Body multi-agent self-adaptive cooperation method and system based on large language model

Abstract

The invention discloses a multi-agent self-adaptive cooperation method and system based on a large language model, and belongs to the technical field of artificial intelligence, and the method comprises the steps: obtaining task description and environment information, and inputting the task description and environment information into the large language model; generating a personalized initial instruction corresponding to the multiple agents; constructing a multi-agent topological state space, and establishing a relation topological structure among the agents; based on a distributed negotiation mechanism, determining an action set through natural language interaction; generating a self-adaptive cooperation strategy according to the action set and the current environment state; the multiple agents execute corresponding actions and obtain feedback, and a cooperation strategy is dynamically adjusted; according to the method, the semantic understanding capability of a large language model is innovatively combined with the technologies of topological state mapping, distributed negotiation and the like, seamless conversion from high-level semantic understanding to specific execution actions is achieved, and the method has the advantages of being high in practicability and high in practicability. And the environmental adaptability, the cooperation efficiency, the safety and the reliability of the multi-agent system are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a somatic multi-agent adaptive collaboration method and system based on a large language model, and more particularly to a method for realizing dynamic collaboration and adaptive decision-making of multi-agents in complex environments using a large language model. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, somatic agent systems have been widely applied in intelligent manufacturing, warehouse logistics, emergency rescue and other fields. Existing multi-agent collaboration systems are mainly based on preset rules or simple reinforcement learning methods, which have obvious limitations in dealing with complex, dynamic and uncertain environments.

[0003] Traditional multi-agent collaboration methods usually adopt a central control mode or a simple distributed protocol, which is difficult to adapt to dynamic changes in the environment. Rule-based methods lack flexibility and cannot handle unexpected situations; while traditional reinforcement learning-based methods are difficult to achieve advanced semantic understanding and long-term planning. In addition, existing methods also have deficiencies in the consistency of negotiation and decision-making among agents, often leading to low collaboration efficiency or conflict behavior.

[0004] In recent years, large language models (LLM) have made breakthroughs in natural language understanding and generation, demonstrating strong contextual understanding and task planning capabilities. However, how to effectively combine the semantic understanding capabilities of large language models with the physical interaction capabilities of multi-agent systems to achieve efficient and adaptive collaboration remains a technical challenge that needs to be addressed. SUMMARY

[0005] The purpose of the present application is to provide a somatic multi-agent adaptive collaboration method and system based on a large language model, aiming to solve the problems of insufficient flexibility, poor environmental adaptability, and difficulty in ensuring collaboration decision-making consistency in existing multi-agent collaboration systems.

[0006] The present application proposes a somatic multi-agent adaptive collaboration method based on a large language model, comprising: Obtaining task description and environment information, and inputting the task description and environment information into a large language model; Based on the large language model, generating individualized initial instructions for each multi-agent, wherein the individualized initial instructions include task objectives, initial environment states, and agent role definitions; Constructing a multi-agent topological state space, mapping the state information of each agent to the topological state space, and establishing a relationship topology structure between agents; Based on a distributed negotiation mechanism, the multiple agents jointly determine a set of actions through natural language interaction, wherein the set of actions is used to guide the collaborative behavior of each agent. Based on the action set and the current environment state, an adaptive cooperation strategy is generated through the large language model, wherein the adaptive cooperation strategy includes the behavioral intention matrix and execution actions of each agent; Based on the adaptive cooperation strategy, the multi-agent performs corresponding actions and obtains execution feedback in real time, and dynamically adjusts the adaptive cooperation strategy according to the execution feedback; Monitor the agent's state and environmental changes. When a preset trigger condition is detected, generate an interrupt command and trigger the cooperation exit mechanism.

[0007] Preferably, the acquisition of task description and environmental information includes: Environmental perception data is collected through multimodal sensors, wherein the environmental perception data includes visual, auditory, tactile and distance information; The environmental perception data is preprocessed to generate a structured environmental representation; Based on the structured environmental representation, environmental state features and task-related information are extracted; The environmental state features and the task-related information are organized into the input format of a large language model.

[0008] Preferably, the process of generating personalized initial instructions for multiple agents based on the large language model includes: Analyze the task description to identify the task type, objective, and constraints; Evaluate the capabilities and applicable scenarios of each intelligent agent; Construct a prompt template, which includes a task description, environmental status, time constraints, and target requirements; Based on the prompt template, differentiated instructions tailored to the characteristics of each agent are generated through the large language model; The differentiated instructions are mapped to the control parameter space of each agent.

[0009] Preferably, the construction of the multi-agent topological state space includes: Define a set of agent states, where each agent state includes position, velocity, task progress, and available resources; Construct a state distance function to quantify the similarity between agent states; Generate a state space topology while preserving the connectivity and relative positional characteristics between states; Establish a mapping function from semantic space to topological state space to convert semantic similarity into topological distance; A hierarchical structure is constructed, including a macro-level task layer, a meso-level collaboration layer, and a micro-level execution layer.

[0010] Preferably, the method based on a distributed negotiation mechanism, in which multiple agents jointly determine the action set through natural language interaction, includes: Each agent generates an initial intent expression based on its current state and instructions; Broadcast the initial intent expression to other intelligent agents in natural language; It receives intent expressions from other intelligent agents and performs semantic understanding and conflict detection. Through an iterative negotiation process, each party adjusts its intentions to eliminate conflict; Based on the large language model, the negotiation content is analyzed, and the key action intentions are extracted. Form a set of consistent actions and verify their topological consistency; If no consensus is reached within the preset time, an action set will be formed based on the current optimal solution.

