Vehicle intelligent agent control method and device, electronic equipment, vehicle and medium

Abstract

The embodiment of the invention provides a vehicle intelligent agent control method and device, electronic equipment, a vehicle and a medium, and the method comprises the steps that a central intelligent agent receives task analysis information sent by a cabin intelligent agent, and the task analysis information is obtained by carrying out semantic analysis on an intelligent agent cooperative control task after the cabin intelligent agent receives the intelligent agent cooperative control task; splitting the agent cooperative control task into a plurality of sub-tasks according to the task analysis information, allocating an execution agent to each sub-task, and generating a task execution plan; the execution intelligent agent comprises an intelligent driving intelligent agent and a cabin intelligent agent; and sending a task execution request corresponding to each sub-task to each execution agent according to the task execution plan, so that each execution agent executes the corresponding sub-task according to the task execution request. According to the invention, the cabin intelligent agent, the intelligent driving intelligent agent and the central intelligent agent cooperatively control the user intention of the task around the intelligent agents and cooperatively execute the task, so that the task execution efficiency and effect are improved.

Description

Technical Field

[0001] This application relates to the field of vehicle control, and more particularly to a vehicle intelligent agent control method, device, electronic device, vehicle, and medium. Background Technology

[0002] Currently, autonomous driving systems and smart cockpit systems are often supported by independent hardware platforms and software stacks, operating in their own isolated control domains. This fragmented design prevents the autonomous driving system and smart cockpit system from working together, leading to inconsistent vehicle behavior and fragmented user experiences during driving. Summary of the Invention

[0003] This application provides a vehicle intelligent agent control method, device, electronic device, vehicle, and medium, aiming to improve the technical problem in the prior art where the autonomous driving system and the intelligent cockpit system operate independently and cannot work together.

[0004] A vehicle intelligent agent control method, comprising: The central intelligent agent receives task parsing information sent by the cockpit intelligent agent. The task parsing information is obtained by the cockpit intelligent agent performing semantic parsing on the intelligent agent collaborative control task after receiving it. The central intelligent agent breaks down the intelligent agent collaborative control task into multiple sub-tasks based on the task parsing information, assigns an execution intelligent agent to each sub-task, and generates a task execution plan; the execution intelligent agents include the intelligent driving intelligent agent and the cockpit intelligent agent, and the task execution plan is used to instruct each execution intelligent agent to execute the corresponding sub-task; The central agent sends task execution requests corresponding to each subtask to each of the execution agents according to the task execution plan, so that each of the execution agents can execute the corresponding subtask according to the task execution request.

[0005] In this embodiment, for intelligent agent collaborative control tasks requiring multiple intelligent agents to work together, the cockpit intelligent agent performs semantic parsing to obtain task parsing information. Then, the central intelligent agent can break down the intelligent agent collaborative control task into multiple sub-tasks based on the task parsing information and assign different sub-tasks to different executing intelligent agents (including the intelligent driving agent and the cockpit intelligent agent) for collaborative completion. In this process, the cockpit intelligent agent possesses complete language understanding and generation capabilities. Using the cockpit intelligent agent for semantic parsing allows for rapid understanding of the user's complex task intent without affecting the real-time performance of the collaborative control task, greatly enhancing the usability and user experience of the cockpit intelligent agent. The central intelligent agent can translate and schedule tasks, enabling deep semantic integration between the cockpit intelligent agent and the intelligent driving agent, rather than a state of isolated, non-interfering operation of physical modules. Through the central control of the central intelligent agent, the cockpit intelligent agent and the intelligent driving agent can achieve the goal of collaboratively executing tasks in the most efficient way, based on the user's intent in the intelligent agent collaborative control task, thus improving task execution efficiency and effectiveness.

[0006] Furthermore, the task parsing information sent by the receiving cockpit intelligent agent includes: After the cockpit agent performs semantic parsing on the received agent collaborative control task, extracts the user intent, task entity, and constraints in the agent collaborative control task, generates task parsing information based on the user intent, task entity, and constraints, and encapsulates the task parsing information into a structured semantic task package, it receives the structured semantic task package containing the task parsing information sent by the cockpit agent.

[0007] In this embodiment, user input access and semantic parsing can be achieved through a cockpit intelligent agent. Specifically, after the user sends an intelligent agent collaborative control task to the cockpit intelligent agent via preset input methods such as voice, touch, or a combination of voice and vision, the cockpit intelligent agent performs semantic understanding on the user-inputted intelligent agent collaborative control task, extracting the user intent, task entities, and constraints. Based on the extracted user intent, task entities, and constraints, it generates task parsing information, encapsulates the task parsing information into a structured semantic task package, and sends it to the central intelligent agent. In this embodiment, the cockpit intelligent agent can parse fuzzy natural language and output a structured representation of the user intent to the central intelligent agent for translation and scheduling, ensuring the semantic accuracy of the intelligent agent collaborative control task.

[0008] Further, the step of dividing the agent-coordinated control task into multiple sub-tasks based on the task parsing information, assigning an executing agent to each sub-task, and generating a task execution plan includes: Based on the task parsing information, the vehicle's surrounding environment information, and the agent capability graph associated with each of the executing agents, the agent collaborative control task is divided into multiple sub-tasks. The agent capability graph includes the executable task types corresponding to each of the executing agents. Assign a corresponding execution agent based on the task type of the subtask; Based on the task parsing information, determine the task execution order and task dependency relationship corresponding to each of the sub-tasks, and generate a task execution plan for all the sub-tasks based on the task execution order and task dependency relationship.

