Agent network communication method based on structured task intent perception and related apparatuses

Abstract

The present disclosure provides an agent network communication method based on structured task intention perception, and relates to the technical field of agent network communication. The method comprises: an application side agent generating a structured task intention according to task input; the structured task intention at least comprising: a target receiving object, semantic knowledge content; a network side agent performing task-oriented global collaboration and resource pre-allocation according to the structured task intention, and generating a network configuration; a communication side agent determining an optimal encoding scheme under the guidance of the network configuration and the structured task intention; mapping the structured task intention into a high-dimensional semantic vector based on the optimal encoding scheme and performing signal transmission; and an execution side agent acquiring a modulation signal through physical perception, generating an estimated semantic vector based on the modulation signal, performing semantic completion and task intention reconstruction on the estimated semantic vector based on a shared semantic knowledge base, and generating an estimated task intention.

Description

Technical Field

[0001] This disclosure relates to the field of intelligent agent network communication technology, and in particular to the fields of distributed artificial intelligence collaborative control and intelligent agent network task collaboration technology. Specifically, this disclosure relates to an intelligent agent network communication method and related apparatus based on structured task intent perception. Background Technology

[0002] Traditional communication systems are primarily based on Shannon's bit-level transmission paradigm, with the core objective of achieving error-free bit-level transmission. However, this paradigm faces significant challenges in highly dynamic and complex wireless environments: Firstly, directly transmitting massive streams of raw data between agents puts immense pressure on spectrum resources; traditional communication systems do not distinguish the importance of information during transmission, resulting in a large amount of irrelevant or redundant raw data being transmitted in the network, further straining spectrum resources and network capacity. Secondly, due to a lack of understanding of the "meaning" of information, traditional systems are prone to a "cliff effect" of drastic performance degradation under adverse channel conditions (such as low signal-to-noise ratio), leading to poor interaction reliability and making it difficult to meet the needs of precise and efficient collaboration between agents.

[0003] In contrast, semantic communication, as an emerging communication paradigm, is considered a key technology for realizing the 6G intelligent agent communication architecture. It no longer pursues precise bit reconstruction, but rather the maximization of the "semantic similarity" of information between the receiver and the sender.

[0004] While semantic communication and agent technologies bring new opportunities to next-generation networks, existing technologies still have the following significant problems: First, information transmission is disconnected from task intent: In agent networks, the transmission of massive amounts of raw data results in huge bandwidth waste and cannot guarantee that the receiver can accurately understand the sender's true intent.

[0005] Secondly, there is a lack of structured expression and understanding of semantic information: existing communication architectures lack standardized and structured encapsulation mechanisms for semantic knowledge content, making it difficult for intelligent agents to efficiently and accurately extract and understand the semantic information after receiving information, thus affecting the efficiency and accuracy of task execution.

[0006] Secondly, the reliability of interaction in low signal-to-noise ratio environments is insufficient: Although semantic communication is theoretically more robust, existing technical solutions have not been able to effectively solve the problem of how to ensure the reliable transmission of key task intentions and semantic information between agents in extremely low signal-to-noise ratio environments.

[0007] Finally, collaborative efficiency is low in dynamic environments. In highly dynamic network environments, agent nodes need to perform real-time path planning and priority calculation based on factors such as task requirements, spatiotemporal constraints, and service levels.

[0008] However, existing technologies lack a communication mechanism that can support efficient collaboration and dynamic adjustment among multiple agent nodes based on structured task intentions, resulting in the overall response speed and task completion rate of the system failing to meet expectations when facing complex and ever-changing real-world scenarios. Summary of the Invention

[0009] This disclosure provides a method, apparatus, electronic device, and computer-readable storage medium for intelligent agent network communication based on structured task intent awareness.

[0010] According to a first aspect of this disclosure, a method for intelligent agent network communication based on structured task intent awareness is provided, applicable to communication scenarios including application-side intelligent agents, network-side intelligent agents, communication-side intelligent agents, and execution-side intelligent agents, characterized in that it includes: The application-side intelligent agent generates a structured task intent based on the task input; the structured task intent includes at least: the target receiving object and semantic knowledge content; The network-side intelligent agent performs task-oriented global coordination and resource pre-allocation based on the structured task intent, and generates network configuration; Guided by the network configuration and the structured task intent, the communication-side agent determines the optimal encoding scheme; based on the optimal encoding scheme, it maps the structured task intent into a high-dimensional semantic vector and transmits the signal. The execution-side agent acquires the modulation signal through physical perception, generates an estimated semantic vector based on the modulation signal, and performs semantic completion and task intent reconstruction on the estimated semantic vector based on a shared semantic knowledge base to generate an estimated task intent.

[0011] Optionally, the method further includes: Based on the estimated task intent, generate executable semantic action instructions; Based on the execution result of the semantic action instruction, a semantic feedback vector containing the execution state is generated; The semantic action instructions include any one of the following: physical control instructions, natural language prompts, and service call parameters; The semantic feedback vector is used by the communication-side agent to perform semantic deviation evaluation and synchronize the update of the shared knowledge base; the semantic feedback vector is used by the network-side agent to perceive the global task achievement status and, based on this, to re-infer and optimize the subsequent macro-scheduling strategy; the semantic feedback vector is used by the application-side agent to evaluate the degree of achievement of the initial intent, thereby forming a complete closed-loop optimization.

[0012] Optionally, the application-side agent generates a structured task intent based on the task input, including: The application-side agent generates a preliminary task intent based on the task input; The application-side agent decomposes the initial task intent into a structured task intent based on historical task knowledge. The structured task intent also includes at least one of the following: expected effect, spatiotemporal constraints, reliability, semantic distortion tolerance, resource overhead, and security level.

[0013] Optionally, the target receiving object can be any of the following: a specified intelligent agent, multiple specified intelligent agents, intelligent agents within a specified spatial range, intelligent agents of a specified role type, or intelligent agents within a specified task group.

[0014] Optionally, the network configuration includes at least one of the following: Optimal macro-path, task priority, and resource pre-allocation scheme; The optimal macroscopic path includes a semantic routing strategy; The resource pre-allocation scheme includes at least one of the following: intra-network aggregation scheme, optimized distribution scheme, and intelligent distribution scheme.

[0015] Optionally, the task priority is calculated based on at least one of the following indicators: task urgency, semantic importance, semantic freshness, latency requirements, reliability requirements, channel and resource status, and target range.

[0016] Optionally, the communication-side agent, guided by the network configuration and the structured task intent, determines the optimal encoding scheme, including: Guided by the network configuration and the structured task intent, the communication-side agent acquires the physical channel state; and invokes the shared semantic knowledge base to determine the optimal encoding scheme based on the physical channel state. The signal transmission includes: transmitting signals through a distributed autonomy and cooperation layer.

[0017] Optionally, when there is only one target receiving object, the signal transmission adopts a unicast method with semantic routing that meets the requirements of latency and reliability; When there are multiple target receiving objects, the signal transmission adopts a broadcast or a semantic distribution method oriented towards multiple nodes; Optionally, the semantic completion includes: semantic vector matching, knowledge graph context completion, and logical consistency verification; The semantic vector matching involves matching the estimated semantic vector with entity, attribute, relation, and event representations in a shared semantic knowledge base to obtain candidate semantic units. The knowledge graph context completion is based on the graph relationships, event templates, scene rules and task contexts between candidate semantic units to infer semantic information that has not been explicitly transmitted or is missing due to channel influence, and obtain the completed semantic units. The logical consistency check utilizes the triple relationships and scene logic in the knowledge graph to perform consistency checks on the "entity-relationship-event-action" link in the completed semantic unit, eliminates logical conflicts, and obtains the checked semantic unit.

[0018] Optionally, the task intent reconstruction includes: identifying elements of the task intent from the semantically completed semantic units; and reconstructing the task intent based on the elements to obtain an estimated task intent.

[0019] Optionally, the communication scenario includes any one of the following: autonomous driving scenario, industrial internet scenario, embodied robot scenario, drone scenario, and satellite scenario.

[0020] According to a second aspect of this disclosure, an intelligent agent network communication device based on structured task intent awareness is provided, applied to a communication scenario including an application-side intelligent agent, a network-side intelligent agent, a communication-side intelligent agent, and an execution-side intelligent agent. Its features include: The intent generation module generates a structured task intent based on the task input by the application-side intelligent agent; the structured task intent includes at least: the target receiving object and semantic knowledge content. The network configuration module involves the network-side intelligent agent performing task-oriented global coordination and resource pre-allocation based on the structured task intent to generate the network configuration. The encoding and transmission module, guided by the network configuration and the structured task intent, determines the optimal encoding scheme; based on the optimal encoding scheme, it maps the structured task intent into a high-dimensional semantic vector and transmits the signal. The intent recovery module involves the execution-side agent acquiring a modulation signal through physical perception, generating an estimated semantic vector based on the modulation signal, and performing semantic completion and task intent reconstruction on the estimated semantic vector based on a shared semantic knowledge base to generate an estimated task intent.

[0021] According to a third aspect of this disclosure, an electronic device is provided, the electronic device comprising: At least one processor; and A memory communicatively connected to at least one of the aforementioned processors; wherein, The memory stores instructions that can be executed by at least one processor, and the instructions are executed by at least one processor to enable at least one processor to perform the method.

[0022] According to a fourth aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause a computer to perform the methods described above.

