Heterogeneous multi-agent adaptive collaborative planning method in communication limited environment

Abstract

The invention provides a heterogeneous multi-agent adaptive collaborative planning method in a communication limited environment, and relates to the technical field of artificial intelligence. If the preset communication condition is triggered, generating a cognitive abstract vector and sending the cognitive abstract vector to other agents; if cognitive abstract vectors sent by other agents are received, potential intentions in the cognitive abstract vectors are analyzed, a model is updated according to the intentions, and the potential intentions at the next moment are updated; if the confidence score is greater than a preset threshold, adding other agents into the current game participant set, and performing game strategy calculation again according to the potential intention at the next moment to obtain a new strategy; if a preset re-planning condition is triggered, re-planning tasks in the current local task queue based on the current game participant set according to the capability declaration mask to obtain a new task allocation scheme; and converting the new strategy into an action instruction and executing the action instruction. The communication dependence is effectively reduced, the intelligent agent capability difference is dynamically integrated, and the collaborative decision-making efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an adaptive cooperative planning method for heterogeneous multi-agent systems in communication-constrained environments. Background Technology

[0002] In existing technologies, multi-agent collaborative decision-making mainly adopts two types of schemes: centralized and distributed.

[0003] In centralized schemes, data from all agents is uploaded to a central control node, which then calculates the optimal strategy and issues commands. While centralized methods can achieve globally optimal performance, they suffer from significant bottlenecks in environments with limited communication or frequent topology changes. Centralized methods are highly dependent on communication and computing resources; interruptions in communication links or increased latency will lead to global scheduling failure.

[0004] In distributed schemes, agents interact with their neighbors through local information to make decisions and coordinate. Although distributed reinforcement learning methods are scalable, they struggle to maintain policy consistency in heterogeneous agent scenarios and remain heavily reliant on neighbor information. Summary of the Invention

[0005] The problem that this invention aims to solve is that existing technologies are highly dependent on resources and it is difficult to maintain policy consistency in heterogeneous intelligent agent scenarios.

[0006] To address the aforementioned problems, in a first aspect, this invention provides an adaptive cooperative planning method for heterogeneous multi-agent systems in communication-constrained environments, comprising: If a preset communication condition is triggered, a cognitive summary vector is generated and sent to other intelligent agents. Each intelligent agent is equipped with at least one of the following modules: perception module, memory module, reasoning module, game decision-making module, and execution module. The cognitive summary vector includes the task identifier, potential intention, confidence score, and capability declaration mask of the current intelligent agent. The capability declaration mask indicates the type of module set in the current intelligent agent. If a cognitive summary vector is received from another agent, the latent intent in the cognitive summary vector is parsed, the model is updated according to the intent, and the updated latent intent for the next time step is obtained. If the confidence score is greater than the preset threshold, other agents are added to the current set of game participants, and the game strategy is recalculated based on the updated potential intentions for the next moment to obtain a new strategy. If the preset replanning condition is triggered, the tasks in the current local task queue will be replanned based on the capability declaration mask and the current set of game participants to obtain a new task allocation scheme and a new task queue. If an execution module is present, the execution module will convert the new strategy into action instructions and execute the action instructions, and save the execution results to the memory module.

[0007] Optionally, the preset communication conditions include the agent entering a preset communication area, the communication time interval reaching a preset duration threshold, or the detection of a high uncertainty event.

[0008] Optionally, the sensing module is used to receive raw observation data and extract feature vectors; The memory module is used to maintain a long-term memory bank, store historical data, and control the memory capacity using a sliding window mechanism. The reasoning module is used to infer the hidden state of the current environment and the potential intentions of other agents based on the current feature vector and historical data. The game decision-making module is used to construct a set of local participants, and based on the set of local participants, solve the optimization model to obtain the optimal strategy; The execution module is used to convert the optimal strategy output into action instructions.

[0009] Optionally, the optimization model is: in, Indicates the optimal strategy; This represents the current policy of the agent. This indicates the current agent's current strategy. Send the corresponding action command. Indicates the current policy of other agents. Indicates other agents in the current policy The corresponding action instruction is given below, where 's' indicates the current hidden state of the environment. Represents the reward function, For policy regularization coefficients, For Kullback-Leibler divergence, To pre-set a priori strategy, This represents the policy corresponding to the current agent when executing action instructions. This represents the preset prior policy corresponding to the current agent when executing action instructions.

[0010] Optionally, the reward function is in, This indicates the reward for completing the task. This indicates a penalty for energy consumption. Indicates a reward for cooperation and collaboration. , and These represent the weights of task completion rewards, energy consumption penalties, and cooperation rewards, respectively.