[0009] In this embodiment, a central intelligent agent decomposes and plans the path for the collaborative control task, thereby sending different sub-tasks from the translated task execution plan to different executing intelligent agents. This allows all executing intelligent agents to collaboratively execute different sub-tasks, improving task execution efficiency and effectiveness. Specifically, the central intelligent agent, based on the received task parsing information, vehicle surrounding environment information, and the intelligent agent capability map that corresponds one-to-one with different executing intelligent agents, breaks down the collaborative control task into multiple sub-tasks. For each sub-task, based on the task type of the sub-task and the executable task types of the executing intelligent agents pre-specified in the intelligent agent capability map, a specific executing intelligent agent can be assigned to each sub-task. Then, based on the user intent, task entity, and constraints corresponding to the collaborative control task in the task parsing information, the task execution order and task dependencies of each sub-task can be determined. Furthermore, based on the task execution order and task dependencies of different sub-tasks, the central intelligent agent generates a task execution plan for managing all sub-tasks.

[0010] Further, the step of sending task execution requests corresponding to each of the sub-tasks to each of the execution agents according to the task execution plan, so that each of the execution agents can execute the corresponding sub-tasks according to the task execution requests, includes: Win Pick The task execution order and task dependencies in the task execution plan determine the target task from all the subtasks, and the execution agent assigned to the target task is determined as the target agent; The system publishes a task execution request for the target task to the target agent and transmits task association information corresponding to the target task to the target agent, so that the target agent can execute the target task according to the task association information.

[0011] In this embodiment, the scheduling and execution of the task execution plan are achieved through the collaboration of the central intelligent agent and the execution intelligent agent. Specifically, the central intelligent agent determines the target task to be executed from all its agents based on the task execution order and task dependencies. Then, the central intelligent agent issues a task execution request for the target task to the target intelligent agent. At the same time, it transmits relevant task-related information (such as map, user preferences, vehicle status, and other contextual information) to the target intelligent agent through the A2A protocol. After that, the execution intelligent agent can execute the target task according to the above-mentioned task-related information and the task execution request.

[0012] Further, the step of issuing a task execution request for the target task to the target agent and transmitting task association information corresponding to the target task to the target agent, so that the target agent can execute the target task according to the task association information, includes: When the target intelligent agent is an intelligent driving agent, a task execution request for the target task is published to the intelligent driving agent via the A2A protocol, and task association information corresponding to the target task is transmitted to the intelligent driving agent. This allows the intelligent driving agent to determine the path planning and control data corresponding to the target task based on the task association information, and then generate at least one structured execution instruction based on the path planning and control data and a preset structured vocabulary. The structured execution instruction is then output to the vehicle control module for execution. The preset structured vocabulary includes several structured vehicle control instructions, and each structured execution instruction is one of the vehicle control instructions.

[0013] In this embodiment, after receiving a task execution request, the intelligent driving agent generates corresponding path planning control data based on the task execution request and task association information corresponding to the sub-task, such as "keeping the lane," "speed limit driving," and "lane change preparation." Then, the path planning data is used to generate structured execution instructions according to a preset structured vocabulary and output to the vehicle control module. Understandably, in this embodiment, the vehicle control module can directly execute the structured execution instructions upon receiving them. The preset structured vocabulary contains only a limited number of structured execution instructions corresponding to a limited number of control actions. These structured execution instructions uniformly adopt a structured format, making them simple and less prone to errors. For example, structured execution instructions could be: "turn left," "speed 30," etc. This application uses a dedicated intelligent driving agent to independently handle the output of vehicle control instructions, and uses a preset structured vocabulary to compress the decoding space of vehicle control instructions, significantly reducing inference latency and ensuring millisecond-level real-time response capabilities during actual vehicle operation. Simultaneously, the space for outputting vehicle control instructions has structural rigidity, improving control stability and safety in complex environments and meeting automotive-grade high reliability requirements. Furthermore, this embodiment enables collaborative information transmission between the central intelligent agent and the intelligent driving agent based on the A2A protocol. The communication specifications built on the A2A protocol provide a standard interface for subsequent system upgrades and model replacements. The intelligent driving agent can be replaced with a higher-capability model or a task-specific model as needed without reconstructing the entire system, greatly improving the maintainability and functional expansion capabilities throughout the system's lifecycle. In the process of realizing the integrated vehicle-cabin-driver task, it has significant advantages in system performance improvement, enhanced interactive experience, and engineering deployment.

[0014] Further, the step of issuing a task execution request for the target task to the target agent and transmitting task association information corresponding to the target task to the target agent, so that the target agent can execute the target task according to the task association information, includes: When the target intelligent agent is a cockpit intelligent agent, the task execution request of the target task is published to the cockpit intelligent agent through the A2A protocol, and the task association information corresponding to the target task is transmitted to the cockpit intelligent agent so that the cockpit intelligent agent can determine and execute the voice broadcast operation and / or cockpit feedback operation corresponding to the target task according to the task association information.

[0015] In this embodiment, after receiving a task execution request, the cockpit agent will synchronously perform voice broadcasting and / or cockpit feedback operations for the target task based on the task execution request and task association information corresponding to the sub-task. For example, it may provide feedback to the user regarding the target task: "A route has been planned for you, estimated to take 30 minutes," and wait for further input or changes from the user. Thus, through the collaboration of the central agent and the cockpit agent, the target task is executed. Furthermore, this embodiment uses the A2A protocol for collaborative information transmission between the central agent and the cockpit agent. The communication standard built on the A2A protocol provides a standard interface for subsequent system upgrades and model replacements. The cockpit agent can be replaced with a higher-capability model or a task-specific model as needed without reconstructing the entire system, greatly improving maintainability and functional expansion capabilities throughout the system's lifecycle. In achieving integrated vehicle-cabin-driver tasks, it offers significant advantages in system performance improvement, enhanced interactive experience, and engineering deployment.