[0023] Compared with existing technologies, this disclosure has the following beneficial effects: First, by using structured task intent perception, it achieves complete semantic expression of task intent, avoiding semantic loss and distortion, and improving the accuracy and reliability of task transmission. Second, each agent node can independently parse task intent and perform localized path planning and priority calculation, reducing the centralized processing burden on the network hub and improving the system's distributed processing capability and response speed. Third, by supporting one or more agents as target receiving objects and combining diverse identification methods for target receiving objects, it achieves flexible support for different transmission needs, improving the efficiency and flexibility of task intent transmission. Fourth, by setting up a collaborative feedback channel and a task dynamic adjustment module, it achieves real-time monitoring and dynamic adjustment of the task execution process, enhancing the system's adaptability and fault tolerance. Fifth, by comprehensively considering multiple factors such as urgency, importance, resource contention, and service level in the priority calculation mechanism, it achieves dynamic optimization of task scheduling, adapting to application scenarios with different levels of urgency and resource conditions.

[0024] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0025] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is a flowchart illustrating a network communication method provided in an embodiment of this disclosure; Figure 2 This is a schematic diagram of the structure of a network communication device provided in an embodiment of this disclosure; Figure 3 This is a block diagram of an electronic device used to implement the network communication method of the embodiments of this disclosure. Detailed Implementation

[0026] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0027] The intelligent agent network communication scenario includes application-side intelligent agents, network-side intelligent agents, communication-side intelligent agents, and execution-side intelligent agents (terminal-side intelligent agents). Driven by shared semantic knowledge, task goal constraints, and feedback results, these intelligent agents achieve intent perception, semantic mapping, and distributed execution through closed-loop collaboration of "perception-memory-reasoning-execution," thereby completing efficient adaptive communication for tasks.

[0028] Task input: Task input is one of the inputs perceived by the application-side intelligent agent (sender) in the perception stage of the "perception-memory-reasoning-execution" closed loop. It refers to the signals or events automatically generated by user needs (such as instructions from human operators) or business systems (such as industrial production line fault detection systems, autonomous driving sensor fusion systems, etc.) to trigger the extraction and generation of task intent.

[0029] Application-side agent: The application-side agent senses the service trigger signal and sends the structured intent to the network hub. The application-side agent is responsible for generating and sending task intents and is the initiator of communication tasks, thus usually playing the role of the sender.

[0030] After receiving a structured intent, the network-side agent determines the optimal macro-path and task priority through a thought chain, triggering semantic encoding and resource pre-allocation at the processing layer. The network-side agent is responsible for forwarding, scheduling, and resource allocation; it acts as both a receiver (receiving intents from the application side) and a sender (issuing instructions to the communication and terminal sides), serving as an intermediate node.

[0031] Guided by the intent, the communication-side agent maps the intent into a high-dimensional semantic vector and transmits the signal. The communication-side agent is responsible for encoding the intent and transmitting the signal, acting as the sender at the physical transmission layer; it also receives scheduling instructions from the network side, thus also acting as the receiver. In other words, the communication-side agent compresses task-related information into a transmissible semantic representation.

[0032] Execution-side (terminal-side) agent: The execution-side agent acquires the modulated signal through physical perception, reconstructs the original intent of the sender based on semantic vectors during the inference phase, and maps it into specific physical control commands. Execution-side agents, such as embodied agents (e.g., vehicles, robotic arms, drones), are responsible for receiving and executing commands and are the executors of the communication task, thus typically playing the role of the receiver.

[0033] The network hub is the functional entity or logical node where the network-side intelligent agent resides. It is responsible for receiving structured task intentions from the application-side intelligent agent and performing operations such as global topology scanning, path planning, priority calculation, resource pre-allocation, semantic routing, intra-network aggregation, and intelligent distribution. It is the core scheduling and decision-making center in the semantic intelligent agent communication network that realizes the closed-loop collaboration of "perception-memory-reasoning-action".

[0034] The semantic intelligent agent communication network architecture of this invention is designed around intent perception, semantic mapping and distributed execution, enabling intelligent agents on the application side, network side, communication side and execution side to form a closed-loop collaboration of "perception-memory-reasoning-execution" driven by shared semantic knowledge, task goal constraints and feedback results.

[0035] Task intent perception is the process performed by the application-side intelligent agent (sender) in the perception phase of the "perception-memory-reasoning-action" closed loop. Its core is to extract and generate structured task intents from user needs or business trigger signals. The application-side intelligent agent perceives user needs or business trigger signals and extracts and generates task intents through the task intent perception phase.

[0036] Semantic mapping is a core step in the semantic agent communication network architecture. It refers to the process by which the communicating agent, guided by intent, encodes structured task intents (such as the target recipient, semantic knowledge content, and expected effect) into a transmittable high-dimensional semantic vector or semantic representation by invoking a shared semantic knowledge base. This process aims to compress task-related information into a compact semantic representation for efficient transmission under low bandwidth and to reconstruct it into an executable action at the receiving end.

[0037] Distributed execution is a core step in the semantic intelligent agent communication network architecture. It refers to the process by which the receiver's embodied intelligent agent (terminal-side agent), after completing semantic reconstruction, maps the original reconstructed intent into specific physical control commands or actions based on the received semantic instructions, prior knowledge stored in its own memory, and environmental feedback, and then executes these commands or actions in the physical world. This process is the final action in the "perception-memory-reasoning-execution" closed loop, ensuring that the actions executed by the receiver correspond to the original intent of the sender. Figure 1 Distributed execution can involve multiple agents executing simultaneously, but whether they do so depends on the task scenario and architecture design. For example, in the Industrial Internet, multiple robots can simultaneously perform semantic interactions and task execution; in drone rescue, multiple drones can simultaneously perform search and rescue missions.

[0038] However, "simultaneous execution" does not mean that all agents perform the same action. Instead, each agent performs actions related to its own task in parallel based on the received semantic instructions and local state, thereby achieving the collaborative completion of the global task.

[0039] Optimal macroscopic path: The optimal macroscopic path is the optimal transmission path from the sender to the receiver (or multiple receivers) determined by the network-side agent through the thought chain during the inference phase. It comprehensively considers latency, reliability, semantic distortion risk, resource consumption, and target task benefits, rather than simply choosing the path with the highest bit rate.

[0040] Task priority: Task priority is the process by which the network-side agent sorts multiple concurrent tasks based on factors such as task urgency, semantic importance, semantic novelty, latency and reliability requirements, channel and resource status, and the range of receiving objects, in order to determine the order of resource allocation.

[0041] Task urgency: Real-time tasks such as accident alarms, collision avoidance, and disaster relief have high priority; ordinary state synchronization or non-real-time knowledge updates have lower priority. Semantic Importance: Semantic features with a greater impact on task outcomes are transmitted first, such as target location, danger level, and critical state changes. Semantic Freshness: When the knowledge at the receiving end is outdated, or its semantic representation is inconsistent with the actual environment state, the priority of related semantic updates increases. Latency and Reliability Requirements: Tasks with high latency and high reliability receive priority for bandwidth, power, computing, and caching resources. Channel and Resource Status: Semantic compression ratio, transmission mode, and resource allocation are dynamically adjusted when there are channel fluctuations or resource constraints. Target Scope: Tasks affecting more agents or critical nodes have higher task impact weights during ranking.

[0042] Resource pre-allocation scheme: The resource pre-allocation scheme is a scheme triggered by the network-side intelligent agent during the action phase to reserve communication, computing, caching and semantic processing resources in advance. It involves the calculation and allocation of core parameters such as link bandwidth, SNR / SINR, available resource blocks, transmit power, node queue length, CPU / GPU load, and model inference latency.

[0043] Semantic routing strategy: The semantic routing strategy involves the network-side agent selecting a path that ensures the core semantics correctly reach the receiver based on the semantic content, task objective, and network state, rather than simply choosing the path with the highest bit rate. For a single receiver, a semantic route that meets latency and reliability requirements is selected; for multiple receivers, broadcasting or multi-node semantic distribution methods are used.

[0044] Intra-network aggregation scheme: When semantic information needs to be forwarded, merged or corrected across nodes, the network-side agent adopts an intra-network aggregation mechanism to merge semantic information from multiple sources at intermediate nodes, reducing duplicate transmissions and improving semantic transmission efficiency.

[0045] Optimize the distribution scheme: The network-side intelligent agent optimizes the distribution path and method of semantic information based on network connectivity, routing status, traffic load, congestion, available resources of edge nodes, and task latency requirements to ensure efficient transmission under the constraints.

[0046] Intelligent distribution scheme: The network-side intelligent agent combines semantic content, receiver characteristics and network status, and adopts intelligent distribution strategies (such as differentiated distribution based on semantic importance, on-demand distribution based on receiver knowledge base status, etc.) to achieve efficient semantic transmission for tasks.

[0047] Distributed Autonomous and Collaborative Layer: The Distributed Autonomous and Collaborative Layer is a functional layer in the semantic agent communication network architecture. It is responsible for modulating high-dimensional semantic vectors and transmitting them to the physical channel in the form of signals through distributed autonomous mechanisms and collaborative protocols after the agents on the communication side have completed semantic encoding. It also supports multiple agents to interact with high-dimensional semantic features simultaneously within the same resource block, realizing task-oriented collaborative transmission and resource scheduling.