[0011] Optionally, the intention update model is in, This represents the latent intention updated by the k-th agent at time t+1. Indicates the trust decay factor. This represents the potential intention of the k-th agent at time t. Let represent the cognitive summary vector of the k-th agent, and decode represent the decoding operation.

[0012] Optionally, the preset replanning conditions include any one of the following: the policy difference between the new policy and the preset prior policy is greater than or equal to a preset change threshold, the current agent's own state is abnormal, and the current agent's task conflicts with the tasks of other agents. The tasks of other agents are the tasks of high-confidence agents parsed from the task identifiers of the received cognitive summary vector.

[0013] Secondly, the present invention also provides an intelligent agent, comprising: The summary vector generation module is used to generate a cognitive summary vector and send it to other intelligent agents if a preset communication condition is triggered. Each intelligent agent is equipped with at least one of the following modules: perception module, memory module, reasoning module, game decision-making module, and execution module. The cognitive summary vector includes the task identifier, potential intention, confidence score, and capability declaration mask of the current intelligent agent. The capability declaration mask indicates the type of module set in the current intelligent agent. The intent update module is used to parse the latent intent in the cognitive summary vector sent by other agents, and obtain the updated latent intent for the next time step according to the intent update model. The game decision-making module is used to add other agents to the current game participant set if the confidence score is greater than a preset threshold, and recalculate the game strategy based on the updated potential intentions for the next moment to obtain a new strategy. The task allocation module is used to re-plan the tasks in the current local task queue based on the capability declaration mask and the current set of game participants if a preset replanning condition is triggered, so as to obtain a new task allocation scheme and a new task queue. If an execution module is present, the execution module will convert the new strategy into action instructions and execute the action instructions, and save the execution results to the memory module.

[0014] This invention provides an adaptive cooperative planning method for heterogeneous multi-agent systems in communication-constrained environments. Compared with existing technologies, it has the following advantages: By employing condition-triggered communication, generating cognitive summary vectors for efficient information exchange, dynamically adjusting policies based on intent update models, and utilizing capability declaration masks for adaptive task replanning, this approach addresses the issues of low efficiency and insufficient robustness in collaborative decision-making under communication constraints and heterogeneous agent environments. It offers advantages such as effectively reducing communication dependencies, dynamically integrating agent capability differences, improving collaborative decision-making efficiency, and enhancing system robustness in communication-constrained and heterogeneous agent environments. This provides a robust and efficient collaborative planning mechanism for heterogeneous multi-agent systems in communication-constrained environments. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating an adaptive cooperative planning method for heterogeneous multi-agent systems in a communication-constrained environment, provided by an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an intelligent agent provided in an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application are described clearly and completely. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0019] like Figure 1 As shown in the embodiment of this application, an adaptive cooperative planning method for heterogeneous multi-agent systems in a communication-constrained environment is provided, comprising: S1: If a preset communication condition is triggered, a cognitive summary vector is generated and sent to other intelligent agents. Each intelligent agent is equipped with at least one of the following modules: perception module, memory module, reasoning module, game decision-making module, and execution module. The cognitive summary vector includes the task identifier, potential intention, confidence score, and capability declaration mask of the current intelligent agent. The capability declaration mask indicates the type of module set in the current intelligent agent. S2: If a cognitive summary vector is received from another agent, the latent intent in the cognitive summary vector is parsed, the model is updated according to the intent, and the updated latent intent for the next time step is obtained. S3: If the confidence score is greater than the preset threshold, other agents are added to the current set of game participants. Based on the updated potential intentions for the next moment, the game strategy is recalculated to obtain a new strategy. S4: If the preset replanning condition is triggered, the tasks in the current local task queue are replanned based on the capability declaration mask and the current set of game participants to obtain a new task allocation scheme and a new task queue.

[0020] S5: If there is an execution module, the execution module will convert the new strategy into action instructions and execute the action instructions, and save the execution results to the memory module.

[0021] In this optional embodiment, the exchange of simplified cognitive summary vectors between agents avoids the transmission of large amounts of raw data, significantly reducing the requirements for communication bandwidth and stability. Through capability declaration masks and preset replanning conditions, the challenges of heterogeneous agents in task allocation and responding to unforeseen events are effectively addressed. The dynamic task reassignment mechanism based on capability matching enables the system to fully utilize the different capabilities of heterogeneous agents, flexibly respond to environmental changes and task anomalies, and significantly improve the adaptability and task completion rate of the multi-agent system in complex, dynamic, and heterogeneous environments.

