The rules recalculate adjacency relationships at each step and update the capacity, latency, and load metrics of each link; logical tree Maintain the "single parent, multiple children" hierarchical constraint to avoid forming loops; Step 32: Each agent immediately broadcasts the minimum necessary link state changes after moving, and refreshes the neighbor list and parent node selection; the selection of the parent node is based on a comprehensive judgment with priority order of "maintaining connectivity, minimum number of hops, sufficient remaining capacity, and low end-to-end latency"; when a parent node link degradation or loss of connection is detected, the child node triggers a fast reselection of the parent node process, and reconnects to the best among the reachable neighbors to ensure the continuity of the backhaul path; the topology and link state are recursively backhauled to the starting base station from bottom to top in a tree structure.

8. The method for on-demand deployment and self-organizing wireless communication of multi-agent systems based on embodied intelligence according to claim 4, characterized in that: Step 4 specifically includes: when the Manhattan distance between a certain intelligent agent and the target user is not greater than a threshold. When the condition is met, it is determined that "the target has been discovered"; if the determination is satisfied, the agent establishes an access relationship with the target user and attaches the target as a leaf node to the logic tree. On the corresponding branch; if not satisfied, the agent enters one of two types of waiting logic: one is to wait for a cooperating agent to approach to enhance the local network capability; the other is to initiate a dispatch request to supplement a new agent near the target or at a key relay location to meet the service quality requirements of the target access.

9. A method for on-demand deployment and self-organizing wireless communication of multiple agents based on embodied intelligence according to claim 4, characterized in that: Step 5 specifically includes: Step 51: After the dispatch request is formed, the information is transmitted along the logic tree. The request message is transmitted from bottom to top to the base station policy center. It includes: the trigger node identifier and location, a summary of the parent-child relationship of the node, statistics of the surrounding physical link quality, candidate relay location suggestions and their reachability verification. Step 52: At the starting base station, a deployment strategy module is set up. Based on the received deployment request and its accompanying node information, a comprehensive evaluation is performed in the order of "coverage gain - connectivity constraints - link quality - resource overhead" to determine whether to respond to the deployment and the specific deployment location. The selected deployment location must meet the following requirements: 1) The Manhattan distance to at least one node in the existing network does not exceed [the specified value]. 1) Ensures physical connectivity upon joining; 2) Significantly increases the target access capacity or reduces end-to-end latency after joining; 3) Does not disrupt the hierarchical stability of the logic tree; After the dispatch command is issued, the new agent joins immediately upon arriving at the designated location. and in The parent node relationship is established, and the routing table and return path are updated synchronously.

Images