[0016] Furthermore, after the central intelligent agent sends task execution requests corresponding to each subtask to each of the executing intelligent agents according to the task execution plan, so that each of the executing intelligent agents can execute the corresponding subtask according to the task execution request, the process further includes: The central agent receives the task execution status fed back by each of the executing agents through the A2A protocol, and updates the task execution plan according to the task execution status.

[0017] In this embodiment, the task execution status corresponding to each subtask will be transmitted back to the central intelligent agent in real time via the A2A protocol, so as to facilitate the scheduling and updating of other subtasks in the task execution plan.

[0018] Furthermore, after the central intelligent agent sends task execution requests corresponding to each subtask to each of the executing intelligent agents according to the task execution plan, so that each of the executing intelligent agents can execute the corresponding subtask according to the task execution request, the process further includes: When the central agent receives a task change request sent by the cockpit agent via the A2A protocol, the central agent obtains the change intent in the task change request; the task change request is generated by the cockpit agent after semantically parsing the received change information and obtaining the change intent based on the parsing. The central intelligent agent determines the task to be changed from all the sub-tasks according to the change intention, and determines the related tasks from all the sub-tasks according to the task execution order and task dependencies in the task execution plan, and then updates the changed task and related tasks in the task execution plan.

[0019] In this embodiment, dynamic coordination of intelligent agent collaborative control tasks can be achieved through multi-turn semantic interactions between the cockpit intelligent agent and the user, as well as information interaction with the central intelligent agent via the A2A protocol. Specifically, when the user issues a task change request during the execution of an intelligent agent collaborative control task to make a new request or change the current user intent (such as "go to the charging station first"), the cockpit intelligent agent will semantically parse the task change request to obtain the change intent and then notify the central intelligent agent. Subsequently, the central intelligent agent determines the corresponding change task based on the change intent and updates and adjusts the task execution plan in real time, such as calling or terminating the change task or related tasks to ensure the stable execution of the task chain. In the above process, the multi-turn interaction process between the cockpit intelligent agent and the user supports memory mechanisms and context references, enhancing task persistence and user experience.

[0020] Furthermore, after the central intelligent agent sends task execution requests corresponding to each subtask to each of the executing intelligent agents according to the task execution plan, so that each of the executing intelligent agents can execute the corresponding subtask according to the task execution request, the process further includes: Upon receiving a task anomaly warning message sent by the intelligent driving agent via the A2A protocol, the central intelligent agent determines the abnormal task from all the sub-tasks based on the task anomaly warning message; the task anomaly warning message is identified by the intelligent driving agent based on real-time vehicle information and surrounding environment information. When the task abnormality warning information is a preset emergency information, the central intelligent agent sends an emergency braking instruction to the intelligent driving agent, so that the intelligent driving agent can generate a structured execution instruction corresponding to the emergency braking instruction according to a preset structured vocabulary, and output the structured execution instruction to the vehicle control module for execution; When the task anomaly warning message is not a preset emergency message, the central agent sends a replanning request containing the abnormal task to the cockpit agent, so that the cockpit agent can respond to the replanning request and interact with the user to perform the replanning operation.

[0021] In this embodiment, through the coordinated operation of the cockpit intelligent agent, the intelligent driving agent, and the central intelligent agent, vehicle safety and system recoverability can be ensured in the event of anomalies during task execution. Even if the cockpit intelligent agent crashes or misjudges semantics, it will not directly affect the control safety of the intelligent driving agent. Conversely, if the driving agent reports a task anomaly warning (such as an impassable environment), the central intelligent agent can block the corresponding emergency to ensure vehicle safety, or send the warning back to the cockpit intelligent agent, prompting the user to re-interact and process it, thereby improving the overall fault tolerance of the system and enhancing its safety robustness and fault isolation capabilities.

[0022] Furthermore, before the central intelligent agent receives the task parsing information sent by the cockpit intelligent agent, it includes: After confirming that the central intelligent agent, the intelligent driving agent, and the cockpit intelligent agent have all started, the central intelligent agent receives the registration requests sent to it by the intelligent driving agent and the cockpit intelligent agent through the A2A protocol, and responds to the registration requests to establish an intelligent agent routing table. The intelligent agent routing table includes the task execution status and the intelligent agent capability map corresponding to the intelligent driving agent and the cockpit intelligent agent respectively.

[0023] This embodiment proposes an A2A protocol system for intelligent cockpits and autonomous driving, standardizing the semantic representation structure, control primitives, and task state feedback mechanisms among multiple agents. This enables the system to possess excellent scalability and module replacement capabilities, significantly outperforming traditional end-to-end black-box models. Specifically, after the system powers on, the cockpit agent, the intelligent driving agent, and the central agent start up respectively. The cockpit agent and the intelligent driving agent register with the central agent through the A2A protocol standard interface, declaring their own capabilities (such as semantic reasoning, path planning, control output, etc.) and resource status (such as the number of CPU cores, memory, inference latency, computing load, energy consumption budget, etc.). Then, the central agent establishes an agent routing table and maintains an agent capability graph (the agent capability graph includes, but is not limited to, agent state, task execution state, and vehicle environment state). The agent capability graph can be used for subsequent task scheduling decisions and maintaining load balancing.

[0024] Furthermore, after the central intelligent agent sends task execution requests corresponding to each subtask to each of the executing intelligent agents according to the task execution plan, so that each of the executing intelligent agents can execute the corresponding subtask according to the task execution request, the process further includes: After receiving the task completion information sent by the executing agent, the central agent updates the task execution status in the agent routing table according to the task completion information and performs a computing resource release operation.