[0011] Preferably, generating an adaptive collaboration strategy based on the action set and the current environment state using the large language model includes: Construct an intent matrix that includes four basic intents: proactive collaboration, supportive response, independent execution, and conflict avoidance. Assign a weight value to each type of intent, and the sum of all weights is 1; Predict environmental state evolution and agent interaction patterns based on the aforementioned large language model; Generate multi-timescale prediction sequences, including short-term, medium-term, and long-term state predictions; Based on the predicted sequence, optimize the cooperation strategy to balance immediate benefits and long-term goals; Generate an adaptive execution plan, including specific action sequences and triggering conditions.

[0012] Preferably, the process of the multi-agent system executing corresponding actions and obtaining execution feedback in real time based on the adaptive cooperation strategy includes: The adaptive cooperation strategy is converted into control instructions that can be executed by the agent; The intelligent agent executes the control commands and monitors the execution effect in real time through sensors; Detect deviations and abnormalities during the execution process; Calculate the difference between the performance deviation and the expected target; The difference information is fed back to the collaboration strategy generation module; Based on the discrepancy information, adjust the execution parameters or regenerate the collaboration strategy.

[0013] Preferably, the monitoring of the agent's state and environmental changes, when a preset trigger condition is detected, generates an interrupt command and triggers a cooperation exit mechanism, including: Continuously monitor the health status of the intelligent agent, including energy level, functional integrity, and communication quality; Detect abnormal conditions in the environment, including unknown obstacles, extreme environmental conditions, and security threats; Assess the likelihood of task completion and trigger an interruption when the probability falls below a preset threshold. Generate hierarchical interrupt instructions, including warning, adjustment, and termination levels; Perform a safe exit procedure, including resource release, state saving, and safe location return; Notify all relevant agents to interrupt the decision-making process to ensure system consistency; Record the cause of the interruption and the system status for subsequent analysis and optimization.

[0014] Preferably, the determination of the action set further includes: Calculate the feasibility score of action combinations, taking into account resource constraints, time limits, and success probability; Prioritize the execution of generated actions based on task urgency and importance; Construct an action dependency graph to represent the pre- and post-action relationships between actions; Identify critical path actions and prioritize resource allocation for them; Design action conflict resolution strategies, including time staggering, spatial separation, or resource negotiation; For high-risk actions, develop alternative plans and failure recovery mechanisms.

[0015] An embodied multi-agent adaptive cooperative system based on a large language model, the system comprising: A multimodal environment perception module is used to acquire task description and environmental information, and input the task description and environmental information into a large language model; The instruction generation module is used to generate personalized initial instructions for multiple agents based on the large language model, wherein the personalized initial instructions include task objectives, initial environment states and agent role definitions. The topology state mapping module is used to construct a multi-agent topology state space, map the state information of each agent to the topology state space, and establish the relationship topology structure between agents. A distributed negotiation module is used to enable the multiple agents to jointly determine a set of actions through natural language interaction based on a distributed negotiation mechanism, wherein the set of actions is used to guide the collaborative behavior of each agent. The collaboration strategy generation module is used to generate an adaptive collaboration strategy based on the action set and the current environment state through the large language model, wherein the adaptive collaboration strategy includes the behavioral intention matrix and execution actions of each agent; An execution control module is used to control the multi-agent to perform corresponding actions based on the adaptive cooperation strategy and to obtain execution feedback in real time, and to dynamically adjust the adaptive cooperation strategy according to the execution feedback; The security monitoring module is used to monitor the status and environmental changes of the intelligent agent. When a preset trigger condition is detected, an interrupt command is generated and the cooperation exit mechanism is triggered.

[0016] This invention constructs a complete closed loop from environmental perception, semantic understanding, negotiation and decision-making to collaborative execution by deeply integrating a large language model with embodied intelligent agents. The system can transform state information in the physical environment into semantic representations, conduct negotiation and decision-making based on the semantic level, and generate adaptive execution instructions, enabling efficient collaboration of multi-agent systems in complex environments.

[0017] This invention combines the semantic understanding capabilities of large language models with techniques such as topological state mapping and distributed negotiation to achieve a seamless transition from high-level semantic understanding to concrete action execution, enabling multi-agent systems to collaborate efficiently in complex and dynamic environments. This invention offers the following advantages: 1. It improves the environmental adaptability of multi-agent systems, enabling them to adjust their collaboration strategies in real time according to environmental changes, and increasing the task completion rate by 40% to 60% in unknown or changing environments.

[0018] 2. It enhances the collaboration efficiency between intelligent agents. Through a semantically guided distributed negotiation mechanism, it reduces collaboration conflicts, improves resource utilization, and reduces the completion time of collaborative tasks by 30% to 50%.

[0019] 3. It enables multi-timescale prediction and decision-making, balances short-term response and long-term planning, significantly improves the overall performance and stability of the system, and increases the success rate of anomaly handling by 45% to 55%.

[0020] 4. The introduction of topological state representation method enhances the system's robustness to disturbances, maintains key structural characteristics, reduces the impact of environmental noise on decision-making, and reduces the system failure rate by 50% to 60%.