[0048] Semantic control instructions: After completing semantic completion and intent reconstruction, the executing agent transforms the reconstructed original intent into various forms of instructions based on the task objective and its own capabilities to drive itself to perform specific operations. This process is not limited to a single physical control instruction but includes multiple forms such as physical control instructions, natural language prompts, and service call parameters. Its core purpose is to enable the receiving agent to execute actions consistent with the original intent of the sending agent. Figure 1 Semantic action instructions are defined, quantifiable actions or tasks. They include any of the following: physical control instructions, natural language prompts, and service call parameters. Physical control instructions: This is the most direct execution method, such as "decelerate" or "steer" in autonomous driving, or "grasp" or "move" low-level hardware control signals in industrial robots. Natural language prompts: These translate the reconstructed intent into human-readable natural language instructions or prompts to guide operators or serve as input for human-machine interaction. For example, in telemedicine, the intent "acute illness" is translated into a prompt "immediately activate emergency protocols." Service call parameters: These translate the intent into parameters for calling specific network services or application functions. For example, in the Industrial Internet, the intent "fault diagnosis" is translated into parameters for calling a specific diagnostic model on an edge server.

[0049] Estimated Task Intent: The estimated task intent is the semantically consistent intent representation reconstructed by the receiving end's communication-side agent or execution-side agent after semantic decoding and semantic completion, and is consistent with the original intent of the sending end. It is an approximate reconstruction of the original structured task intent of the sending end, containing core elements such as semantic knowledge content, expected effect, target object, and spatiotemporal constraints, and is used to guide the receiving end in executing tasks consistent with the original intent of the sending end. Figure 1The action or task performed.

[0050] Semantic completion: Semantic completion is the process by which the receiving end's communication-side agent or execution-side agent, after receiving an incomplete or channel-affected estimated semantic vector, combines prior knowledge graphs and contextual information from a shared semantic knowledge base to infer and supplement missing semantic information in order to restore the complete semantic meaning of the sender's original intent.

[0051] Semantic Feedback Vector: The semantic feedback vector is a set of information generated by the receiving agent (such as an embodied agent) after it perceives the execution result in real time during task execution. It contains information about its current understanding state (such as current understanding state, task completion status, missing semantic parts, and next transmission requirements). This vector is used to drive closed-loop optimization of the system, specifically including: the communication-side agent using it to evaluate semantic deviation and synchronize updates of the shared knowledge base; the network-side agent using it to perceive the global task achievement status and re-reasoning and optimizing subsequent macro-scheduling strategies; and the application-side agent using it to evaluate the degree of achievement of the initial intent, thus forming a complete "perception-memory-reasoning-execution" closed-loop optimization.

[0052] Semantic Distortion: Semantic distortion is an indicator used to determine the deviation between the task intent understood by the receiver and the original intent intended by the sender. When calculating semantic distortion, the original task intent from the sender is first broken down into several categories of key information. After receiving the signal, the receiver completes and reconstructs an estimated intent, which also includes these categories of information. The system converts each category of information from both the sender and receiver into a high-dimensional semantic vector and compares them within the same semantic space. A local distortion is calculated for each category of information: the closer to 0, the more accurate the reconstruction; the closer to 1, the greater the deviation.

[0053] Semantic deviation score: Semantic deviation score assesses not only whether the receiver understands the message, but also how well it executes the message and whether the understanding and execution are consistent. The inputs for calculating semantic deviation score include: the original task intent from the sender, the estimated task intent reconstructed by the receiver, and the feedback result from the executor after completing the task. These three are converted into semantic vectors and scored comprehensively across three dimensions: Reconstruction deviation: the difference in semantic similarity between the original intent and the intent understood by the receiver; Execution deviation: the difference between the desired effect of the original intent and the actual execution result (the better the effect, the smaller the deviation); Consistency deviation: whether the intent understood by the receiver is consistent with its actual action (e.g., understanding "decelerate" but actually "accelerate" results in a large consistency deviation). These three deviations are weighted and summed to obtain the semantic deviation score.

[0054] Improved KL divergence: Traditional KL divergence compares the differences between two probability distributions, but directly applying it to semantic communication has three problems: directional asymmetry, overflow when the probability is zero, and failure to distinguish information importance. Improved KL divergence specifically measures the difference in semantic distributions between the sender and receiver: First, it performs smoothing: adding a minimum value to all semantic probabilities to avoid infinitely large calculations due to zero probabilities; then, it ensures bidirectional symmetry: instead of calculating only one-way differences, it averages both directions to eliminate directional bias; it assigns higher weights to key semantics (urgent, high-value, high-reliability) to make the differences in important information have a greater impact on the result; improved KL divergence considers not only probability differences but also semantic relationship differences (conceptual hierarchy, closeness of relationship) and logical conflicts (whether it violates common sense, rules, or causal relationships); finally, it normalizes the overall result to a finite interval to avoid extreme values ​​affecting the judgment. Ultimately, improved KL divergence can simultaneously reflect probability differences, semantic relationship differences, logical conflicts, and information importance, making it suitable for semantic retransmission, knowledge base updates, and network resource scheduling.

[0055] Figure 1 A flowchart illustrating a network communication method provided in an embodiment of this disclosure is shown. Figure 1 As shown in the embodiments of this disclosure, the communication method includes the following steps.

[0056] Step S110: The application-side agent generates a structured task intent based on the task input; the structured task intent includes at least: the target receiving object and semantic knowledge content. Step S120: The network-side intelligent agent performs task-oriented global coordination and resource pre-allocation based on the structured task intent to generate network configuration; In step S130, the communication-side agent, guided by the network configuration and the structured task intent, determines the optimal encoding scheme; based on the optimal encoding scheme, it maps the structured task intent into a high-dimensional semantic vector and transmits the signal. Step S140: The execution agent acquires the modulation signal through physical perception, generates an estimated semantic vector based on the modulation signal, and performs semantic completion and task intent reconstruction on the estimated semantic vector based on the shared semantic knowledge base to generate an estimated task intent.

[0057] First, this solution achieves efficient task-oriented intent transmission. By transforming raw task input into structured task intent and performing semantic mapping and execution throughout the entire link, the system no longer needs to transmit massive amounts of raw data (such as video streams and raw sensor waveforms), but instead transmits concise semantic features. This significantly reduces the consumption of physical bandwidth, making it particularly suitable for communication scenarios with low bandwidth, high latency, or high interference (such as industrial IoT and drone swarm communication).

[0058] Secondly, this scheme enhances the robustness and anti-interference capability of communication. At the receiving end (execution side), the agent utilizes a shared semantic knowledge base to semantically complete incomplete or damaged semantic vectors. Even if physical channels cause signal distortion or partial data packet loss, the receiving end can infer the sender's original intent. This mechanism is similar to contextual understanding in human language communication, significantly improving the reliability of communication in harsh environments.

[0059] Furthermore, this solution optimizes the dynamic scheduling and utilization of network resources. The network-side agent is no longer a simple data forwarder, but rather performs global coordination and resource pre-allocation based on received intents. The network can dynamically adjust routing strategies and resource allocation (such as bandwidth and power) according to the urgency, semantic importance, and target object of the task. This avoids congestion caused by "best-effort" transmission in traditional networks, ensuring the priority and transmission quality of critical tasks.

[0060] Furthermore, this solution enhances the intelligence level of cross-domain collaboration. The various intelligent agents (application, network, communication, and execution) in the architecture collaborate through a closed loop of "perception-memory-reasoning-execution." In particular, the execution-side agents can feed back execution results to the network side to optimize subsequent scheduling strategies. This forms a dynamic system with self-optimization capabilities. The network is no longer a statically configured pipeline, but an intelligent agent network that can adjust in real time based on task feedback, thereby continuously improving the success rate of task execution.

[0061] Finally, this solution supports distributed cooperative execution of multiple agents. By mapping intents to general high-dimensional semantic vectors, different execution-side agents (such as heterogeneous robots, drones, and vehicles) can understand and execute the same instructions based on a shared knowledge base. This solves the communication protocol incompatibility problem caused by hardware heterogeneity in multi-agent systems, achieving efficient semantic interoperability between different devices.

[0062] Optionally, the method further includes: Based on the estimated task intent, generate executable semantic action instructions; Based on the execution result of the semantic action instruction, a semantic feedback vector containing the execution state is generated; The semantic action instructions include any one of the following: physical control instructions, natural language prompts, and service call parameters; The semantic feedback vector is used by the communication-side agent to perform semantic deviation evaluation and synchronize the update of the shared knowledge base; the semantic feedback vector is used by the network-side agent to perceive the global task achievement status and, based on this, to re-infer and optimize the subsequent macro-scheduling strategy; the semantic feedback vector is used by the application-side agent to evaluate the degree of achievement of the initial intent, thereby forming a complete closed-loop optimization.

[0063] The semantic feedback vector has multiple functions: it enables the communication-side agent to conduct semantic deviation assessment and realize synchronous iterative updates of the shared knowledge base; it enables the network-side agent to perceive the global task completion status and then reason and optimize the subsequent macro-scheduling strategy; and it enables the application-side agent to evaluate the degree of realization of the initial intention, thereby forming a complete semantic communication closed-loop optimization link.

[0064] It should be noted that the semantic feedback vector is not limited to a fixed-format mathematical vector, but rather is a comprehensive information carrier that carries the receiver's current semantic understanding state, task execution status, missing semantic content, and subsequent transmission supplementation needs. Its core purpose is to provide feedback on the semantic information already acquired by the receiver, the remaining information gaps, and the content that the sender is expected to further supplement the transmission.

[0065] Semantic distortion assessment is performed by the communication agent to determine the correctness of semantic transmission, that is, to measure the degree of deviation between the semantics reconstructed by the receiver and the original intent of the sender.