[0022] By employing condition-triggered communication, generating cognitive summary vectors for efficient information exchange, dynamically adjusting policies based on intent update models, and utilizing capability declaration masks for adaptive task replanning, this approach addresses the issues of low efficiency and insufficient robustness in collaborative decision-making under communication constraints and heterogeneous agent environments. It offers advantages such as effectively reducing communication dependencies, dynamically integrating agent capability differences, improving collaborative decision-making efficiency, and enhancing system robustness in communication-constrained and heterogeneous agent environments. This provides a robust and efficient collaborative planning mechanism for heterogeneous multi-agent systems in communication-constrained environments.

[0023] The following is a detailed description of each step.

[0024] To address the differences in computing resources, sensor configurations, and software capabilities among agents in a heterogeneous multi-agent system, a customizable multi-brain-region cognitive architecture is deployed for each heterogeneous agent i (i=1,2,…,N). The complete brain-like cognitive architecture comprises five functional sub-modules: a perception module, a memory module, a reasoning module, a game-theoretic decision-making module, and an execution module. The perception module receives raw observation data and extracts feature vectors. The memory module maintains a long-term memory bank, stores historical data, and uses a sliding window mechanism to control memory capacity. The reasoning module infers the current environmental state and the potential intentions of other agents based on the current feature vectors and historical data. The game-theoretic decision-making module constructs a set of local participants and solves an optimization model based on this set to obtain the optimal strategy. The execution module converts the optimal strategy output into action commands.

[0025] The perception module receives raw observations from sensors or environmental interfaces and extracts feature vectors. This perception module can be a sensor data processing unit, for example, using image recognition algorithms, radar signal processing algorithms, or sonar data analysis algorithms to transform raw, unprocessed sensor inputs (such as images, sound, and distance data) into structured feature representations that can be used by subsequent modules.

[0026] The memory module maintains a long-term memory bank to store historical states, interaction records, and task contexts, and uses a sliding window mechanism to control the memory capacity.

[0027] The inference module, based on current feature vectors and historical data from a memory bank, infers the hidden state of the environment and the potential intentions of other agents through causal models or Bayesian networks. Specifically, this module can construct and continuously update an internal model based on a probabilistic graphical model, using perceived features and historical memory (historical data) as observational evidence. It then uses variational inference or belief propagation algorithms to calculate the most likely explanation for the currently observed environmental state and the potential intentions of other agents. Essentially, its analysis process transforms uncertain, partial observations, combined with contextual memory, into a structured, probabilistic belief estimate of the global situation. The inference module can also be an inference engine based on machine learning models (such as neural networks or Bayesian networks), analyzing input feature vectors and historical data through a trained model to predict unobserved environmental states or infer behavioral patterns of other agents. Furthermore, the inference module can also be a system based on logical reasoning or a rule engine, performing state estimation and intention recognition based on preset logical rules and a knowledge base, combined with currently perceived information and historical data in memory.

[0028] The game decision-making module constructs a set of local participants. Assume that its policy follows a parameterized distribution. , In the current strategy The following intention The corresponding actions are determined, and the following optimization problem is solved to generate the policy. The optimized model is in, Indicates the optimal strategy; This represents the current policy of the agent. This indicates the current agent's current strategy. Send the corresponding action command. Indicates the current policy of other agents. Indicates other agents in the current policy The corresponding action instruction is given below, where 's' indicates the current hidden state of the environment. Represents the reward function, For policy regularization coefficients, This is the Kullback-Leibler divergence. Additionally, Let be the individual reward function of agent i, reflecting the action instruction it takes in the current state s. And other intelligent agents take action The benefits at any given time can be designed by comprehensively considering factors such as task completion, energy consumption, and synergistic gains. For example: ,in Rewards for completing the task. As a penalty for energy consumption, The rewards for collaboration and cooperation can be specifically set according to the application scenario. , and These represent the weights of task completion rewards, energy consumption penalties, and cooperation rewards, respectively. The regularization coefficient is used to balance reward maximization and policy stability. Pre-defined prior strategies, such as safety and conservative strategies, are used to constrain the scope of strategy updates. Kullback-Leibler divergence measures the difference between the current policy and the prior policy. It is used to suppress policy mutations and enhance robustness. The calculation formula is: in, This represents the policy corresponding to the current agent when executing action instructions. This represents the preset prior policy corresponding to the current agent when executing action instructions.