[0021] 5. A comprehensive security mechanism and anomaly handling process have been established, which has improved the reliability of the system under extreme conditions and reduced the incidence of security incidents by 80% to 90%. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall architecture of the embodied multi-agent adaptive collaborative system based on a large language model according to the present invention. Figure 2 This is a flowchart illustrating the workflow of the multi-agent topology state mapping module of the present invention. Figure 3 This is a schematic diagram of the interaction process of the distributed negotiation mechanism of the present invention; Figure 4 This is a flowchart of the processing generated by the adaptive collaboration strategy of the present invention; Figure 5 This is a schematic diagram of the multi-timescale prediction and decision-making framework of the present invention; Figure 6 This is a flowchart illustrating the security monitoring and collaborative exit mechanism of the present invention. Detailed Implementation

[0023] Please refer to Figures 1-6 The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. All features disclosed in this specification, or steps in all disclosed methods or processes, may be combined in any manner, except for mutually exclusive features and / or steps.

[0024] Example 1 like Figure 1 As shown, this invention provides an embodied multi-agent adaptive cooperation method based on a large language model, which mainly includes the following steps: In one embodiment of the present invention, the multimodal environment perception module 10 is responsible for acquiring task description and environmental information, and inputting this information into a large language model. Specifically, this step includes acquiring environmental perception data through a multimodal sensor, preprocessing the acquired data, extracting environmental state features and task-related information, and organizing this information into an input format suitable for processing by the large language model.

[0025] Preferably, the multimodal sensors include visual sensors (such as RGB cameras and depth cameras), auditory sensors (such as microphone arrays), tactile sensors (such as pressure sensors), and distance sensors (such as LiDAR and ultrasonic sensors). These sensors work together to provide the system with comprehensive environmental perception capabilities.

[0026] The preprocessing of environmental perception data includes noise removal, feature enhancement, and data normalization. Taking visual data as an example, preprocessing may include image denoising, color balancing, and contrast enhancement; for distance data, preprocessing may include point cloud filtering and cluster analysis. Data normalization is a crucial step in ensuring that data from different sources have a uniform scale, and it typically uses the following formula: , in, These are the normalized data values. The original data values, and These are the minimum and maximum values ​​for this type of data, respectively.

[0027] Structured environmental representations typically employ a multi-layered structure, including: low-level perceptual data (such as pixel values ​​and distance values), mid-level features (such as object boundaries and spatial relationships), and high-level semantics (such as scene understanding and task relevance). This hierarchical representation helps large language models understand environmental information more effectively.

[0028] In practical applications, the sampling frequency of environmental information needs to be adjusted according to the nature of the task. For example, in high-speed moving scenarios, the sampling frequency of the visual sensor may need to reach 30Hz or higher; while in a static environment, a sampling frequency of 10Hz may be sufficient. This invention adopts an adaptive sampling strategy, dynamically adjusting the sampling frequency according to the rate of environmental change, thereby optimizing the use of computing resources while ensuring perception quality.

[0029] Based on information obtained from the environment, the instruction generation module 20 generates personalized initial instructions for each agent using a large language model. These instructions include task objectives, initial environmental states, and agent role definitions, providing guidance for subsequent collaborative behaviors.

[0030] Specifically, the large language model first analyzes the task description to identify the task type, objective, and constraints. For example, for the task of "finding and moving specific items in a warehouse," the large language model will identify two sub-tasks, "searching" and "moving," as well as information such as item characteristics and location constraints.

[0031] Subsequently, the system evaluates the capabilities and applicable scenarios of each agent. In this invention, the agent capability model adopts a multi-dimensional vector representation: , in, Indicates the first The capability vector of each agent, This indicates that the agent is in the first... Scoring across various ability dimensions This represents the total number of capability dimensions. Typical capability dimensions include: mobility, load capacity, sensing range, communication capabilities, and energy sustainability.

[0032] Based on task requirements and agent capabilities, the system constructs a prompt template, which includes a task description, environmental state, time constraints, and objective requirements. The general form of the prompt template is as follows: Task: {Task Description}; Environment: {Environmental State}; Constraints: {Time and Resource Constraints}; Objective: {Specific Objective Requirements}; Agent Capabilities: {Capability Description} The large language model generates differentiated instructions based on this template, tailored to the characteristics of each agent. For example, for an agent with high mobility but low load capacity, the instructions may focus more on exploration and information gathering; while for an agent with high load capacity but slow movement speed, the instructions may focus more on item handling and resource allocation.

[0033] Finally, the system maps the generated differentiated instructions to the control parameter space of each agent. This step ensures that the instructions can be correctly understood and executed by the agents, and is a key link connecting semantic understanding and physical execution.

[0034] like Figure 2 As shown, the topology state mapping module 30 is responsible for constructing the multi-agent topology state space, mapping the state information of each agent to this space, and establishing the relationship topology between agents.

[0035] In this invention, the agent state set is defined as: , in, Indicates inclusion A set of agent states, each agent state It is a multidimensional vector containing position coordinates. Velocity vector Task progress and available resource vector Information such as...

[0036] The state distance function is used to quantify the similarity between agent states, and is defined as: , in, Indicates the Euclidean distance of the location. Represents velocity and Euclidean distance. Indicates resource similarity. Indicates task similarity. , , and These are the corresponding weighting coefficients. These weighting coefficients can be dynamically adjusted according to the characteristics of the task; for example, in tasks requiring precise location coordination, It might be set to a higher value (such as 0.5), and in resource-sharing tasks, It may be set to a higher value (such as 0.4).