[0066] Synchronous updates of a shared knowledge base are essentially a process of continuous alignment and unification of semantic context between the sending and receiving ends. The shared semantic knowledge base typically undergoes initial semantic alignment through pre-training before communication and is dynamically updated during business operations. This updating and synchronization is performed by the communication agent. For example, the communication agent first verifies the consistency of semantic understanding between the two parties: if the degree of semantic distortion is low, the existing knowledge base remains unchanged; if the degree of semantic distortion is high, new concepts, semantic relationships, semantic codebooks, task experience, and environmental information are synchronously updated, enabling the knowledge bases of both parties to achieve semantic convergence again.

[0067] In the 6G intelligent and simplified architecture proposed in this application, the evaluation of task completion effect by the application intelligent agent breaks through the traditional communication evaluation paradigm that relies solely on bit-level key performance indicators (KPIs, such as latency and bandwidth), and constructs a multi-dimensional intent achievement evaluation system with key value indicators (KVI) at its core, specifically including three core indicators: By leveraging semantic feature extraction operators, the initial task intent of the sending end and the execution result feedback of the receiving end are mapped to a unified high-dimensional semantic space. Cosine similarity or improved KL divergence is used to quantify the semantic distance between the two, accurately measuring the degree of matching at the semantic level, and effectively solving the technical pain point of "the instruction is successfully delivered but there is a discrepancy in the semantic understanding between the two parties".

[0068] Industry-specific utility functions: For differentiated business needs in vertical industries (such as the grasping accuracy of embodied intelligent robotic arms and the success rate of drone obstacle avoidance), nonlinear customized utility functions are constructed; based on the completion status of actual execution parameters in physical space, they are mapped back to digital space to achieve quantitative evaluation of information value (VoI).

[0069] Resource utilization gain ratio: This measures the communication bit resources required to achieve a unit of intended value, characterizing the system's "intelligence and simplicity" level and judging whether the communication system can efficiently achieve the preset business application intent with minimal transmission overhead.

[0070] To resolve semantic ambiguities in multi-agent interaction and achieve deep consensus between sender and receiver, this solution uses a standardized computation process to achieve precise end-to-end intent logic alignment. The core mechanism is as follows: Closed-loop variational alignment: Before the semantic vector is sent, the sending agent performs a preliminary simulation and deduction of the intention execution effect based on the local environment observation model and the receiving behavior predictor, and predicts potential semantic deviations in advance.

[0071] Causal consistency constraint: Introducing the causal graph analysis method to constrain the execution sequence of the receiver's actions, ensuring that it strictly matches the sender's initial intent in terms of temporal and causal logic; by minimizing the mutual information difference between the initial intent distribution and the actual execution distribution, precise alignment of cross-agent end-to-end intents is achieved.

[0072] The storage of multi-agent collaborative experience is not simply log archiving, but a process of knowledge internalization, refinement, and reuse oriented towards the continuous iterative evolution of agents. Its dynamic accumulation follows the following computational rules: Reward conversion and weight update rules: A reinforcement learning reward decoupling mechanism is adopted. When the evaluation of intent achievement is higher than the preset threshold, the memory unit weights corresponding to the collaboration strategy are dynamically updated based on the improved Bellman equation to strengthen the weight ratio of high-quality collaboration experience.

[0073] Knowledge abstraction and redundancy elimination rules: Calculate the mutual information between newly added collaborative experience and existing entries in the memory bank; if the mutual information is higher than the threshold, it is determined to be the same type of experience, and the entries are merged and the features are enhanced; if the mutual information is low, it is used as a special scenario experience to open a new memory branch, realizing dynamic expansion of memory units and lightweight intelligent storage.

[0074] Importance sampling retention mechanism: For high-quality collaborative experiences in extreme channel environments, high-priority service tasks, etc., a larger decay period coefficient (lambda) is assigned to increase the retention time of such high-value experiences, so that they have a higher retrieval priority in the long-term reasoning and decision-making of the agent.

[0075] Optionally, the application-side agent generates a structured task intent based on the task input, including: The application-side agent generates a preliminary task intent based on the task input; The application-side agent decomposes the initial task intent into a structured task intent based on historical task knowledge. Optionally, the structured task intent mainly includes the following components: First, the target recipients. This refers to which intelligent agents or network nodes the intent needs to be sent to. It can be a specific intelligent agent, or intelligent agents within a certain spatial range, role type, or task group. For example, in an autonomous driving accident warning scenario, "vehicles within 300 meters behind" are the target recipients.

[0076] Second, the semantic knowledge content to be transmitted. This refers to the core semantic information that needs to be conveyed, rather than the complete raw data. This content can include events, targets, attributes, relationships, locations, times, state changes, etc. Examples include "An accident has occurred 50 meters ahead," "Location and severity of the accident," and "The road may be blocked."

[0077] Third, the expected effect. This refers to the cognitive outcome or action that the receiver should take after receiving the semantic information. Examples include "slowing down," "detouring," "collaborative perception," "improving target tracking accuracy," and "reconfiguring resources."

[0078] Fourth, spatiotemporal constraints. This refers to the effective area, effective time, event location, range of recipients, and task deadline of the intent. Examples include "50 meters ahead," "300 meters behind," "near the current time," and "low-latency broadcast."

[0079] Fifth, task service requirements. These include latency, reliability, semantic distortion tolerance, resource overhead, and security level. Tasks such as accident alarms, obstacle avoidance, and emergency collaboration have higher requirements for latency and reliability; tasks involving normal environment state synchronization or non-real-time knowledge updates have relatively lower requirements.

[0080] Sixth, semantic importance or urgency is used by the network side to determine the resource guarantee priority of the task corresponding to the intent. For example, danger alarms, collision risks, and sharing of critical targets are high semantic importance tasks; ordinary status updates and low-value background information are low priority tasks.

[0081] Optionally, the structured task intent may also include at least one of the following: expected effect, spatiotemporal constraints, reliability, semantic distortion tolerance, resource overhead, and security level.

[0082] Optionally, the target receiving object can be any of the following: a specified intelligent agent, multiple specified intelligent agents, intelligent agents within a specified spatial range, intelligent agents of a specified role type, or intelligent agents within a specified task group.

[0083] Optionally, the network configuration includes at least one of the following: Optimal macro-path, task priority, and resource pre-allocation scheme; The optimal macroscopic path includes a semantic routing strategy; The resource pre-allocation scheme includes at least one of the following: intra-network aggregation scheme, optimized distribution scheme, and intelligent distribution scheme.

[0084] Optionally, when the network-side agent determines the optimal macroscopic path during the inference phase, the thought chain can be understood as a decision chain "from business intent to network action": it first judges the core objectives of the communication task, such as low latency, high reliability, low energy consumption, or high semantic fidelity; then it reads the network topology, link quality, node load, edge computing power, cache, and semantic knowledge base deployment location to form an understanding of the current network state; subsequently, it generates several candidate paths on the network graph, instead of only selecting the shortest path; then it comprehensively evaluates each path, which simply means comparing the costs of "latency, congestion, energy consumption, reliability, and semantic distortion," which can be formally understood as a weighted result of "total cost = several indicators × weight"; finally, it selects the path that best meets the task objectives and reserves communication, computing, caching, and semantic processing resources in advance. Core parameters involved in resource pre-allocation include link bandwidth, signal-to-noise ratio (SNR / SINR), channel state information (CSI), available resource blocks, transmit power, node queue length, CPU / GPU load, model inference latency, knowledge base hit rate, end-to-end latency, packet loss rate, semantic distortion, and information freshness. Commonly used computational methods include graph search, K-shortest path, multi-objective optimization, reinforcement learning, and digital twin prediction. For example, in autonomous driving accident warnings, the shortest path might pass through congested roadside units (RSUs), while slightly more distant RSUs have idle computing power and traffic semantic knowledge bases. The network agent might choose the latter because it can complete the semantic processing and broadcasting of "accident location, risk level, and detour suggestions" faster. The capabilities of the network agent include topology perception, network digital twins, graph routing, path planning, and resource scheduling, which correspond to the above process.

[0085] Optionally, the task priority is calculated based on at least one of the following indicators: task urgency, semantic importance, semantic freshness, latency requirements, reliability requirements, channel and resource status, and target range.

[0086] Optionally, the communication-side agent, guided by the network configuration and the structured task intent, determines the optimal encoding scheme, including: Guided by the network configuration and the structured task intent, the communication-side agent acquires the physical channel state; and invokes the shared semantic knowledge base to determine the optimal encoding scheme based on the physical channel state. The signal transmission includes: transmitting signals through a distributed autonomy and cooperation layer.

[0087] The shared semantic knowledge base invoked by the communication agent is essentially a "semantic background and encoding prior" jointly owned by the sender and receiver. Its core content includes source semantic knowledge, task semantic knowledge, channel / transmission knowledge, and semantic representation rules. For example, in a traffic scenario, it would include concepts such as vehicles, pedestrians, accidents, lanes, risk levels, and action suggestions, along with their relationships, as well as representations such as semantic vectors, semantic tags, knowledge graphs, semantic codebooks, and context templates. During the inference phase, when determining the optimal encoding scheme, the communication agent needs to read feedback metrics such as CSI, SNR / SINR, bandwidth, available resource blocks, bit error rate, packet loss rate, historical retransmission count, moving speed, blocking probability, beam changes, as well as semantic similarity, semantic distortion, and task completion rate. The specific process is as follows: First, the task is parsed to determine whether it is security control, perception sharing, or ordinary transmission; then, relevant concepts are retrieved from the shared knowledge base to determine which semantics must be transmitted and which can be completed by the receiver; then, the original data is compressed into semantic vectors, labels, or "index + residual" using a semantic encoder; subsequently, the coding strength, compression ratio, and redundancy ratio are selected according to the channel state, with less redundancy and higher compression when the channel is good, and enhanced protection of key semantics when the channel is poor, using joint source-channel coding or generative semantic coding when necessary; finally, the receiver combines the knowledge base decoding and uses semantic similarity to determine whether retransmission is needed. The communication agent is responsible for semantic encoding and decoding, knowledge base negotiation and synchronization, while the semantic coding processing layer emphasizes mechanisms such as shared semantic knowledge base, CSI feedback, semantic hybrid automatic repeat request (HARQ), and semantic distortion.