[0025] In this embodiment, resources can be released promptly after the task is completed, thereby reducing system energy consumption and computing resource usage, and improving system throughput and maintenance efficiency. Specifically, after a subtask is completed, the executing agent corresponding to the subtask reports the task completion information to the central agent; after all subtasks corresponding to each agent collaborative control task are completed (each subtask has a pre-defined completion standard, such as reaching the destination or the end of broadcasting), each executing agent also needs to report the task completion information to the central agent. After receiving the task completion information sent by the executing agents, the central agent updates the task execution status in the agent routing table and releases the relevant computing resources. After all subtasks in the agent routing table are completed, all agents will return to the standby state, ready to handle the next agent collaborative control task.

[0026] A vehicle intelligent agent control device, comprising: The receiving module is used for the central intelligent agent to receive task parsing information sent by the cockpit intelligent agent. The task parsing information is obtained by the cockpit intelligent agent after receiving the intelligent agent collaborative control task and performing semantic parsing on the intelligent agent collaborative control task. The plan generation module is used by the central intelligent agent to break down the intelligent agent collaborative control task into multiple sub-tasks according to the task parsing information, assign an execution intelligent agent to each sub-task, and generate a task execution plan; the execution intelligent agents include the intelligent driving intelligent agent and the cockpit intelligent agent, and the task execution plan is used to instruct each of the execution intelligent agents to execute the corresponding sub-task; The execution module is used by the central agent to send task execution requests corresponding to each subtask to each execution agent according to the task execution plan, so that each execution agent can execute the corresponding subtask according to the task execution request.

[0027] An electronic device includes a controller and a memory, wherein, Memory, used to store computer programs; The controller is used to execute the program stored in the memory to implement the above-mentioned vehicle intelligent agent control method.

[0028] A vehicle that includes the aforementioned electronic equipment.

[0029] A computer-readable storage medium storing a computer program, which, when executed by a controller, implements the above-described vehicle intelligent agent control method. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart of a vehicle intelligent agent control method provided in an embodiment of this application; Figure 2 This is a flowchart illustrating step S20 of a vehicle intelligent agent control method provided in an embodiment of this application; Figure 3 This is a flowchart illustrating step S30 of a vehicle intelligent agent control method provided in an embodiment of this application; Figure 4 This is a flowchart illustrating a vehicle intelligent agent control method according to another embodiment of this application; Figure 5 This is a flowchart illustrating a vehicle intelligent agent control method according to another embodiment of this application; Figure 6 This is a schematic diagram of the structure of a vehicle intelligent agent control device provided in an embodiment of this application; Figure 7 This is a schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0032] To make the technical problems, technical solutions, and beneficial effects solved by this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] In one embodiment, please refer to Figure 1 A vehicle intelligent agent control method is provided, the method comprising the following steps S10-S30: S10. The central intelligent agent receives task parsing information sent by the cockpit intelligent agent. The task parsing information is obtained by the cockpit intelligent agent after receiving the intelligent agent collaborative control task and performing semantic parsing on the intelligent agent collaborative control task. The intelligent agent collaborative control task refers to a task that requires multiple intelligent agents to perform collaboratively, such as a task that requires the cockpit intelligent agent, the central intelligent agent, and the intelligent driving intelligent agent to perform collaboratively.

[0035] As an example, in step S10, receiving the task parsing information sent by the cockpit agent includes: after the cockpit agent performs semantic parsing on the received agent collaborative control task, extracts the user intent, task entity, and constraints in the agent collaborative control task, generates task parsing information based on the user intent, task entity, and constraints, and encapsulates the task parsing information into a structured semantic task package, receiving the structured semantic task package containing the task parsing information sent by the cockpit agent.

[0036] Understandably, during the process of users sending collaborative control tasks to the cockpit agent through preset input methods such as voice, touch, or a combination of voice and vision, the cockpit agent can automatically trigger guided feedback based on the user's silence duration. That is, after a user sends a collaborative control task via voice, the cockpit agent responds and prompts for user feedback. However, if the user does not respond and the silence duration exceeds a preset threshold, the cockpit agent will automatically trigger guided feedback. For example, if a user sends a collaborative control task to the cockpit agent via voice, such as "Navigate to the destination parking lot," and the cockpit agent replies, "I will recommend three parking lots for your destination. Which one would you like to go to?" if the user does not respond and the silence duration exceeds the preset threshold, the cockpit agent will proactively initiate another guided question: "Are you still there? Would you like me to introduce these three parking lots again?" In this embodiment, the cockpit intelligent agent enables user input access and semantic parsing. Specifically, the user sends an intelligent agent collaborative control task to the cockpit intelligent agent via preset input methods such as voice, touch, or a combination of voice and vision (e.g., the user outputs a voice command like "Navigate to the company, pick up a package on the way"). The cockpit intelligent agent then performs semantic understanding on the user's input, extracting the user's intent (e.g., navigation, picking up a package), task entities (e.g., location, action), and constraints (e.g., route, order). Based on the extracted user intent, task entities, and constraints, it generates task parsing information, encapsulates this information into a structured semantic task package, and sends it to the central intelligent agent. In this embodiment, the cockpit intelligent agent can parse fuzzy natural language and output a structured representation of the user's intent, ensuring semantic accuracy for the intelligent agent collaborative control task.