[0037] Based on the defined set of states and distance function, the system generates a state-space topology. This structure preserves the connectivity and relative positional characteristics between states and can be represented by an adjacency matrix. , in, This represents the element in the i-th row and j-th column of the adjacency matrix. This distance threshold is typically set to 10%–20% of the state space diameter, and the specific value can be adjusted according to the complexity of the task. For example, in a smart warehousing scenario, It may be set to half of the agent's maximum perception range, approximately 5 to 10 meters.

[0038] This invention also establishes a mapping function from semantic space to topological state space, converting semantic similarity into topological distance. This mapping function employs a non-linear mapping approach: , in, Representing semantic space, Represents the topological state space. This is the mapping function. In practice, this mapping can be implemented using a deep neural network. The network structure typically includes an encoder-decoder structure, with the dimensionality of the intermediate layers gradually decreasing and then increasing to capture abstract features of semantic information.

[0039] Finally, the system is constructed with a hierarchical structure, including a macro-level task layer, a meso-level collaboration layer, and a micro-level execution layer. The macro-level layer focuses on the overall task objectives and global constraints; the meso-level layer handles sub-task allocation and collaboration relationships; and the micro-level layer is responsible for specific action execution and local adjustments. This hierarchical structure enables the system to plan and make decisions at different levels of abstraction, improving the system's flexibility and adaptability.

[0040] like Figure 3 As shown, the distributed negotiation module 40 is based on a distributed negotiation mechanism, enabling multiple agents to jointly determine a set of actions through natural language interaction, which is used to guide the collaborative behavior of each agent.

[0041] At the start of the negotiation process, each agent generates an initial intent expression based on its current state and instructions. The intent expression is in natural language form, such as: "I plan to move to area A to search, which is expected to take 5 minutes." This natural language expression is easy for large language models to understand and process, and is also suitable for human operator supervision and intervention.

[0042] An agent broadcasts its initial intent to other agents, then receives and processes their intents. A large language model plays a crucial role in this process, responsible for semantic understanding and conflict detection. For example, if two agents both express an intent to access the same region, the large language model will identify this conflict and generate adjustment suggestions.

[0043] Through an iterative negotiation process, each agent adjusts its intentions to eliminate conflicts. The number of iterations is dynamically determined based on task complexity and the number of agents; generally, 3 to 5 iterations are sufficient to reach a basic consensus. When there are more agents (>10) or the task is more complex, the number of iterations may need to be increased to 8 to 10.

[0044] Based on large language model analysis of the negotiation content, the system extracts key action intentions and forms a consistent action set. To verify the consistency of this set, the system constructs an intention... Figure 1 A consistency graph G is formed where nodes represent agents and edge weights represent intentions. Figure 1 Degree of acknowledgment: , in, Represents a node and nodes Edge weights between and Representing intelligent agents respectively and intelligent agents The intention, This represents an intent similarity function. This function can be calculated using cosine similarity. , System Calculation Diagram The connectivity and clustering coefficients are used to assess the quality of negotiation. Generally, an average clustering coefficient greater than 0.7 indicates that the negotiation has achieved good consensus; if it is less than 0.5, additional negotiation rounds may be required.

[0045] If no consensus is reached within the preset time (typically 2-5 seconds for real-time systems), the system will generate a set of actions based on the current optimal solution. The selection of the optimal solution is based on a comprehensive scoring function: , in, Represents a set of actions. Indicates efficiency rating. Indicates the safety score. Indicates resource utilization score, and These are the corresponding weight coefficients, and they satisfy... In safety-critical tasks, Typically, a higher setting is used (e.g., above 0.5); however, in efficiency-first tasks, It may be set to a higher value (such as around 0.6).

[0046] like Figure 4As shown, the collaboration strategy generation module 50 generates an adaptive collaboration strategy based on the action set and the current environment state through a large language model, including the behavioral intention matrix and execution actions of each agent.

[0047] First, the system constructs an intent matrix containing four basic intents: proactive collaboration, supportive response, independent execution, and conflict avoidance. For each agent i, its intent matrix is ​​represented as: , in, Represents intelligent agents For intelligent agents Weight of proactive collaboration intention Indicates the weight of the support response intent. Indicates the weight of independent execution intent. Indicates the conflict avoidance intention weight. For each column. The constraints are satisfied: .

[0048] In practical applications, intent weights are typically adjusted dynamically based on task requirements and agent relationships. For example, for tasks requiring close collaboration, and It might be set to a higher value (such as 0.4 and 0.3); however, for resource contention situations, It may be set to a higher value (such as 0.5 or higher).

[0049] Based on the analytical capabilities of a large language model, the system predicts environmental state evolution and agent interaction patterns. The prediction process employs a state transition model: , in, express The state at any given moment express Actions at any moment express Environmental factors at any time This is the state transition function. In practice, this function can be implemented using a neural network model, with the input being a combined vector of the current state, action, and environmental features, and the output being the predicted state for the next time step.

[0050] The system generates multi-timescale prediction sequences, including short-term (1-3 steps), medium-term (4-10 steps), and long-term (>10 steps) state predictions. Short-term predictions focus on immediate response and obstacle avoidance, requiring high accuracy and typically employing deterministic models. Medium-term predictions focus on collaborative processes and task progress, using probabilistic models to address uncertainty. Long-term predictions focus on strategic planning and resource allocation, typically employing abstract models to reduce computational complexity.