[0088] When network agents perform re-inference and optimization of macro-scheduling strategies, they need to integrate multi-level and multi-dimensional feedback data for comprehensive calculations, covering: channel state information (CSI), signal-to-noise ratio / signal-to-interference-plus-noise ratio (SNR / SINR), beam alignment error and link blocking probability at the physical channel layer; end-to-end latency, latency jitter, throughput, packet loss rate, queue occupancy rate and link load status at the network transmission layer; semantic distortion, semantic information age, semantic similarity, semantic importance graph, and differences in semantic knowledge bases between sender and receiver at the semantic layer; task completion accuracy, quality of user experience (QoE), intent satisfaction and service level agreement (SLA) violation degree at the task intent layer; and internal feedback indicators such as the agent's own tool invocation results, introspection and reflection information, process reward signals and multi-agent collaborative consistency.

[0089] The specific optimization process can be summarized as follows: First, the aforementioned multi-source heterogeneous feedback data is time-aligned, feature-encoded, and weighted to construct a globally unified real-time network state representation. Then, based on the degree of performance deviation, the level of semantic timeliness degradation, and the amplitude of network state fluctuations, it is determined whether to trigger re-inference of the scheduling strategy. If re-inference is triggered, the relevant context is retrieved from the shared semantic knowledge base, network topology, historical scheduling trajectory, and digital twin simulation environment to locate the deep-seated causes of performance degradation and resource imbalance. Subsequently, the network agent generates multiple candidate macro-scheduling schemes, and relies on self-evaluation, self-reflection mechanisms, and multi-agent collaborative review to complete the screening, correction, and priority ranking of candidate strategies, and output the optimized macro-scheduling strategy. Before the strategy is officially issued, its future execution performance is verified through digital twin or simulation to ensure that it meets the requirements of transmission power, transmission capacity, end-to-end latency, energy consumption constraints, and SLA compliance. Finally, the verified macro-scheduling strategy is issued to the lower-level real-time control and scheduling module for implementation, and the strategy execution results are written back to the agent experience memory and shared knowledge base to achieve knowledge accumulation and iterative updates.

[0090] Therefore, relying on the autonomous closed-loop operation mechanism of "perception-memory-reasoning-action", the network intelligent agent can achieve adaptive macro-schedule optimization based on semantic value, business task intent and real-time network resource status.

[0091] Optionally, when there is only one target receiving object, the signal transmission adopts a unicast method with semantic routing that meets the requirements of latency and reliability; When there are multiple target receiving objects, the signal transmission adopts a broadcast or a semantic distribution method oriented towards multiple nodes; Optionally, the semantic completion includes: semantic vector matching, knowledge graph context completion, and logical consistency verification; The semantic vector matching involves matching the estimated semantic vector with entity, attribute, relation, and event representations in a shared semantic knowledge base to obtain candidate semantic units. The knowledge graph context completion is based on the graph relationships, event templates, scene rules and task contexts between candidate semantic units to infer semantic information that has not been explicitly transmitted or is missing due to channel influence, and obtain the completed semantic units. The logical consistency check utilizes the triple relationships and scene logic in the knowledge graph to perform consistency checks on the "entity-relationship-event-action" link in the completed semantic unit, eliminates logical conflicts, and obtains the checked semantic unit.

[0092] Optionally, the task intent reconstruction includes: identifying elements of the task intent from the semantically completed semantic units; and reconstructing the task intent based on the elements to obtain an estimated task intent.

[0093] The estimated task intent can be used to calculate semantic distortion. This estimated task intent can directly guide the execution-side agent to generate semantic action instructions and drive the task to be executed. It can also serve as a benchmark to be compared with the original structured task intent of the sending end, accurately calculating the semantic distortion. This provides a key basis for semantic distortion assessment, dynamic updating of the shared knowledge base, and optimization of network resource scheduling.

[0094] The complete process of restoring semantic vectors to the original intent is as follows: The receiving communication agent first performs semantic decoding on the received signal to obtain the estimated semantic vector; then, relying on the shared semantic knowledge base, it completes semantic matching at the entity, attribute, relationship, event, and task levels; on this basis, it completes the missing semantics and performs logical consistency verification; finally, it outputs the structured task intent, or further transforms it into natural language prompts, service call parameters, or device control commands, driving the receiving agent to execute actions that match the original intent of the sending end.

[0095] For example, if the sender only transmits core semantic features such as "ahead, vehicle abnormality, congestion, high risk," the receiver can combine traffic scene knowledge graphs to reconstruct them into complete semantics: "A traffic accident has occurred ahead; vehicles behind should be warned to slow down and detour." This reconstruction of the original intent does not involve restoring the original bitstream or reconstructing the original video frame by frame; rather, it accurately reproduces the semantic meaning that the sender expects the receiver to understand and execute.

[0096] In the semantic encoding and processing layer, the operational logic is as follows: the sending end extracts core semantic information based on the communication task issued by the intent layer and encodes it into a semantic vector; the receiving end reconstructs the communication task through a reverse decoding process. The core performance evaluation of this stage is the semantic distortion between the semantics reconstructed by the receiving end and the original intent of the sending end, rather than the bit error rate metric in traditional communication. The communication agent also follows this operational mechanism: the sending end generates and encodes the semantic vector, the receiving end reconstructs the original semantics through a reverse process, and the semantic background of both parties is unified and aligned through protocol negotiation.

[0097] Optionally, the communication scenario includes any one of the following: autonomous driving scenario, industrial internet scenario, embodied robot scenario, drone scenario, and satellite scenario.

[0098] Figure 2 A schematic diagram of the network communication device is shown, such as... Figure 2 As shown, an agent network communication device 200 based on structured task intent perception is applied to a communication scenario including an application-side agent, a network-side agent, a communication-side agent, and an execution-side agent. Its features include: The intent generation module 210 generates a structured task intent based on the task input by the application-side intelligent agent; the structured task intent includes at least: the target receiving object and semantic knowledge content. The network configuration module 220, the network-side intelligent agent performs task-oriented global coordination and resource pre-allocation 40 according to the structured task intent, and generates network configuration; The encoding and transmission module 230, guided by the network configuration and the structured task intent, determines the optimal encoding scheme; based on the optimal encoding scheme, it maps the structured task intent into a high-dimensional semantic vector and transmits the signal. The intent recovery module 240 executes the intelligent agent to obtain the modulation signal through physical perception, and generates an estimated semantic vector based on the modulation signal; it then performs semantic completion and task intent reconstruction on the estimated semantic vector based on a shared semantic knowledge base to generate an estimated task intent.

[0099] The network communication device and the network communication method are one-to-one. Their specific implementation methods and technical effects will not be elaborated further.

[0100] This embodiment discloses a specific application of a vehicle intelligent agent network communication method in an autonomous driving scenario. Based on the semantic intelligent agent communication architecture of this invention, the application achieves a high degree of autonomous cooperation between vehicles, infrastructure, and the network. It is particularly suitable for emergency avoidance scenarios when a sudden accident occurs ahead. The following vehicle, based on its intention of "accident ahead, emergency avoidance," and combined with its own position and road conditions, plans a specific avoidance trajectory. The detailed implementation process is as follows: In the key 6G scenario of autonomous driving, the semantic intelligent agent communication architecture of the present invention can be used to realize highly autonomous collaboration between vehicles, infrastructure and networks. In this architecture, the vehicle, as an embodied intelligent agent, acquires information about the complex physical environment around it in real time through the perception phase. The intent layer identifies potential dangers (such as sudden accidents ahead) based on the perceived multimodal data and sets a communication target of "emergency avoidance".

[0101] At this point, the communication-side agent invokes the shared knowledge base at the semantic layer to encode the "emergency avoidance" intention into a compact semantic vector, and then performs optimal path scheduling through the network agent. The path planning process is as follows: First, the target node set is determined based on the target receiving object and spatiotemporal constraints. In this embodiment, the target receiving object is vehicles within a 300-meter radius behind. Therefore, the requirement to "broadcast accident information to vehicles within a 300-meter radius behind" is interpreted as a semantic distribution task targeting multiple vehicle nodes within this specific spatial range, and the target node set is defined as all driving vehicles within a 300-meter radius behind.

[0102] Secondly, the network-side agent selects a suitable transmission path or set of cooperative nodes based on network connectivity, routing status, traffic load, congestion, available resources at edge nodes, and task latency requirements. At the same time, information such as channel status, SNR, and CSI is used as auxiliary inputs for the communication agent or adaptive semantic scheduling module to adjust the semantic compression ratio, transmission mode, and resource allocation, ensuring efficient transmission of semantic information.