[0029] When the current agent needs to recalculate its game strategy, the game decision-making module uses this optimization model to find the optimal strategy. The core of this model is to maximize a comprehensive objective function. This objective function not only considers the reward function the agent can obtain by executing action instruction a in the current hidden state s, but also uses Kullback-Leibler divergence to measure the difference between the current strategy and the preset prior strategy. During the optimization process, the model considers the current agent's current strategy and its corresponding action instruction, as well as estimates of the current strategies and their corresponding action instructions of other agents. Through iterative optimization, the agent can find an optimal strategy that maximizes its own and collaborative rewards while avoiding drastic policy fluctuations and maintaining a certain degree of stability. This mechanism enables agents to perform collaborative planning more effectively when facing limited communication and environmental uncertainty, avoiding inefficient or conflicting behaviors caused by improper strategy calculation. This optimization model provides a quantitative and operable policy generation framework for the game decision-making module, ensuring that in the process of multi-agent collaboration, policy generation is not only based on immediate rewards, but also comprehensively considers long-term goals, policy stability and prediction of the behavior of other agents, thereby improving the collaborative efficiency and robustness of the entire system.

[0030] The execution module is responsible for translating the strategy output into specific action commands and monitoring the task execution status, such as whether subtasks have been completed or whether any failures have occurred. This execution module can be an action command generator, transforming abstract strategies (such as "move towards the target") into concrete control signals (such as motor speed and servo angle). Alternatively, the execution module can be a drive unit that directly interfaces with the underlying actuators (such as robot joint controllers or UAV flight control systems), responsible for transmitting action commands to the physical actuators and monitoring the execution process.

[0031] Each heterogeneous agent matches five functional sub-modules based on its own hardware structure and software architecture characteristics. During adaptive game decision-making and adaptive planning, it is not necessary to possess all five modules; existing modules can be used for cooperative matching. During system startup or task initialization, each agent declares its set of implemented functional modules. This declaration information can be embedded in subsequent communication summaries for other agents to perform capability matching and collaborative modeling.

[0032] The agent's internal functions are clearly divided into modules such as perception, memory, reasoning, game-theoretic decision-making, and execution, providing solid internal support for adaptive collaborative planning methods for heterogeneous multi-agent systems in communication-constrained environments. This modular design enables the agent to efficiently and accurately process raw observation data, manage historical information, infer environmental states and the intentions of others, formulate optimal strategies, and execute actions. This significantly improves the efficiency and robustness of the agent in information processing, decision-making, and task execution in complex dynamic environments, effectively solving the accuracy and real-time problems of collaborative planning that may arise in the absence of explicit internal mechanisms, and ensuring the effective implementation and adaptability of the entire collaborative planning method.

[0033] S1: If a preset communication condition is triggered, a cognitive summary vector is generated and sent to other intelligent agents. Each intelligent agent has at least one of the following modules: perception module, memory module, reasoning module, game decision-making module, and execution module. The cognitive summary vector includes the task identifier, potential intention, confidence score, and capability declaration mask of the current intelligent agent. The capability declaration mask indicates the type of module set in the current intelligent agent.

[0034] Specifically, the preset communication conditions include an agent entering a preset communication area, a communication time interval reaching a preset duration threshold, or the detection of a high-uncertainty event. For example, the agent can be configured to communicate once every fixed time period to ensure periodic information updates. Alternatively, communication can be triggered when the agent completes a specific task to inform other agents of the change in its task status. Furthermore, communication can be triggered when the agent detects a significant change in its own state, such as an energy level falling below a certain threshold.

[0035] First, preset communication conditions are set, such as entering a preset communication area, a time interval reaching a threshold, or detecting a high-uncertainty event. When the conditions are met, agent i generates a structured, fixed-length cognitive summary vector. Its fields strictly correspond to the five standard modules, but only the outputs of the implemented modules are filled in. Fields corresponding to unimplemented modules are set to predefined null values, such as all-zero vectors, special flag bits, or NaN encoding. The specific field structure can be referenced as follows (arranged in a fixed order): 1. Task Identifier: Indicates the subtask being executed, originating from the game decision-making module or the execution module; 2. Perceptual features: If they exist, they are composed of dimensionality-reduced feature vectors and come from the perceptual module; 3. Latent Intent: If it exists, it consists of compressed latent intents, derived from the perception module; 4. Attention Focus Coordinates: A vector collaboratively generated by the inference, game-theoretic decision-making, and execution modules, typically represented as the center (x, y, z) of the target region or an identifier of a key entity. When generating these coordinates, the inference module first infers the "key region" based on the environmental state; the game-theoretic decision-making module calculates the "strategy center of gravity" based on the game equilibrium solution; and the execution module provides feedback on the "intervention point" based on the current task execution status. The system weights and fuses these three outputs or takes the union of their values ​​to ultimately determine the coordinates that best represent the current decision focus.