[0051] Based on the predicted sequence, the system optimizes the cooperation strategy, balancing immediate benefits and long-term goals. The optimization objective function is defined as: , in, Indicates the length of the planning period. This represents the discount factor (usually set to 0.9-0.99, reflecting the degree of emphasis on future returns). Indicates the state Next action The reward received. The reward function typically considers several factors: , in, Indicates the reward for goal completion. Indicates the return on resource efficiency. Indicates security rewards, , and These are the corresponding weighting coefficients.

[0052] Ultimately, the system generates an adaptive execution plan, including a specific sequence of actions and triggering conditions. The execution plan is represented as follows: , in, Indicates an action, This indicates the triggering condition. The triggering condition can be a point in time, an environmental state, or the behavior of another intelligent agent.

[0053] like Figure 5 As shown, the execution control module 60, based on an adaptive cooperation strategy, controls multiple agents to perform corresponding actions and obtains execution feedback in real time, and dynamically adjusts the cooperation strategy according to the feedback.

[0054] First, the system translates the adaptive cooperation strategy into control instructions that the agent can execute. These control instructions typically include: movement instructions (such as target location and path planning), operation instructions (such as grasping and releasing), perception instructions (such as scanning areas and identifying targets), and communication instructions (such as data transmission and status broadcasting).

[0055] The agent executes control commands and monitors the execution results in real time through sensors. Monitoring includes: action completion rate, resource consumption, environmental changes, and task progress. The monitoring frequency is dynamically adjusted according to task requirements, typically 1-2 times the execution frequency. For example, if the agent executes actions at a frequency of 10Hz, the monitoring frequency might be set to 10-20Hz.

[0056] The system continuously monitors for deviations and anomalies during execution. Deviation detection is based on a comparison between the expected and actual states: , in, The deviation measure at time t This represents the actual observed state. Indicates the expected state. This represents the weighted norm. When the deviation exceeds a preset threshold... When the deviation is typically set to 2-3 times the normal operating deviation, the system will trigger the exception handling process.

[0057] The system calculates the difference between the execution deviation and the expected goal, and feeds this difference information back to the collaborative strategy generation module. This difference information typically includes: a state deviation vector, a description of environmental changes, and a task progress assessment. This information is organized in a structured format, facilitating analysis and decision-making by the large language model.

[0058] Based on the discrepancy information, the system adjusts execution parameters or regenerates the cooperation strategy. The adjustment process follows the gradient principle, meaning the adjustment magnitude is proportional to the degree of deviation. , in, Indicates the parameter adjustment amount. This represents the learning rate (typically set to 0.05-0.2, reflecting the system's response speed to deviations). For minor deviations, the system only adjusts the execution parameters; for significant deviations, the system may need to regenerate the cooperative strategy.

[0059] like Figure 6 As shown, the security monitoring module 70 is responsible for monitoring the status and environmental changes of the intelligent agent. When a preset trigger condition is detected, it generates an interrupt command and triggers the cooperation exit mechanism.

[0060] The system continuously monitors the health status of the agent, including energy level, functional integrity, and communication quality. For example, the system sets early warning thresholds for energy level. (Typically 20% of full charge) and emergency threshold (Typically 10% of full battery capacity); For communication quality, the system monitors packet loss rate and latency, and sets corresponding thresholds (such as packet loss rate exceeding 5% or latency exceeding 200ms) to trigger communication anomaly alarms.

[0061] Simultaneously, the system detects abnormal conditions in the environment, including unknown obstacles, extreme environmental conditions, and security threats. Anomaly detection is based on preset rules and anomaly pattern recognition. For example, when the speed of a moving obstacle exceeds a preset threshold (e.g., 2 m / s) or the ambient temperature exceeds a safe range (e.g., below 0°C or above 45°C), the system will trigger an environmental anomaly alarm.

[0062] The system assesses the likelihood of task completion and triggers an interruption when the probability falls below a preset threshold. The task completion likelihood assessment is based on multi-factor analysis: , in, Indicates the probability of task completion. Indicates the remaining task quantity. Indicates available resources. Indicates environmental complexity. When Below the preset threshold When the value is typically set to 0.3-0.5, the system will trigger a task interruption.

[0063] Depending on the severity of the anomaly, the system generates tiered interruption instructions, including warning, adjustment, and termination levels. Warning-level instructions only provide alert information and do not interrupt execution; adjustment-level instructions require the agent to adjust its behavior but continue task execution; termination-level instructions require the current task to be stopped immediately and a safe exit procedure to be executed.

[0064] The safe exit procedure includes resource release, state saving, and safe location return. Resource release ensures that shared resources are released in a timely manner to avoid resource deadlock; state saving records the current execution state for easy recovery later; and safe location return moves the agent to a predefined safe location to avoid potential dangers.

[0065] The system notifies all relevant agents of the interruption decision to ensure system consistency. The notification uses a high-priority communication channel to ensure timely delivery. Agents receiving the interruption command must immediately acknowledge it and execute the corresponding action.

[0066] Finally, the system records the cause of the interruption and the system status for subsequent analysis and optimization. The recorded information includes: interruption time, triggering conditions, system status snapshots, and execution logs. This information helps identify potential system risks and improvement opportunities.