[0037] S20. The central intelligent agent, based on the task parsing information, divides the agent-coordinated control task into multiple sub-tasks, assigns an execution intelligent agent to each sub-task, and generates a task execution plan. The execution intelligent agents include the intelligent driving agent and the cockpit agent. The task execution plan is used to instruct each execution intelligent agent to execute the corresponding sub-task. Further, in this embodiment, the maximum depth of dividing the agent-coordinated control task into multiple sub-tasks can be set to within 5 steps to avoid infinite recursion in the division process. When dividing sub-tasks, the central intelligent agent also needs to comprehensively consider the context state, task feasibility, and task dependencies between sub-tasks. The task execution plan can be generated based on each sub-task and its corresponding task execution order and task dependencies.

[0038] S30. The central intelligent agent sends task execution requests corresponding to each sub-task to each of the executing intelligent agents according to the task execution plan, so that each of the executing intelligent agents can execute the corresponding sub-task according to the task execution request. The task execution plan includes, but is not limited to, sub-tasks, map status corresponding to sub-tasks, vehicle positioning, path planning information, historical task trajectories, and other data. Understandably, the path planning information mentioned above can be initially planned according to user intent and task issuance time, and then needs to be adjusted according to priority strategies. Priority strategies include rules such as safety priority, time priority, and fuel consumption priority. This priority measurement can be preset or adjusted according to user historical preferences or the user's explicit requirements in the agent-coordinated control task (and its subsequent associated instruction information). User historical preferences can be determined based on the path planning data of the user's historically executed sub-tasks using a preset path preference judgment model. The preset path preference judgment model can be trained based on a neural network model, which will not be elaborated here.

[0039] In this embodiment, for intelligent agent collaborative control tasks requiring multiple intelligent agents to work together, the cockpit intelligent agent performs semantic parsing to obtain task parsing information. Then, the central intelligent agent can break down the intelligent agent collaborative control task into multiple sub-tasks based on the task parsing information and assign different sub-tasks to different executing intelligent agents (including the intelligent driving agent and the cockpit intelligent agent) for collaborative completion. In this process, the cockpit intelligent agent possesses complete language understanding and generation capabilities. Using the cockpit intelligent agent for semantic parsing allows for rapid understanding of the user's complex task intent without affecting the real-time performance of the collaborative control task, greatly enhancing the usability and user experience of the cockpit intelligent agent. The central intelligent agent can translate and schedule tasks, enabling deep semantic integration between the cockpit intelligent agent and the intelligent driving agent, rather than a state of isolated, non-interfering operation of physical modules. Through the central control of the central intelligent agent, the cockpit intelligent agent and the intelligent driving agent can achieve the goal of collaboratively executing tasks in the most efficient way, based on the user's intent in the intelligent agent collaborative control task, thus improving task execution efficiency and effectiveness. The central intelligent agent, cockpit intelligent agent, and intelligent driving intelligent agent in this application can all perform collaborative optimization around the user intent corresponding to the intelligent agent collaborative control task. For example, when a user expresses driving preferences (such as "avoid highways") to the cockpit intelligent agent through voice dialogue, the central intelligent agent can adjust the path planning and control strategy in real time based on the task parsing information after semantic understanding fed back by the cockpit intelligent agent, and then instruct the vehicle control module to execute the corresponding driving actions, realizing a truly user intent-driven driving behavior. This application not only provides a brand-new cockpit-driver integrated paradigm at the architectural level, but also brings significant improvements in system performance, safety, and user satisfaction in actual deployment and use, and has broad engineering application prospects and industrial promotion value.

[0040] As an example, such as Figure 2 As shown, in step S20, the process of dividing the agent collaborative control task into multiple sub-tasks based on the task parsing information, assigning an execution agent to each sub-task, and generating a task execution plan includes steps S201-S203: S201. Based on the task parsing information, the vehicle's surrounding environment information, and the agent capability map associated with each of the executing agents, the agent collaborative control task is divided into multiple sub-tasks. The agent capability map includes executable task types corresponding to each of the executing agents. Since the agent capability map pre-defines the executable task types of the executing agents, it can serve as the basis for dividing the agent collaborative control task into multiple sub-tasks and assigning executing agents to the sub-tasks. That is, only when the task type of a sub-task after the agent collaborative control task is divided is an executable task type corresponding to at least one executing agent can the sub-task be successfully assigned to the corresponding executing agent and executed by the assigned executing agent.

[0041] S202. Assign a corresponding execution agent according to the task type of the sub-task; wherein, the execution agent to be assigned to the sub-task includes, but is not limited to, the intelligent driving agent and the cockpit agent, and may also include the central agent or other agents, which are not limited here.

[0042] In one embodiment, step S202, which involves allocating the corresponding execution agent based on the task type of the sub-task, includes: when the task type of the sub-task is path planning and control, determining that the execution agent corresponding to the sub-task is the intelligent driving agent; and when the task type of the sub-task is information broadcasting and interaction, determining that the execution agent corresponding to the sub-task is the cockpit agent. This embodiment describes the specific process of allocating the execution agent for the sub-task. That is, when the task type of the sub-task is path planning and control, since the intelligent driving agent's agent capability map has already indicated that its executable task type is path planning and control, the corresponding execution agent for this sub-task needs to be the intelligent driving agent; while when the task type of the sub-task is information broadcasting and interaction, since the cockpit agent's agent capability map has already indicated that its executable task type is information broadcasting and interaction, the corresponding execution agent for this sub-task needs to be the cockpit agent. Understandably, the central agent needs to allocate and schedule tasks according to the agent state of the execution agent and the path planning of the task to ensure sequential order and concurrent resource coordination.

[0043] S203. Determine the task execution order and task dependency relationship corresponding to each of the sub-tasks based on the task parsing information, and generate a task execution plan for all the sub-tasks based on the task execution order and task dependency relationship.