[0036] 5. Confidence score This reflects the reliability of the current cognitive state and is comprehensively evaluated by available modules. It is a scalar value derived from the comprehensive evaluation of all available modules, reflecting the reliability of the current overall cognitive state. It is a function of multiple indicators, such as: feature reconstruction error of the perception module, intention prediction entropy of the inference module, policy distribution entropy of the game decision-making module, and historical task success rate of the execution module. During scoring, these uncertainty indicators can be integrated into a confidence score between 0 and 1 using a lightweight scoring network or a weighted average formula.

[0037] 6. Capability declaration mask A binary vector indicating the presence of each module in a module set. It is an N-bit binary vector (e.g., a 5-bit vector for 5 modules), with each bit corresponding to a functional module in sequence, such as... This can represent the presence of a perception module, a reasoning module, and an execution module. The value of each bit (1 or 0) directly declares whether the agent has implemented the corresponding module. During parsing, the receiver can accurately determine the capability boundaries of the sender j based on the mask. For example, if the parsing reveals that the sender j's capability mask is [1,0,0,0,1], it indicates that the sender j only possesses "perception" and "execution" capabilities, and lacks complex "memory," "reasoning," and "game-playing" capabilities. Based on this, the receiver i can dynamically adjust its modeling method for j's behavior—no longer attempting complex intention reasoning for j, but instead treating it as an agent following fixed instructions or simple reflection strategies, thereby simplifying its own decision-making model and achieving more accurate and efficient collaboration.

[0038] The cognitive summary vector is broadcast or forwarded to neighboring agents, thereby spreading throughout the communication link network.

[0039] S2: If a cognitive summary vector is received from another agent, the potential intent in the cognitive summary vector is parsed, the model is updated according to the intent, and the updated potential intent for the next time step is obtained.

[0040] Specifically, parsing potential intentions can be accomplished by consulting a predefined intention mapping table, which encodes the received intentions into an intention representation that the agent can understand.

[0041] After receiving the cognitive summary vector from another agent k, agent i first needs to parse the intent embedding and confidence in the cognitive summary vector, and update the intent parameters for the next time step. The intent update model is as follows: in, This represents the latent intention updated by the k-th agent at time t+1. Indicates the trust decay factor. This represents the potential intention of the k-th agent at time t. Let represent the cognitive summary vector of the k-th agent, and decode represent the decoding operation.

[0042] A trust decay factor is used to adjust the retention rate of historical potential intentions, while newly decoded intentions are used to correct or update this historical understanding. This mechanism allows agents to progressively update their understanding of other agents' intentions, avoiding drastic fluctuations caused by relying solely on the latest information or lags caused by over-reliance on historical information. In this way, even under conditions of communication uncertainty or incomplete information, agents can maintain relatively stable and accurate estimates of other agents' intentions, thus providing more reliable input for subsequent game strategy calculations and improving the robustness and effectiveness of overall collaborative planning.

[0043] S3: If the confidence score is greater than a preset threshold, other agents are added to the current set of game participants. Based on the updated potential intentions for the next moment, the game decision-making module recalculates the game strategy to obtain a new strategy. The formation of the set of game participants enables the current agent to identify partners with whom it can cooperate.

[0044] Specifically, if the confidence score exceeds a preset threshold, i.e. Then add agent k to the current set of game participants. The updated latent intentions are used to reconstruct the optimization model. Based on the updated latent intention parameters of other agents in the next time step, the game optimization problem is solved again to obtain a new strategy.

[0045] S4: If the preset replanning condition is triggered, then based on the capability declaration mask, the tasks in the current local task queue are replanned according to the current set of game participants to obtain a new task allocation scheme and a new task queue. Otherwise, the local task queue is not replanned.

[0046] Specifically, the preset replanning conditions include any one of the following: the policy difference between the new policy and the preset prior policy is greater than or equal to a preset change threshold; the current agent's state is abnormal; and the current agent's task conflicts with the tasks of other agents. The tasks of other agents are the tasks of high-confidence agents parsed from the task identifiers of the received cognitive summary vector. If the policy change (i.e., the Kullback-Leibler divergence) exceeds the preset change threshold, it is marked as a "policy mutation," triggering task replanning. The policy change can be calculated using the Kullback-Leibler divergence mentioned in the game decision module.