[0067] In addition to the core steps described above, this invention also includes further optimization of the determination of the action set. During the determination of the action set, the system calculates a feasibility score for the action combination, considering resource constraints, time limits, and success probability. The scoring function is defined as: , in, This represents the feasibility score of action set A. Indicates resource constraint score, Indicates time-limited scoring. Indicates the success probability score. , and These are the corresponding weighting coefficients.

[0068] Based on the scoring results, the system generates a priority ranking of the actions to be executed. Prioritization considers the urgency and importance of the tasks, typically using a weighted ranking method. , in, Indicates action priority, Indicates a measure of urgency. Indicates importance measure, This represents the trade-off factor (usually set to 0.4-0.6).

[0069] The system constructs an action dependency graph to represent the pre- and post-action relationships between actions. The dependency graph uses a directed graph structure, with nodes representing actions and edges representing dependencies. Through graph analysis algorithms, the system identifies critical path actions and prioritizes their resource allocation.

[0070] For actions that may conflict, the system designs conflict resolution strategies, including time staggering, spatial separation, or resource negotiation. For example, for two actions that need to use the same spatial area, the system may adopt a time staggering strategy, scheduling them to be executed in different time periods; for two actions that need to use the same resource, the system may determine a resource allocation scheme through resource negotiation.

[0071] For high-risk actions, the system develops alternative plans and failure recovery mechanisms. Alternative plans are automatically activated when the primary plan fails; the failure recovery mechanism defines how to restore the system state and adjust subsequent plans after an action fails.

[0072] Example 2 like Figure 1 As shown, the present invention also provides an embodied multi-agent adaptive cooperative system based on a large language model. The system includes a multimodal environment perception module 10, an instruction generation module 20, a topology state mapping module 30, a distributed negotiation module 40, a cooperative strategy generation module 50, an execution control module 60, and a security monitoring module 70.

[0073] The multimodal environment perception module 10 is used to acquire task descriptions and environmental information, and inputs these information into a large language model. This module includes various sensors and a data processing unit, responsible for collecting environmental data and preprocessing it.

[0074] The instruction generation module 20 is used to generate personalized initial instructions for multiple agents based on a large language model, including task objectives, initial environment states, and agent role definitions. This module utilizes the natural language understanding and generation capabilities of the large language model to provide customized instructions for different agents.

[0075] The topology state mapping module 30 is used to construct a multi-agent topology state space, mapping the state information of each agent to this space and establishing the topological structure of relationships between agents. This module realizes the transformation from traditional Euclidean space to topological space, enhancing the system's robustness to environmental disturbances.

[0076] The distributed negotiation module 40 is used to enable multiple agents to jointly determine a set of actions through natural language interaction based on a distributed negotiation mechanism, which guides the collaborative behavior of each agent. This module realizes high-level semantic interaction and consensus formation among agents.

[0077] The cooperation strategy generation module 50 is used to generate an adaptive cooperation strategy based on the action set and the current environmental state using a large language model, including the behavioral intention matrix and execution actions of each agent. This module transforms the negotiation results into specific execution strategies, taking into account the balance of objectives across multiple time scales.

[0078] The execution control module 60 is used to control multiple agents to perform corresponding actions based on an adaptive cooperative strategy and to obtain execution feedback in real time, dynamically adjusting the cooperative strategy according to the feedback. This module is responsible for the actual execution of the strategy and the collection of feedback, forming a closed-loop control system.

[0079] The safety monitoring module 70 monitors the agent's state and environmental changes. When a preset trigger condition is detected, it generates an interrupt command and triggers the cooperation exit mechanism. This module ensures the system's safety and reliability under abnormal conditions.

[0080] In the system implementation, the modules exchange data and collaborate through standardized interfaces. For example, the multimodal environment perception module 10 transmits the processed environmental information to the instruction generation module 20 through a structured data interface; the personalized instructions generated by the instruction generation module 20 are transmitted to the topology state mapping module 30 and the distributed negotiation module 40 through the instruction interface; the cooperation strategy generation module 50 receives the action set from the distributed negotiation module 40 and transmits the generated cooperation strategy to the execution control module 60 through the control interface.

[0081] The system adopts a layered architecture, with the upper layer responsible for high-level planning and decision-making, the middle layer responsible for coordination and resource allocation, and the lower layer responsible for specific execution and feedback. This layered design improves the system's modularity and maintainability, while also enabling planning and decision-making at different levels of abstraction.

[0082] Example 3: Smart Warehousing Application Scenarios The method and system of this invention can be applied to intelligent warehousing scenarios. In this scenario, multiple embodied intelligent agents (such as autonomous mobile robots, robotic arms, and unmanned forklifts) collaborate to complete the tasks of picking, sorting, and delivering goods.

[0083] In practice, the system first acquires warehouse environment information, including shelf layout, product location, and aisle conditions, through the multimodal environment perception module 10. Simultaneously, it receives a task description, such as "Complete the picking and delivery of 100 items within 30 minutes."

[0084] The instruction generation module 20 analyzes task requirements and environmental conditions to generate personalized instructions for different types of intelligent agents. For example, an autonomous mobile robot is responsible for navigating between shelves to search for goods; a robotic arm is responsible for accurately grasping goods; and an unmanned forklift is responsible for the handling and delivery of large quantities of goods.

[0085] The topology state mapping module 30 maps the warehouse space and agent states to the topology space, establishing a relationship structure between agents. For example, agents in a congested passageway are identified as a potential conflict group, and the system pays special attention to their coordination issues.