[0044] In this embodiment, a central intelligent agent is used to decompose and plan the path for the intelligent agent collaborative control task. Specifically, the central intelligent agent, based on the received task parsing information, vehicle surrounding environment information (such as current location and traffic conditions), and the intelligent agent capability map that corresponds one-to-one with different executing intelligent agents, breaks down the intelligent agent collaborative control task into multiple sub-tasks (such as "setting destination", "planning route", "requesting left turn control", etc.). For each sub-task, based on the task type of the sub-task and the executable task types of the executing intelligent agents pre-specified in the intelligent agent capability map, a specific executing intelligent agent can be assigned to each sub-task. Then, based on the user intent (e.g., navigation, picking up a package), task entity (e.g., location, action), and constraints (e.g., following the route, order, etc.) corresponding to the intelligent agent collaborative control task in the task parsing information of the structured semantic task package, the task execution order and task dependencies corresponding to the sub-tasks can be determined. Furthermore, based on the task execution order and task dependencies of different sub-tasks, the central intelligent agent generates a task execution plan for managing all sub-tasks. Understandably, the task execution plan can be represented graphically as a task plan diagram.

[0045] Understandably, in this embodiment, graph search or scheduling algorithms can be used to determine the task execution order among subtasks. Further, the task execution plan for all subtasks includes the task priority corresponding to each subtask, which can be used to determine or adjust the task execution order. This task priority can be initially determined based on the task assignment time, and then adjusted based on factors such as task type (e.g., navigation type has a higher priority than weather broadcast type), user preferences (user preferences can be determined using a preset preference judgment model based on the user's historical execution order of subtasks. For example, if the preset preference judgment model determines that the user likes to listen to music first upon boarding, the priority of the music listening type subtask can be increased; the preset preference judgment model can be trained based on a neural network model, which will not be elaborated here). The task execution plan includes not only the subtasks corresponding to the current agent collaborative control task, but also the subtasks that have not yet been completed in historical collaborative control tasks (agent collaborative control tasks executed before the current time) that the central agent has already started executing, so that the central agent can uniformly manage all subtasks according to the task execution plan. Furthermore, if a subtask has a low priority and its execution time in the task execution plan has expired for a period exceeding the preset expiration time, it can be directly deleted from the task execution plan. The execution plans for other related subtasks in the task execution plan will then be adjusted accordingly to ensure the smooth execution of other subtasks while removing redundant data. However, if the subtask is pre-marked as important, indicating its significance, even if it has a low priority and its execution time in the task execution plan has expired for a period exceeding the preset expiration time, an expiration notification must still be sent to the user via the cockpit agent. This allows the user to confirm whether the subtask needs to be deleted, preventing the loss of important tasks.

[0046] As an example, after step S203, that is, after generating the task execution plan for all the sub-tasks according to the task execution order and task dependencies, the method further includes: Extract the task intent corresponding to the intelligent agent collaborative control task from the task execution plan, determine the similarity between the task intent and the user intent of the task parsing information, and trigger a secondary confirmation operation when the similarity is lower than a preset consistency threshold. The secondary confirmation operation includes: displaying the task intent to the user through the cockpit intelligent agent in a preset interaction mode (voice, touch, voice and vision combination, etc.) so that the user can confirm whether to execute the intelligent agent collaborative control task according to the task intent.

[0047] In other words, the user's initial intent is identified by the cockpit agent, which generates task parsing information based on it and reports it to the central agent. The central agent may rewrite the task execution plan a second time when generating the plan (for example, if the user's intent is to go to the supermarket, the central agent will rewrite it as follows: call the navigation system -> destination: supermarket). Therefore, in this embodiment, the central agent judges the similarity between the task intent and the user intent in the task parsing information. Then, when a secondary confirmation operation is triggered, the cockpit agent needs to confirm with the user whether to execute the agent-coordinated control task according to the task intent in the task execution plan. If the user agrees, the task execution plan is modified. If the user does not agree, a task modification operation is triggered, and the task execution plans for all the sub-tasks need to be regenerated.

[0048] As an example, such as Figure 3 As shown, in step S30, sending task execution requests corresponding to each subtask to each execution agent according to the task execution plan, so that each execution agent can execute the corresponding subtask according to the task execution request, includes: S301, obtained Pick The task execution order and task dependencies in the task execution plan determine the target task from all the subtasks, and the execution agent assigned to the target task is determined as the target agent; S302, a task execution request for the target task is published to the target intelligent agent, and task association information corresponding to the target task is transmitted to the target intelligent agent so that the target intelligent agent can execute the target task according to the task association information.

[0049] In this embodiment, the scheduling and execution of the task execution plan are achieved through the collaboration of the central intelligent agent and the execution intelligent agent. Specifically, the central intelligent agent determines the target task to be executed from all its agents based on the task execution order and task dependencies. Then, the central intelligent agent issues a task execution request for the target task to the target intelligent agent. At the same time, it transmits relevant task-related information (such as map, user preferences, vehicle status, and other contextual information) to the target intelligent agent through the A2A protocol. After that, the execution intelligent agent can execute the target task according to the above-mentioned task-related information and the task execution request.

[0050] As an example, in step S302, the step of issuing a task execution request for the target task to the target agent and transmitting task association information corresponding to the target task to the target agent so that the target agent can execute the target task according to the task association information includes: When the target intelligent agent is an intelligent driving agent, the system publishes a task execution request for the target task to the intelligent driving agent via the A2A (Agent-to-Agent Protocol) and transmits task association information corresponding to the target task to the intelligent driving agent. This allows the intelligent driving agent to determine the path planning and control data corresponding to the target task based on the task association information. Then, based on the path planning and control data and a preset structured vocabulary, the system generates at least one structured execution instruction and outputs the structured execution instruction to the vehicle control module for execution. The preset structured vocabulary includes several structured vehicle control instructions, and each structured execution instruction is one of the vehicle control instructions.