[0047] A local task queue is maintained, initially allocated offline by the global task decomposer or obtained through initial communication negotiation. The global task decomposer is a centralized computing module located at a ground station, command center, or a leader agent. It receives top-level tasks and, based on a pre-input environment model, estimated agent capabilities, and optimization objectives, automatically runs a task decomposition algorithm such as mixed-integer programming to generate a preliminary, non-final task allocation scheme. "Offline" refers to pre-planning performed before system startup or task commencement. This scheme serves as a "seed plan" for the initial actions of each agent and is subsequently dynamically adjusted by an event-driven mechanism within pre-defined replanning conditions during operation. Obtaining tasks through initial communication negotiation means that, without pre-planning or after communication is restored, the agent group exchanges their respective capability declaration masks through an initial broadcast communication and, based on this, runs a distributed consensus protocol (such as a voting-based auction algorithm) to negotiate the initial task allocation.

[0048] During the task, the local task queue maintains its current task queue plan unless a policy mutation occurs. Replanning is initiated when any of the following events occur (i.e., any of the preset replanning conditions are met): abnormal state of the task queue itself, such as battery level <20% or sensor failure; receiving a high-confidence external digest that causes a task conflict, such as two agents simultaneously committing to the same subtask; or a policy mutation indicating that the original plan is not feasible.

[0049] The replanning approach considers modeling in the form of a constraint satisfaction problem: the variables are subtask allocations, the domain is the set of available agents, and constraints include capability matching, spatiotemporal conflict avoidance, and resource limits. The solver employs a lightweight backtracking search or a heuristic greedy algorithm to output a new task allocation scheme and update the task queue. Specifically, the lightweight backtracking search refers to a search algorithm that has undergone key optimizations on the standard backtracking framework to control computational complexity: it adopts a most-constrained variable-first heuristic, prioritizing the allocation of subtasks with the fewest available agents and the smallest domain to trigger failures early and reduce search branches; simultaneously, in the forward check, only first-order constraint propagation is performed, that is, only constraints directly related to the currently assigned variables are checked, achieving a balance between solution speed and pruning effect. Furthermore, the heuristic greedy algorithm employs a minimum-cost incremental allocation strategy, without backtracking, but instead traversing all unassigned subtasks, selecting the agent that can be allocated with the minimum "cost" at each step, and confirming the allocation through task-agent pairs. The “cost” here can comprehensively measure factors such as task distance, resource consumption, and the degree of conflict with other agents’ plans, and can be calculated in real time using a simple linear weighting function.

[0050] By introducing policy differences between the new strategy and the preset prior strategy, abnormal states of the current agent, and conflicts between the current agent's task and other agents' tasks as preset replanning conditions, the scheme in this application enables refined and adaptive control of the multi-agent collaborative planning process. This allows the agent system to identify critical moments requiring task adjustments in a timely and accurate manner, avoiding unnecessary resource waste caused by replanning. It also ensures rapid response and strategy adjustment in critical situations such as drastic environmental changes, agent malfunctions, or task allocation conflicts. In particular, by considering only high-confidence tasks of other agents to determine conflicts, misjudgments and ineffective replanning due to uncertain information are effectively avoided, significantly improving the robustness, adaptability, and task completion efficiency of the multi-agent system in communication-constrained environments.

[0051] S5: If there is an execution module, the execution module will convert the new strategy into action instructions and execute the action instructions, and save the execution results to the memory module.

[0052] Specifically, the execution module executes action instructions and writes the results (including success / failure, time consumption, resource consumption, etc.) into the memory bank of the memory module for subsequent reasoning and policy prior updates. The entire process is executed cyclically with a fixed control cycle, forming a closed loop of "perception, reasoning, game theory, communication, planning, execution, and optimization".

[0053] In summary, compared with existing technologies, it has the following beneficial effects: 1. In communication-constrained environments, this method only requires exchanging semantic summaries of 5%–10% of the original data volume to maintain effective collaboration, reduce communication overhead, enhance reliability, and is suitable for extreme scenarios such as deep sea, disaster areas, and battlefields.

[0054] 2. Traditional decentralized reinforcement learning methods are prone to falling into suboptimal strategies due to a lack of equilibrium guarantees. By employing explicit game theory modeling and inference of the opponent's intentions, the success rate of tasks in mixed-motivation tasks, such as some goal-conflict tasks, is improved.

[0055] 3. The multi-brain region cognitive architecture enables single agents to have stronger contextual understanding and long-term planning capabilities. In dynamic environments, the response time for task replanning is reduced to 1 / 3 of that of traditional task allocation methods, and no central scheduling node is required, significantly enhancing fault tolerance.