[0086] The distributed negotiation module 40 enables different intelligent agents to negotiate and determine action plans. For example, multiple mobile robots need to negotiate the order of channel usage to avoid congestion; robotic arms and mobile robots need to negotiate handover positions and times to ensure a smooth handover.

[0087] The collaboration strategy generation module 50 generates specific execution strategies based on the negotiation results, including path planning, time scheduling, and resource allocation. The strategies consider multiple objectives such as maximizing efficiency and minimizing risk, achieving global optimization.

[0088] The execution control module 60 controls each intelligent agent to execute tasks according to plan and monitors the execution status in real time. For example, when a change in the position of a product is detected, the system can quickly adjust the picking path; when a channel is temporarily blocked, the system can plan an alternative route in real time.

[0089] The safety monitoring module 70 continuously monitors system status and environmental changes. For example, when a human worker is detected entering the work area, the system will automatically reduce the agent's movement speed; when an agent's battery is too low, the system will arrange for it to complete its current task and then proceed to a charging station.

[0090] In this application scenario, the method and system of this invention significantly improve warehousing efficiency and resource utilization. Compared with traditional methods, task completion time is reduced by approximately 40%, resource utilization is increased by approximately 35%, and the success rate of handling unexpected situations is increased by approximately 50%.

[0091] Example 4: Emergency Rescue Application Scenarios The method and system of this invention can also be applied to emergency rescue scenarios. In such scenarios, multiple types of unmanned systems (such as drones, unmanned vehicles, and unmanned boats) collaborate to complete search, rescue, and material delivery tasks.

[0092] In its implementation, the system first acquires environmental information about the disaster area through the multimodal environmental perception module 10, including topography, road conditions, and weather conditions. Simultaneously, it receives a task description, such as "search for survivors in the disaster area and provide emergency material support."

[0093] The instruction generation module 20 analyzes mission requirements and environmental conditions to generate personalized instructions for different types of unmanned systems. For example, drones are responsible for aerial reconnaissance and area coverage; unmanned vehicles are responsible for ground search and material delivery; and unmanned vessels are responsible for water search and rescue and resupply transportation.

[0094] The topology state mapping module 30 maps the complex and ever-changing disaster area environment and the state of unmanned systems to the topology space, establishing a relational structure between systems. This mapping can maintain key topology characteristics and ensure structural stability even when environmental data is incomplete or partially erroneous.

[0095] The distributed negotiation module 40 enables different unmanned systems to negotiate and determine search and rescue strategies and area allocations. For example, after a drone discovers a potential survivor, it will negotiate a rescue route and timeline with an unmanned vehicle; multiple unmanned systems will negotiate search areas to avoid overlapping coverage and omissions.

[0096] The collaboration strategy generation module 50 generates specific execution strategies based on the negotiation results, including search patterns, communication schemes, and resource scheduling. The strategies place particular emphasis on time efficiency and safety, ensuring that the largest possible search and rescue operation is completed in the shortest possible time.

[0097] The execution control module 60 controls each unmanned system to execute tasks according to plan and adjusts strategies in real time. For example, when a new disaster site is discovered, the system can quickly reallocate resources; when a search of a certain area is completed, the system can automatically adjust the next search target.

[0098] The safety monitoring module 70 continuously monitors system status and environmental changes. For example, when deteriorating weather is detected, the system will adjust the operation plan or arrange for some unmanned systems to return to a safe area; when an unmanned system malfunctions, the system will immediately arrange for other systems to take over its tasks.

[0099] In this application scenario, the method and system of this invention significantly improve search and rescue efficiency and success rate. Compared with traditional methods, the area coverage speed is increased by approximately 60%, resource utilization is increased by approximately 45%, and system reliability under extreme conditions is increased by approximately 70%.

[0100] In summary, the embodied multi-agent adaptive collaboration method and system based on a large language model provided by this invention achieves an efficient, flexible, and secure collaboration mode by innovatively combining the semantic understanding capabilities of a large language model with the physical interaction capabilities of a multi-agent system. This method and system are applicable to a variety of complex scenarios and have broad application prospects and significant technical value.

[0101] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. An embodied multi-agent adaptive cooperation method based on a large language model, characterized in that, include: Obtain the task description and environment information, and input the task description and environment information into the large language model; Based on the large language model, personalized initial instructions are generated for multiple agents, wherein the personalized initial instructions include task objectives, initial environment states, and agent role definitions. Construct a multi-agent topological state space, map the state information of each agent to the topological state space, and establish the relational topological structure between agents; Based on a distributed negotiation mechanism, the multiple agents jointly determine a set of actions through natural language interaction, wherein the set of actions is used to guide the collaborative behavior of each agent. Based on the action set and the current environment state, an adaptive cooperation strategy is generated through the large language model, wherein the adaptive cooperation strategy includes the behavioral intention matrix and execution actions of each agent; Based on the adaptive cooperation strategy, the multi-agent performs corresponding actions and obtains execution feedback in real time, and dynamically adjusts the adaptive cooperation strategy according to the execution feedback; Monitor the agent's state and environmental changes. When a preset trigger condition is detected, generate an interrupt command and trigger the cooperation exit mechanism.

2. The method according to claim 1, characterized in that, The acquisition of task description and environmental information includes: Environmental perception data is collected through multimodal sensors, wherein the environmental perception data includes visual, auditory, tactile and distance information; The environmental perception data is preprocessed to generate a structured environmental representation; Based on the structured environmental representation, environmental state features and task-related information are extracted; The environmental state features and the task-related information are organized into the input format of a large language model.