[0051] In this embodiment, after receiving a task execution request, the intelligent driving agent generates corresponding path planning control data based on the sub-task, the corresponding task execution request, and task association information. This data includes phrases like "maintain lane," "speed limit driving," and "lane change preparation." Then, the path planning data is used to generate structured execution instructions according to a preset structured vocabulary and output to the vehicle control module. In this embodiment, the vehicle control module can directly execute the structured execution instructions upon receiving them. The preset structured vocabulary contains only a limited number of structured execution instructions corresponding to a limited number of control actions. These instructions uniformly adopt a structured format, making them simple and less prone to errors. For example, structured execution instructions could be phrases like "turn left" or "speed 30." This application uses a dedicated intelligent driving agent to independently handle the output of vehicle control instructions. The preset structured vocabulary compresses the decoding space of vehicle control instructions, significantly reducing inference latency and ensuring millisecond-level response capabilities during actual vehicle operation. Simultaneously, the space for outputting vehicle control instructions possesses structural rigidity, improving control stability and safety in complex environments and meeting automotive-grade high reliability requirements.

[0052] Understandably, in the above embodiments, dynamic scheduling of vehicle-level semantic tasks can be achieved through cockpit agents. That is, the cockpit agent supports complex task collaborative processing guided by multi-turn human-machine interaction. For example, the cockpit agent can perform semantic parsing on semantically complex agent collaborative control tasks such as "giving someone a ride," "navigating and adjusting the air conditioning," and "picking up a package while avoiding congested routes." Thus, the cockpit agent possesses complete language understanding and generation capabilities. Combined with a large semantic vocabulary space, it can support language tasks such as multi-turn semantic interaction, complex intent recognition, context preservation, and task orchestration. This enables a process where the cockpit agent performs semantic parsing and reasoning, the central agent orchestrates and translates, and the executive agents, such as the intelligent driving agent and the cockpit agent, implement the process, breaking through the previous passive response interaction mode centered on physical components in existing vehicle systems. Compared to the traditional approach of forcibly incorporating both autonomous driving control tasks and interaction tasks into a unified large vocabulary and model, this embodiment has a structural decoupling attribute between a small vocabulary and a large vocabulary decoding space, which can optimize the semantic understanding and generalization capabilities of the interaction system. Specifically, this application deploys the intelligent driving agent and the cockpit agent as dedicated models with different decoding vocabulary sizes. The intelligent driving agent uses a pre-set structured small vocabulary to generate structured execution instructions, ensuring the real-time performance and stability of vehicle control command generation. The cockpit agent uses a large vocabulary model for semantic parsing during interaction, supporting complex semantic reasoning and human-computer interaction. The two complement each other without conflict, achieving semantic-level collaboration between the cockpit agent and the intelligent driving agent. Based on the A2A protocol, this application achieves task-level decoupling and semantic-level fusion between the cockpit agent and the driving agent. A central agent is introduced to realize semantic-level task scheduling and conversion. Through an independent central agent, the semantic task execution paths of the cockpit agent and the intelligent driving agent are coordinated, realizing intent translation, state feedback, and task decomposition, significantly improving cross-task communication efficiency and semantic consistency. This central agent effectively solves problems such as multi-task semantic conflicts and control path uncertainty in a unified model, overcoming the technical bottlenecks of reasoning conflicts, vocabulary redundancy, and non-real-time control inherent in traditional "single-model multi-task" architectures.

[0053] As an example, in step S302, the step of issuing a task execution request for the target task to the target agent and transmitting task association information corresponding to the target task to the target agent so that the target agent can execute the target task according to the task association information includes: When the target intelligent agent is a cockpit intelligent agent, the task execution request of the target task is published to the cockpit intelligent agent through the A2A protocol, and the task association information corresponding to the target task is transmitted to the cockpit intelligent agent so that the cockpit intelligent agent can determine and execute the voice broadcast operation and / or cockpit feedback operation corresponding to the target task according to the task association information.

[0054] In this embodiment, after receiving a task execution request, the cockpit agent will synchronously perform voice broadcasting and / or cockpit feedback operations for the target task based on the task execution request and task association information corresponding to the sub-task. For example, it may provide feedback to the user regarding the target task: "A route has been planned for you, estimated to take 30 minutes," and wait for further input or changes from the user. Thus, through the collaboration of the central agent and the cockpit agent, the target task is executed. Furthermore, this embodiment uses the A2A protocol for collaborative information transmission between the central agent and the cockpit agent. The communication standard built on the A2A protocol provides a standard interface for subsequent system upgrades and model replacements. The cockpit agent can be replaced with a higher-capability model or a task-specific model as needed without reconstructing the entire system, greatly improving maintainability and functional expansion capabilities throughout the system's lifecycle. In achieving integrated vehicle-cabin-driver tasks, it offers significant advantages in system performance improvement, enhanced interactive experience, and engineering deployment.

[0055] Understandably, during the execution of the aforementioned target tasks, dynamic scheduling of semantic tasks at the vehicle level can be achieved through the cockpit intelligent agent. Specifically, the cockpit intelligent agent supports collaborative processing of complex tasks guided by multi-round human-machine interaction. For example, users can perform semantic compound tasks such as "giving someone a ride along the way," "navigating and adjusting the air conditioning," and "picking up a package without going through congested routes" at any time. These tasks can be inferred by the cockpit intelligent agent, orchestrated by the central intelligent agent, and implemented by the intelligent driving agent, breaking through the passive response interaction mode of existing in-vehicle systems that are centered on physical components.