[0056] like Figure 2 As shown in the embodiment of this application, an intelligent agent includes: The summary vector generation module 10 is used to generate a cognitive summary vector and send it to other intelligent agents if a preset communication condition is triggered. Each intelligent agent is equipped with at least one of the following modules: perception module 20, memory module 30, reasoning module 40, game decision module 50, and execution module 60. The cognitive summary vector includes the task identifier, potential intention, confidence score, and capability declaration mask of the current intelligent agent. The capability declaration mask indicates the types of modules set in the current intelligent agent. The intent update module 70 is used to, if it receives a cognitive summary vector sent by another agent, parse the potential intent in the cognitive summary vector, and obtain the updated potential intent for the next moment according to the intent update model; If the confidence score is greater than the preset threshold, the strategy update module 80 adds other agents to the current game participant set, and recalculates the game strategy based on the potential intention of the next time step to obtain a new strategy. The task allocation module 90 is used to re-plan the tasks in the current local task queue based on the capability declaration mask and the current set of game participants if a preset replanning condition is triggered, so as to obtain a new task allocation scheme and a new task queue.

[0057] If there is an execution module 60, the execution module 60 will convert the new strategy into action instructions and execute the action instructions, and save the execution result to the memory module 30.

[0058] Specifically, by combining cognitive summary vectors and intent update models in a dynamic collaborative manner, and by introducing a capability declaration mask mechanism, the system effectively reduces information transmission volume and maintains policy consistency in communication-constrained environments, thereby improving the collaborative efficiency and robustness of heterogeneous multi-agent systems. In communication-constrained environments, agents do not need to continuously exchange complete observation data; instead, they achieve collaboration through the exchange of high-level semantic information. This allows the system to maintain basic collaborative functions even when communication links are interrupted or latency increases. The capability declaration mask mechanism enables agents to identify the capability configurations of other agents. When a preset replanning condition is triggered, the system can dynamically adjust task allocation based on the capability declaration mask, assigning tasks to agents with the corresponding capabilities. This fully utilizes the diversity of heterogeneous agents and improves task completion rates.

[0059] Application Example 1: Consider three heterogeneous UAVs performing a forest fire monitoring mission: UAV-A (infrared thermal imaging, 30-minute endurance), UAV-B (visible light + lidar, high-speed maneuverability), and UAV-C (communication relay, no detection capability). The mission area is divided by mountains, with communication coverage only at the mountaintop. After system initialization, each UAV deploys the six-step process described above. During the non-communication phase, UAV-A detects an abnormal heat source. Its inference module, combining wind direction and historical fire patterns, infers the direction of fire spread and estimates that UAV-B may be approaching from the north. When UAV-A reaches the communication coverage area at the mountaintop, its communication module is activated and immediately triggers a communication event: generating a summary and broadcasting it. At this time, UAV-B, which is either within the communication coverage area or entering the area, successfully receives the summary within this communication window. In the next control cycle, UAV-B updates its belief in UAV-A's intention based on the received information, incorporates it into the game, reassesses the payoffs of "continuing patrol" versus "turning to support," and ultimately chooses to support. Meanwhile, as A's battery level dropped to 22%, its task planner automatically handed over subsequent monitoring to B and generated a return-to-home command. Only one summary was exchanged throughout the entire process; no images or tracks were transmitted.

[0060] Application Example 2: In an urban underground pipeline inspection scenario, five heterogeneous robots (including a pipe crawler, a flying drone, and a ground mobile platform) need to collaboratively locate leak points. Due to metal shielding, communication is only briefly restored at the inspection wellhead. Each robot runs the method of this invention: the crawler infers the leak range through pressure gradients, generating a summary containing "suspected range [320m, 350m]" and a confidence level of 0.88; after receiving the data at the wellhead, the drone verifies it using its own acoustic data and establishes a cooperative game with the crawler in its game theory module to decide whether to hover to assist in precise positioning; the task planner detects that the ground platform's battery is low and automatically transfers its task to the drone.