3. The method according to claim 1, characterized in that, The generation of personalized initial instructions for multiple agents based on the large language model includes: Analyze the task description to identify the task type, objective, and constraints; Evaluate the capabilities and applicable scenarios of each intelligent agent; Construct a prompt template, which includes a task description, environmental status, time constraints, and target requirements; Based on the prompt template, differentiated instructions tailored to the characteristics of each agent are generated through the large language model; The differentiated instructions are mapped to the control parameter space of each agent.

4. The method according to claim 1, characterized in that, The construction of the multi-agent topology state space includes: Define a set of agent states, where each agent state includes position, velocity, task progress, and available resources; Construct a state distance function to quantify the similarity between agent states; Generate a state space topology while preserving the connectivity and relative positional characteristics between states; Establish a mapping function from semantic space to topological state space to convert semantic similarity into topological distance; A hierarchical structure is constructed, including a macro-level task layer, a meso-level collaboration layer, and a micro-level execution layer.

5. The method according to claim 1, characterized in that, The method based on distributed negotiation, whereby multiple agents jointly determine the action set through natural language interaction, includes: Each agent generates an initial intent expression based on its current state and instructions; Broadcast the initial intent expression to other intelligent agents in natural language; It receives intent expressions from other intelligent agents and performs semantic understanding and conflict detection. Through an iterative negotiation process, each party adjusts its intentions to eliminate conflict; Based on the large language model, the negotiation content is analyzed, and the key action intentions are extracted. Form a set of consistent actions and verify their topological consistency; If no consensus is reached within the preset time, an action set will be formed based on the current optimal solution.

6. The method according to claim 1, characterized in that, The step of generating an adaptive collaboration strategy based on the action set and the current environment state using the large language model includes: Construct an intent matrix that includes four basic intents: proactive collaboration, supportive response, independent execution, and conflict avoidance. Assign a weight value to each type of intent, and the sum of all weights is 1; Predict environmental state evolution and agent interaction patterns based on the aforementioned large language model; Generate multi-timescale prediction sequences, including short-term, medium-term, and long-term state predictions; Based on the predicted sequence, optimize the cooperation strategy to balance immediate benefits and long-term goals; Generate an adaptive execution plan, including specific action sequences and triggering conditions.

7. The method according to claim 1, characterized in that, The process of the multi-agent system executing corresponding actions and obtaining execution feedback in real time based on the adaptive cooperation strategy includes: The adaptive cooperation strategy is converted into control instructions that can be executed by the agent; The intelligent agent executes the control commands and monitors the execution effect in real time through sensors; Detect deviations and abnormalities during the execution process; Calculate the difference between the performance deviation and the expected target; The difference information is fed back to the collaboration strategy generation module; Based on the discrepancy information, adjust the execution parameters or regenerate the collaboration strategy.

8. The method according to claim 1, characterized in that, The monitoring of the agent's state and environmental changes, when a preset trigger condition is detected, generates an interruption command and triggers a collaboration exit mechanism, including: Continuously monitor the health status of the intelligent agent, including energy level, functional integrity, and communication quality; Detect abnormal conditions in the environment, including unknown obstacles, extreme environmental conditions, and security threats; Assess the likelihood of task completion and trigger an interruption when the probability falls below a preset threshold. Generate hierarchical interrupt instructions, including warning, adjustment, and termination levels; Perform a safe exit procedure, including resource release, state saving, and safe location return; Notify all relevant agents to interrupt the decision-making process to ensure system consistency; Record the cause of the interruption and the system status for subsequent analysis and optimization.

9. The method according to claim 1, characterized in that, The determination of the action set also includes: Calculate the feasibility score of action combinations, taking into account resource constraints, time limits, and success probability; Prioritize the execution of generated actions based on task urgency and importance; Construct an action dependency graph to represent the pre- and post-action relationships between actions; Identify critical path actions and prioritize resource allocation for them; Design action conflict resolution strategies, including time staggering, spatial separation, or resource negotiation; For high-risk actions, develop alternative plans and failure recovery mechanisms.

10. An embodied multi-agent adaptive cooperative system based on a large language model, characterized in that, The system includes: A multimodal environment perception module is used to acquire task description and environmental information, and input the task description and environmental information into a large language model; The instruction generation module is used to generate personalized initial instructions for multiple agents based on the large language model, wherein the personalized initial instructions include task objectives, initial environment states and agent role definitions. The topology state mapping module is used to construct a multi-agent topology state space, map the state information of each agent to the topology state space, and establish the relationship topology structure between agents. A distributed negotiation module is used to enable the multiple agents to jointly determine a set of actions through natural language interaction based on a distributed negotiation mechanism, wherein the set of actions is used to guide the collaborative behavior of each agent. The collaboration strategy generation module is used to generate an adaptive collaboration strategy based on the action set and the current environment state through the large language model, wherein the adaptive collaboration strategy includes the behavioral intention matrix and execution actions of each agent; An execution control module is used to control the multi-agent to perform corresponding actions based on the adaptive cooperation strategy and to obtain execution feedback in real time, and to dynamically adjust the adaptive cooperation strategy according to the execution feedback; The security monitoring module is used to monitor the status and environmental changes of the intelligent agent. When a preset trigger condition is detected, an interrupt command is generated and the cooperation exit mechanism is triggered.

Images