[0103] Furthermore, in this embodiment, the receiving objects are multiple rear vehicles, therefore a broadcast or multi-node semantic distribution method is adopted. If semantic information needs to be forwarded, merged, or corrected across nodes, semantic relay, intra-network aggregation, or collaborative reasoning mechanisms are combined to reduce redundant transmissions and improve semantic transmission efficiency. It should be noted that if multiple candidate paths exist, the network-side agent prioritizes the path that ensures the correct arrival of core semantics, rather than simply choosing the path with the highest bit rate. The path selection process comprehensively considers semantic distortion risk, latency constraints, reliability requirements, resource consumption, and target task benefits.

[0104] The priority calculation in the above path selection process is performed based on the following parameters: 1. Task urgency: In this embodiment, accident alarms and collision avoidance are real-time tasks with high priority; while normal state synchronization or non-real-time knowledge updates have lower priority. Therefore, priority is given to ensuring the transmission of semantic information related to accident alarms and emergency avoidance. 2. Semantic importance: Semantic features that have a greater impact on the outcome of the emergency avoidance mission are transmitted first, such as core semantic information like accident location, danger level, and key state changes in this embodiment; 3. Semantic freshness: When the knowledge received by the following vehicle is outdated, or its semantic representation is inconsistent with the actual environmental state of the accident ahead, the priority of relevant semantic updates is increased to ensure that the accident information obtained by the following vehicle is real-time and effective. 4. Latency and Reliability Requirements: The emergency avoidance task in this embodiment is a high-latency, high-reliability task, which prioritizes the acquisition of bandwidth, power, computing and caching resources to ensure the fast and accurate transmission of semantic information; 5. Channel and resource status: Dynamically adjust semantic compression ratio, transmission mode and resource allocation when the channel fluctuates or resources are scarce to adapt to the complex autonomous driving communication environment; 6. Scope of receiving objects: In this embodiment, the semantic information is received by multiple vehicles within 300 meters behind, affecting multiple intelligent agents. Therefore, it has a higher task impact weight when prioritizing.

[0105] The aforementioned priority ranking and resource scheduling sections correspond to three specific methods: query semantic scheduling allocates communication and computing resources based on downstream tasks (i.e., emergency avoidance tasks for vehicles behind) or user query intent; semantic importance-aware scheduling allocates physical resources based on the spatiotemporal value of semantic features (such as accident location and hazard level); and adaptive semantic scheduling dynamically adjusts the semantic compression ratio and transmission mode based on channel conditions. Related work also includes task-based QoE-based resource allocation, feature importance-based power allocation, AoSI-based multi-source scheduling, and adjusting semantic compression and transmission methods based on real-time CSI fluctuations to further ensure the reliability and efficiency of semantic transmission.

[0106] In this embodiment, multiple intelligent agents collaborate, with their responsibilities clearly defined as follows: the sending agent (i.e., the vehicle involved in the accident ahead) is responsible for extracting the intent of "accident ahead, emergency avoidance" and encoding and transmitting it; the network-side agent is responsible for global path planning and resource scheduling, i.e., completing the aforementioned path planning and priority calculation operations; the receiving agent (i.e., the vehicle within 300 meters behind) is independently responsible for local execution path planning based on its own state, environmental perception, and received semantic instructions. It is important to note that although multiple intelligent agents—the sending agent, the network agent, and the receiving agent—are involved, the "execution path planning" action is only completed by the receiving agent, belonging to its local decision-making stage within its "reasoning-action" closed loop.

[0107] The specific process of semantic reconstruction and action execution at the receiving end is as follows: The receiver embodied intelligent agent (the vehicle behind) at the collaborative terminal obtains the modulation signal through physical perception; in the perception stage, semantic features are extracted and semantically completed by combining the prior knowledge graph stored in the memory stage; then in the reasoning stage, the original intention of the sending end (i.e., "accident ahead, emergency avoidance") is reconstructed based on the semantic vector and mapped into specific physical control commands.

[0108] In this embodiment, the recipient of the semantic information is the vehicle behind, and the knowledge content transmitted is the location and severity of the accident. The desired effect is to cause the vehicle behind to slow down or change its route, thus achieving emergency avoidance. Specifically, in the action phase, after receiving the semantic instruction, the vehicle behind autonomously decides on an avoidance trajectory and executes the action by combining the reasoning phase with its own memory phase of stored historical road condition experience, ultimately completing the emergency avoidance operation and avoiding a collision with the vehicle in front of the accident.

[0109] As can be seen from the above implementation process, the semantic intelligent agent communication architecture and vehicle intelligent agent network communication method of the present invention, through multi-agent collaboration and closed-loop cognitive mechanism, ensure that the key safety intention of "accident ahead, emergency avoidance" can be accurately transmitted even in the case of network congestion or poor signal, which greatly improves driving safety in autonomous driving scenarios.

[0110] This embodiment discloses a specific application of an intelligent agent network communication method in industrial internet scenarios and embodied robot scenarios. Based on the semantic intelligent agent communication architecture disclosed herein, this application is adapted to the complex long-cycle task collaboration needs of multiple embodied intelligent agents (such as robotic arms and inspection robots) in smart factories. It can effectively realize automatic error correction and robust execution of tasks when the production line fails. A typical application scenario is: the robotic arm replans the grasping path based on the feedback of "part position offset". The implementation details of this method are described in detail below in conjunction with the specific application process.

[0111] In a smart factory environment, multiple embodied intelligent agents collaborate to complete production tasks. The core foundation of their collaboration relies on the semantic intelligent agent communication architecture disclosed herein. The entire communication and task execution process forms a closed-loop cognitive mechanism of "perception-reasoning-memory-execution-feedback," with the semantic perception stage being the foundation of the entire closed loop. The specific implementation process is as follows: The terminal embodied intelligent agent (such as a robotic arm or inspection robot) first acquires raw multimodal perception data through cameras, LiDAR, IMU, tactile sensors, etc.; then, through a semantic encoder or multimodal encoder, it maps heterogeneous data such as images, point clouds, motion states, and tactile information to a unified semantic space; and then, based on the current structured task intent, it extracts task-related targets, attributes, relationships, events, positions, state changes, and environmental changes to form semantic vectors or structured semantic representations.

[0112] It is worth noting that embodied intelligent agents do not need to transmit complete video streams or complete raw sensor data, but rather transmit environmental changes or target features, thereby reducing the burden on air interface resources. In smart factory scenarios, embodied intelligent agents such as robotic arms and inspection robots need to transmit multimodal information such as vision and touch in real time when performing tasks such as parts grasping and fault inspection. Semantic communication enables them to transmit environmental changes or target features, rather than complete video streams, effectively optimizing resource consumption.

[0113] Semantic feature extraction is clearly task-oriented. The terminal does not compress all perceived information evenly, but selects key semantics based on the structured task intent. Specifically, in the robot collaboration task (robotic arm grasping, inspection robot inspection) of this embodiment, the focus is on extracting the target object (such as the part to be grasped), the operable area, the location of obstacles, the state of other intelligent agents, and action constraints. In addition, the task-oriented extraction logic in different scenarios can be used as a reference: for example, in the traffic accident alarm task, the focus is on extracting the accident vehicle, road location, congestion status, severity, scope of impact, and suggested actions; in the UAV collaborative inspection task, the focus is on extracting abnormal targets, target locations, motion states, regional relationships, and collaborative node requirements.

[0114] When a production line malfunctions (such as a part misalignment in this embodiment, preventing the robotic arm from completing its preset grasping action), the network agent plays a crucial scheduling role. Through modular division multiple access (MDA) technology in the distributed autonomy and collaboration layer, it supports multiple robots simultaneously interacting with high-dimensional semantic features within the same resource block, ensuring efficient and conflict-free semantic information transmission between multiple agents. At this point, the semantic feedback system immediately drives the inference phase to replan the route and initiates an automatic task error correction mechanism. Throughout this process, the network agent's path planning and resource scheduling are key to ensuring efficient error correction.

[0115] The path planning and resource scheduling of network intelligent agents need to be comprehensively executed in combination with production task requirements, agent collaboration constraints, and various network and resource parameters. First, the set of embodied intelligent agents participating in the collaboration is identified. In this embodiment, the collaborative task is part picking and fault handling on the production line. The embodied intelligent agents participating in the collaboration include multiple robotic arms and inspection robots. Therefore, the relevant task requirements are parsed into collaborative semantic interaction tasks for this type of intelligent agent. Subsequently, the network intelligent agent selects a suitable transmission path or set of collaborative nodes based on network connection relationships, routing status, traffic load, congestion, available resources of edge servers, and task latency requirements. At the same time, it relies on modular division multiple access technology to optimize the efficiency of high-dimensional semantic feature interaction, ensuring that multiple robots complete semantic interaction synchronously.

[0116] For multi-agent collaborative scenarios, a semantic interaction and collaborative scheduling approach oriented towards multiple nodes is adopted. If semantic information needs to be forwarded, fused, or corrected across agents, semantic relay, intra-network aggregation, or collaborative reasoning mechanisms are combined to reduce redundant transmissions and improve semantic transmission efficiency. It is worth noting that during path selection, network agents prioritize paths that ensure the correct transmission of core fault and task semantics, rather than simply pursuing bit rate. Key considerations include semantic distortion risk, latency constraints, reliability requirements, resource consumption, and production task benefits. The scheduling order is determined through priority calculation.

[0117] This embodiment employs three specific scheduling methods: query semantic scheduling allocates communication and computing resources based on downstream production tasks (fault handling, part retrieval and error correction) or user query intent; semantic importance-aware scheduling allocates physical resources based on the spatiotemporal value of semantic features; and adaptive semantic scheduling dynamically adjusts the semantic compression ratio and transmission mode based on channel conditions. Furthermore, related auxiliary scheduling work includes resource allocation based on task QoE, power allocation based on feature importance, multi-source scheduling based on AoSI, and adjusting semantic compression and transmission methods based on real-time CSI fluctuations to further improve the reliability and efficiency of semantic transmission.