[0061] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0062] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A heterogeneous multi-agent adaptive cooperative planning method in a communication-constrained environment, characterized in that, include: If a preset communication condition is triggered, a cognitive summary vector is generated and sent to other intelligent agents. Each intelligent agent is equipped with at least one of the following modules: perception module, memory module, reasoning module, game decision-making module, and execution module. The cognitive summary vector includes the task identifier, potential intention, confidence score, and capability declaration mask of the current intelligent agent. The capability declaration mask indicates the type of module set in the current intelligent agent. If a cognitive summary vector is received from another agent, the latent intent in the cognitive summary vector is parsed, the model is updated according to the intent, and the updated latent intent for the next time step is obtained. If the confidence score is greater than the preset threshold, other agents are added to the current set of game participants, and the game strategy is recalculated based on the updated potential intentions for the next moment to obtain a new strategy. If the preset replanning condition is triggered, the tasks in the current local task queue will be replanned based on the capability declaration mask and the current set of game participants to obtain a new task allocation scheme and a new task queue. If an execution module is present, the execution module will convert the new strategy into action instructions and execute the action instructions, and save the execution results to the memory module.

2. The adaptive cooperative planning method for heterogeneous multi-agent systems in communication-constrained environments as described in claim 1, characterized in that, The preset communication conditions include the agent entering the preset communication area, the communication time interval reaching a preset duration threshold, or the detection of a high uncertainty event.

3. The adaptive cooperative planning method for heterogeneous multi-agent systems in a communication-constrained environment as described in claim 1, characterized in that, The sensing module is used to receive raw observation data and extract feature vectors; The memory module is used to maintain a long-term memory bank, store historical data, and control the memory capacity using a sliding window mechanism. The reasoning module is used to infer the hidden state of the current environment and the potential intentions of other agents based on the current feature vector and historical data. The game decision-making module is used to construct a set of local participants, and based on the set of local participants, solve the optimization model to obtain the optimal strategy; The execution module is used to convert the optimal strategy output into action instructions.

4. The adaptive cooperative planning method for heterogeneous multi-agent systems in a communication-constrained environment as described in claim 1, characterized in that, The optimization model is as follows: in, Indicates the optimal strategy; This represents the current policy of the agent. This indicates the current agent's current strategy. Send the corresponding action command. Indicates the current policy of other agents. Indicates other agents in the current policy The corresponding action instruction is given below, where 's' indicates the current hidden state of the environment. Represents the reward function, For policy regularization coefficients, For Kullback-Leibler divergence, To pre-set a priori strategy, This represents the policy corresponding to the current agent when executing action instructions. This represents the preset prior policy corresponding to the current agent when executing action instructions.

5. The adaptive cooperative planning method for heterogeneous multi-agent systems in a communication-constrained environment as described in claim 4, characterized in that, The reward function is: in, This indicates the reward for completing the task. This indicates a penalty for energy consumption. Indicates a reward for cooperation and collaboration. , and These represent the weights of task completion rewards, energy consumption penalties, and cooperation rewards, respectively.

6. The adaptive cooperative planning method for heterogeneous multi-agent systems in a communication-constrained environment as described in claim 1, characterized in that, The intent update model is as follows: in, This represents the latent intention updated by the k-th agent at time t+1. Indicates the trust decay factor. This represents the potential intention of the k-th agent at time t. Let represent the cognitive summary vector of the k-th agent, and decode represent the decoding operation.

7. The adaptive cooperative planning method for heterogeneous multi-agent systems in a communication-constrained environment as described in claim 1, characterized in that, The preset replanning conditions include any one of the following: the policy difference between the new policy and the preset prior policy is greater than or equal to a preset change threshold, the current agent's own state is abnormal, and the current agent's task conflicts with the tasks of other agents. The tasks of other agents are the tasks of high-confidence agents parsed from the task identifiers of the received cognitive summary vector.

8. An intelligent agent, characterized in that, include: The summary vector generation module is used to generate a cognitive summary vector and send it to other intelligent agents if a preset communication condition is triggered. Each intelligent agent is equipped with at least one of the following modules: perception module, memory module, reasoning module, game decision-making module, and execution module. The cognitive summary vector includes the task identifier, potential intention, confidence score, and capability declaration mask of the current intelligent agent. The capability declaration mask indicates the type of module set in the current intelligent agent. The intent update module is used to parse the latent intent in the cognitive summary vector sent by other agents, and obtain the updated latent intent for the next time step according to the intent update model. The game decision-making module is used to add other agents to the current game participant set if the confidence score is greater than a preset threshold, and recalculate the game strategy based on the updated potential intentions for the next moment to obtain a new strategy. The task allocation module is used to re-plan the tasks in the current local task queue based on the capability declaration mask and the current set of game participants if a preset replanning condition is triggered, so as to obtain a new task allocation scheme and a new task queue. If an execution module is present, the execution module will convert the new strategy into action instructions and execute the action instructions, and save the execution results to the memory module.

Images