[0118] In this embodiment, the responsibilities of the multiple agents are clearly defined, forming a collaborative closed loop: the application agent is responsible for offloading complex computational logic to the edge server, simplifying the computational burden on the embodied agent; the network agent is responsible for global path planning, resource scheduling, and semantic interaction management among the multiple agents, coordinating various scheduling tasks; the receiving embodied agent (robotic arm, inspection robot) is independently responsible for local execution path planning and task adjustment based on its own state, production line perception data, and received semantic instructions. It should be noted that the "execution path replanning" action is only completed by the receiving embodied agent, belonging to the local decision-making link in its "reasoning-action" closed loop, ensuring the autonomy and flexibility of task execution.

[0119] Semantic reconstruction and action execution at the receiving end are key steps in achieving automatic task error correction. The specific process is as follows: the embodied intelligent agent (robotic arm, inspection robot) at the collaborative terminal acquires the modulated signal through physical perception; semantic features are extracted in the perception stage, and semantic completion of the features is performed by combining the industrial process knowledge graph stored in the memory stage, and the intelligent agent is supported to perform analogical reasoning based on this knowledge graph; then, in the reasoning stage, the original intent of the sending end (such as "part position offset, replan the grasping path") is reconstructed based on the semantic vector, and mapped into specific physical control commands.

[0120] In this embodiment, the semantic information is received by embodied intelligent agents such as robotic arms and inspection robots. The core knowledge content transmitted includes part position offset information, fault type, and fault location. The desired effect is that the robotic arm replans its grasping path, the inspection robot accurately locates the fault, and ultimately achieves automatic error correction and robust execution of the production task. Specifically, when the robotic arm finds that the preset grasping action cannot be completed (part position offset) during the execution phase, the semantic feedback system drives the reasoning phase to replan the route. The robotic arm combines the industrial process knowledge graph stored in its memory phase and historical execution experience to autonomously decide on a new grasping path and execute it, successfully completing the part grasping task and verifying the effectiveness of this method.

[0121] In summary, the semantic agent communication architecture and agent network communication method disclosed herein, through multi-agent collaboration and closed-loop cognitive mechanisms, combined with modular division multiple access technology and industrial process knowledge graphs, effectively solve the problems of low efficiency of multi-agent collaboration and untimely fault handling in industrial Internet scenarios. Even under network congestion or poor signal conditions, it can accurately convey key task intentions such as "part position offset, replanning the grasping path", significantly improving the robustness and efficiency of production tasks in industrial Internet scenarios.

[0122] This embodiment, based on the 6G intelligent simplified semantic communication architecture of this application, takes emergency search and rescue scenarios such as earthquakes and floods as an example to construct a cluster collaborative search and rescue system composed of dozens to hundreds of drones. The drones simultaneously act as communication-side intelligent agents and embodied execution intelligent agents, performing tasks such as environmental perception of the disaster area, target recognition, information transmission, and collaborative search and rescue. The implementation process is as follows: In extreme disaster relief environments, ground communication infrastructure is severely damaged, and wireless channels are significantly affected by mountain obstruction, electromagnetic interference, and multipath fading, resulting in unstable channel quality. Furthermore, the high-speed movement and dynamic reorganization of drones lead to frequent changes in network topology. Traditional bitstream-based high-definition video and image backhaul modes consume large amounts of bandwidth and suffer severe packet loss, making it difficult to ensure reliable transmission of search and rescue information.

[0123] In this solution, each UAV first enters the perception layer, where it collects images of the disaster area using onboard vision and infrared sensors. Instead of transmitting the original high-definition image bitstream back, it extracts core semantic features for search and rescue, such as trapped personnel, collapsed buildings, dangerous areas, and rescue channels, and encodes them into lightweight semantic vectors. Upon entering the intent layer, the application-side agent uniformly issues a global task intent of "prioritizing the search and rescue of trapped personnel, avoiding secondary disasters, and rapidly transmitting information back." Through an end-to-end intent alignment mechanism, the search and rescue targets, priorities, and execution logic of all UAVs in the cluster are unified, eliminating ambiguities in task understanding among multiple UAVs. At the collaborative scheduling layer, the network-side agent, based on a semantic-aware resource scheduling strategy, collects in real time the remaining battery power, location information, and whether trapped personnel have been identified for each UAV, and prioritizes the UAVs: for UAVs with low battery power but having already detected key disaster targets, channel time slots, power resources, and beam resources are allocated first to ensure priority transmission of their semantic vectors; for inspection UAVs with sufficient battery power but having not detected targets, transmission resource consumption is reduced, achieving refined and intelligent resource scheduling.

[0124] After the semantic vector is transmitted via a wireless link, the ground receiving agent performs semantic decoding on the received signal. In the event of strong interference or loss of some semantic features, it relies on a shared semantic knowledge base to complete semantic matching of entities, events, and tasks. Through semantic completion and logical consistency verification, it reconstructs the complete search and rescue intent and outputs structured search and rescue instructions or control parameters to guide subsequent rescue operations.

[0125] This embodiment achieves collaborative scheduling and intent alignment driven by semantics in disaster relief environments characterized by low bandwidth, strong interference, and dynamic topology changes. It eliminates the need to transmit massive amounts of raw data and relies solely on core semantic vectors to complete collaborative search and rescue missions, significantly reducing communication overhead and effectively improving the reliability and success rate of mission execution for large-scale UAV swarms in extreme scenarios.

[0126] This embodiment focuses on the application of satellite semantic communication intelligent agents in a space-ground collaborative manner. Based on the 6G intelligent simplified semantic communication architecture of this application, and taking scenarios such as wide-area coverage of space-ground collaboration, emergency communication in remote areas, and deep space exploration as examples, a semantic communication system is constructed, consisting of a low-Earth orbit satellite constellation, geostationary orbit satellites, and ground terminals. The satellite cluster acts as the core communication-side intelligent agent, simultaneously undertaking embodied perception and semantic transmission tasks. This addresses the technical pain points of traditional satellite communication, such as limited bandwidth, large latency fluctuations, and inefficient space-ground collaboration. The implementation process is as follows: In satellite-to-ground communication scenarios, communication links between satellites and ground terminals, and between satellites themselves, are affected by factors such as cosmic radiation, ionospheric scintillation, and atmospheric attenuation, resulting in extremely unstable channel quality. Furthermore, the high-speed movement of low-orbit satellites leads to frequent switching of satellite-to-ground links and dynamic changes in network topology. At the same time, traditional satellite communication uses bitstream transmission mode, which requires the transmission of massive amounts of remote sensing images, environmental monitoring data, and terminal control commands, consuming a large amount of onboard bandwidth resources. Moreover, it is prone to data packet loss and command misinterpretation due to link interference, making it difficult to meet the real-time and reliability requirements of scenarios such as wide-area collaboration and emergency response.

[0127] In this solution, the satellite constellation, acting as a collection of communication-side intelligent agents and embodied intelligent agents, systematically completes semantic perception, intent alignment, collaborative scheduling, and semantic transmission according to the layered logic of the 6G intelligent and simplified architecture. The specific implementation process is as follows: 1. Semantic Extraction at the Perception Layer: Each satellite collects environmental data (such as meteorological parameters, geological features, and terminal distribution), emergency event information (such as disasters in remote areas and distress signals at sea), and onboard equipment status (such as remaining power and link quality) within its coverage area through airborne remote sensing sensors and signal monitoring modules. Instead of transmitting massive bitstreams of raw remote sensing images and monitoring data, core semantic features are extracted, such as "earthquake in remote mountainous areas, people trapped, signal interruption," "insufficient onboard power, strong link interference," and "distressed targets at sea, coordinate positioning," and encoded into lightweight semantic vectors, significantly reducing the bandwidth usage of satellite-to-ground transmission.

[0128] 2. Global Alignment at the Intent Layer: Based on the global task requirements of satellite-ground collaboration, the application-side intelligent agent issues a unified task intent that prioritizes emergency communication, accurately transmits core semantics, and dynamically adapts to link status. Through an end-to-end intent alignment mechanism, it completes intent coordination between the satellite cluster and the ground control center, as well as between satellites. This clarifies the task division of each satellite (e.g., low-Earth orbit satellites are responsible for near-Earth semantic perception and terminal access, while geostationary orbit satellites are responsible for semantic relay and wide-area coverage), eliminates semantic ambiguities between satellites and ground, and between satellites, and ensures that all intelligent agents work collaboratively around the same task objective.

[0129] 3. Semantic-Aware Scheduling at the Collaboration Layer: As the core scheduling unit for satellite-ground collaboration, the network-side intelligent agent uses a semantic-aware resource scheduling strategy to collect the real-time operational status of each satellite (remaining power, link quality, semantic transmission priority) and dynamically allocate resources based on the communication needs of ground terminals. The principle of "semantic value priority and efficient resource utilization" is strictly followed. For satellites transmitting critical semantics (such as distress signals and disaster relief instructions), those with insufficient onboard power, or those with poor link quality, priority is given to allocating onboard channel resources, power resources, and relay links to ensure the priority transmission of core semantic vectors. For satellites transmitting non-critical semantics and those with good equipment status, the resource allocation ratio is adjusted reasonably to achieve intelligent and simplified scheduling of onboard resources and avoid resource waste.

[0130] 4. Semantic Transmission and Intent Reconstruction: The satellite transmits the encoded semantic vector to the ground control center or target terminal via satellite-to-ground links and inter-satellite links. The receiving communication agent performs semantic decoding on the received signal to obtain the estimated semantic vector. To address the issue of partial semantic loss caused by link interference, the satellite relies on a shared satellite-to-ground semantic knowledge base (including satellite-to-ground communication protocols, emergency scenario semantic databases, satellite equipment parameters, etc.) to complete entity, attribute, relationship, and task semantic matching. Through semantic completion technology and logical consistency verification, the complete task intent is reconstructed.

[0131] For example, when a low-orbit satellite detects the core semantic features of "earthquake and people trapped" in a remote mountainous area, some semantic information is lost during transmission due to ionospheric interference. The ground receiving end, combined with the emergency rescue semantic knowledge base, can complete and reconstruct the complete intent of "an earthquake has occurred in a remote mountainous area, many houses have collapsed, people are trapped, and it is necessary to urgently dispatch rescue forces and open emergency communication links," thus driving the ground rescue terminal and other cooperating satellites to perform corresponding actions.

[0132] Meanwhile, the ground control center generates semantic feedback vectors from the execution results of semantic action commands (such as the dispatch status of rescue forces and the status of emergency link opening), and sends them back to the satellite cluster. This allows the communication-side intelligent agents to perform semantic deviation assessment and synchronize updates to the shared knowledge base, enables the network-side intelligent agents to optimize subsequent space-ground resource scheduling strategies, and allows the application-side intelligent agents to assess the degree of achievement of the initial intent, thus forming a closed-loop optimization of space-ground collaboration.

[0133] This embodiment uses satellite clusters as communication-side intelligent agents and embodied intelligent agents, relying on core technologies such as semantic extraction, intent alignment, and semantic awareness scheduling. It breaks through the limitations of traditional satellite communication that relies on bit stream transmission. In satellite-to-ground communication scenarios with limited bandwidth, unstable channels, and dynamic topology changes, it achieves reliable transmission of core semantics and satellite-to-ground collaboration with minimal communication overhead. This significantly improves the intelligence, real-time performance, and reliability of satellite communication systems and can be widely applied to satellite-to-ground collaboration scenarios such as emergency communication, deep space exploration, and wide-area environmental monitoring.

[0134] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0135] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0136] The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the communication method provided in the embodiments of this disclosure.

[0137] The readable storage medium is a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to perform the communication method provided in the embodiments of this disclosure.

[0138] The computer program product includes a computer program that, when executed by a processor, implements the communication methods provided in the embodiments of this disclosure.

[0139] Figure 3 A schematic block diagram of an example electronic device 500 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0140] like Figure 3 As shown, device 500 includes a computing unit 501, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 502 or a computer program loaded from storage unit 508 into random access memory (RAM) 503. RAM 503 may also store various programs and data required for the operation of device 500. The computing unit 501, ROM 502, and RAM 503 are interconnected via bus 504. Input / output (I / O) interface 505 is also connected to bus 504.

[0141] Multiple components in device 500 are connected to I / O interface 505, including: input unit 506, such as keyboard, mouse, etc.; output unit 507, such as various types of monitors, speakers, etc.; storage unit 508, such as disk, optical disk, etc.; and communication unit 509, such as network card, modem, wireless transceiver, etc. Communication unit 509 allows device 500 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0142] The computing unit 501 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 501 performs the various methods and processes described above, such as communication methods. For example, in some embodiments, the communication method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program may be loaded and / or installed on device 500 via ROM 502 and / or communication unit 509. When the computer program is loaded into RAM 503 and executed by the computing unit 501, one or more steps of the communication method described above may be performed. Alternatively, in other embodiments, the computing unit 501 may be configured to perform the communication method by any other suitable means (e.g., by means of firmware).

[0143] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0144] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0145] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0146] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0147] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0148] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0149] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0150] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for intelligent agent network communication based on structured task intent awareness, applied to communication scenarios including application-side intelligent agents, network-side intelligent agents, communication-side intelligent agents, and execution-side intelligent agents, characterized in that, include: The application-side agent generates structured task intents based on task input; The structured task intent includes at least: the target receiving object and semantic knowledge content; The network-side intelligent agent performs task-oriented global coordination and resource pre-allocation based on the structured task intent, and generates network configuration; Guided by the network configuration and the structured task intent, the communication-side agent determines the optimal encoding scheme; based on the optimal encoding scheme, it maps the structured task intent into a high-dimensional semantic vector and transmits the signal. The execution-side agent acquires the modulation signal through physical perception, generates an estimated semantic vector based on the modulation signal, and performs semantic completion and task intent reconstruction on the estimated semantic vector based on a shared semantic knowledge base to generate an estimated task intent.

2. The intelligent agent network communication method according to claim 1, characterized in that, Also includes: Based on the estimated task intent, generate executable semantic action instructions; Based on the execution result of the semantic action instruction, a semantic feedback vector containing the execution state is generated; The semantic action instructions include any one of the following: physical control instructions, natural language prompts, and service call parameters; The semantic feedback vector is used by the communication-side agent to perform semantic deviation evaluation and synchronize the update of the shared knowledge base; the semantic feedback vector is used by the network-side agent to perceive the global task achievement status and, based on this, to re-infer and optimize the subsequent macro-scheduling strategy. The semantic feedback vector is used by the application-side agent to evaluate the degree of achievement of the initial intent, thereby forming a complete closed-loop optimization.

3. The intelligent agent network communication method according to claim 1, characterized in that, The application-side agent generates a structured task intent based on the task input, including: The application-side agent generates a preliminary task intent based on the task input; The application-side agent decomposes the initial task intent into a structured task intent based on historical task knowledge. The structured task intent also includes at least one of the following: expected effect, spatiotemporal constraints, reliability, semantic distortion tolerance, resource overhead, and security level.

4. The intelligent agent network communication method according to claim 3, characterized in that, The target receiving object can be any of the following: a specified intelligent agent, multiple specified intelligent agents, an intelligent agent within a specified spatial range, an intelligent agent with a specified role type, or an intelligent agent within a specified task group.

5. The intelligent agent network communication method according to claim 1, characterized in that, The network configuration includes at least one of the following: Optimal macro-path, task priority, and resource pre-allocation scheme; The optimal macroscopic path includes a semantic routing strategy; The resource pre-allocation scheme includes at least one of the following: intra-network aggregation scheme, optimized distribution scheme, and intelligent distribution scheme.

6. The intelligent agent network communication method according to claim 5, characterized in that, The task priority is calculated based on at least one of the following indicators: task urgency, semantic importance, semantic freshness, latency requirements, reliability requirements, channel and resource status, and target range.

7. The intelligent agent network communication method according to claim 1, characterized in that, The communication-side agent, guided by the network configuration and the structured task intent, determines the optimal encoding scheme, including: Guided by the network configuration and the structured task intent, the communication-side agent acquires the physical channel state; and invokes the shared semantic knowledge base to determine the optimal encoding scheme based on the physical channel state. The signal transmission includes: transmitting signals through a distributed autonomy and cooperation layer.

8. The intelligent agent network communication method according to claim 1, characterized in that, When there is only one target receiving object, the signal transmission adopts a unicast method with semantic routing that meets the requirements of latency and reliability; When there are multiple target receiving objects, the signal transmission adopts a broadcast or a semantic distribution method oriented towards multiple nodes.

9. The intelligent agent network communication method according to claim 1, characterized in that, The semantic completion includes: semantic vector matching, knowledge graph context completion, and logical consistency verification; The semantic vector matching involves matching the estimated semantic vector with entity, attribute, relation, and event representations in a shared semantic knowledge base to obtain candidate semantic units. The knowledge graph context completion is based on the graph relationships, event templates, scene rules and task contexts between candidate semantic units to infer semantic information that has not been explicitly transmitted or is missing due to channel influence, and obtain the completed semantic units. The logical consistency check utilizes the triple relationships and scenario logic in the knowledge graph to perform consistency checks on the "entity-relationship-event-action" link in the completed semantic unit, eliminates logical conflicts, and obtains the checked semantic unit.

10. The intelligent agent network communication method according to claim 1, characterized in that, The task intent reconstruction includes: identifying elements of the task intent from semantically completed semantic units; and reconstructing the task intent based on the elements to obtain an estimated task intent.

11. The intelligent agent network communication method according to claim 1, characterized in that, The communication scenarios include any one of the following: autonomous driving scenario, industrial internet scenario, embodied robot scenario, drone scenario, and satellite scenario.

12. A smart agent network communication device based on structured task intent perception, applied to a communication scenario including application-side smart agents, network-side smart agents, communication-side smart agents, and execution-side smart agents, characterized in that, include: The intent generation module enables the application-side agent to generate structured task intents based on task input. The structured task intent includes at least: the target receiving object and semantic knowledge content; The network configuration module involves the network-side intelligent agent performing task-oriented global coordination and resource pre-allocation based on the structured task intent to generate the network configuration. The encoding and transmission module, guided by the network configuration and the structured task intent, determines the optimal encoding scheme; based on the optimal encoding scheme, it maps the structured task intent into a high-dimensional semantic vector and transmits the signal. The intent recovery module involves the execution-side agent acquiring a modulation signal through physical perception, generating an estimated semantic vector based on the modulation signal, and performing semantic completion and task intent reconstruction on the estimated semantic vector based on a shared semantic knowledge base to generate an estimated task intent.

13. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-11.

14. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-11.

Images