[0056] In the above embodiments, the communication specifications built on the A2A protocol provide a standard interface for subsequent system upgrades and model replacements. The cockpit intelligent agent and the intelligent driving agent can be replaced with higher-capability models or task-specific models as needed without reconstructing the entire system, greatly improving the maintainability and functional expansion capabilities throughout the system's lifecycle. Furthermore, the central intelligent agent, acting as a bridge module, can also support the introduction of third-party intelligent agents (such as parking and energy management agents) to achieve full-scenario expansion, enhancing the system's scalability and module reusability. In addition, this application can simplify engineering deployment and computing resource allocation strategies. That is, the multi-agent decoupled architecture allows each module to be deployed on heterogeneous computing resources. For example, the intelligent driving agent can be deployed on an automotive-grade MCU (Microcontroller Unit), and the cockpit intelligent agent can be deployed on a high-performance SoC (System on Chip) or cloud inference module, supporting flexible computing power scheduling and operating environment optimization, improving overall system deployment efficiency, energy management capabilities, and resource utilization.

[0057] As an example, after step S302, that is, after the central intelligent agent sends a task execution request corresponding to each of the sub-tasks to each of the executing intelligent agents according to the task execution plan, so that each of the executing intelligent agents can execute the corresponding sub-task according to the task execution request, the process further includes: the central intelligent agent receiving the task execution status fed back by each of the executing intelligent agents through the A2A protocol, and updating the task execution plan according to the task execution status.

[0058] In this embodiment, the task execution status corresponding to each subtask will be transmitted back to the central intelligent agent in real time via the A2A protocol, so as to facilitate the scheduling and updating of other subtasks in the task execution plan.

[0059] Furthermore, the task execution status can include information on whether a subtask failed, the load status of the corresponding execution agent (such as CPU utilization, inference queue length, etc.), and the waiting time for subtask scheduling. When the waiting time for a subtask exceeds the maximum waiting time, or when a subtask fails, if the failed subtask has a low priority, the central agent can issue an instruction to the cockpit agent, which will then send an alarm message to the user via a multi-turn dialogue, allowing the user to confirm whether to reschedule the subtask. Conversely, if the failed subtask has a high priority or involves security issues, the central agent can issue an instruction to the cockpit agent, which will then send an alarm message to the user via a multi-turn dialogue, allowing the user to confirm whether to reschedule the subtask, or can directly reschedule the subtask. Task execution failure is considered to occur when there is a conflict between the newly planned subtask and an earlier planned but not yet executed subtask, or when some subtasks are terminated due to not meeting execution conditions. For example, a previous subtask might have required opening a window, but it was not executed because it was preempted by other higher-priority subtasks. Now that it's raining, even if the historical subtask needs to be executed, it's not suitable to do so, and in this case, the subtask will be considered to have failed.

[0060] Furthermore, this application predefines the task failure status codes corresponding to all subtasks that can provide feedback on task execution status, as well as the corresponding processing actions. That is, if a subtask fails to execute, the corresponding processing action can be determined based on the task failure status code, and then the failed task can be processed according to the processing action.

[0061] As an example, such as Figure 4 As shown, after step S30, that is, after the central intelligent agent sends task execution requests corresponding to each sub-task to each of the executing intelligent agents according to the task execution plan, so that each executing intelligent agent can execute the corresponding sub-task according to the task execution request, the process further includes: S40, when the cockpit agent receives a task change request sent via the A2A protocol, the central agent obtains the change intent in the task change request; the task change request is generated by the cockpit agent after semantically parsing the received change information and obtaining the change intent based on the parsing. S50, the central intelligent agent determines the change task from all the sub-tasks according to the change intention, and determines the associated task from all the sub-tasks according to the task execution order and task dependency relationship in the task execution plan, and then updates the change task and associated task in the task execution plan.

[0062] In this embodiment, dynamic coordination of intelligent agent collaborative control tasks can be achieved through multi-round semantic interaction between the cockpit intelligent agent and the user, as well as information interaction with the central intelligent agent via the A2A protocol. Specifically, when the user issues a task change request during the execution of an intelligent agent collaborative control task to make a new request or change the current user intent (such as "go to the charging station first"), the cockpit intelligent agent will semantically parse the task change request to obtain the change intent and then notify the central intelligent agent. Subsequently, the central intelligent agent determines the corresponding change task based on the change intent and updates and adjusts the task execution plan in real time, such as calling or terminating the change task or related tasks to ensure the stable execution of the task chain. In the above process, the multi-round interaction process between the cockpit intelligent agent and the user supports a memory mechanism and context reference, enhancing task persistence and user experience. Among them, the cockpit intelligent agent and the user maintain the number of dialogue history rounds (default 5 rounds), that is, the memory mechanism supports remembering 5 rounds of dialogue (the previous four rounds and the current round), and one round of dialogue refers to one back-and-forth dialogue between the user and the cockpit intelligent agent.

[0063] If the intent to change a task change request conflicts with the user intent corresponding to a previous agent collaborative control task or subtask, it can be handled by automatically overwriting the previous agent collaborative control task or subtask, requesting confirmation from the user on how to handle it, or keeping both in parallel.

[0064] In some embodiments, the vehicle intelligent agent control method further includes: the central intelligent agent acquiring the session connection duration between the cockpit intelligent agent and the user, and controlling the user interruption control delay threshold (ms) based on the session connection duration to ensure that the cockpit intelligent agent can respond to interruption requests in a timely manner. This enables a balance between semantic continuity and system responsiveness between multi-turn semantic interactions, providing support for LLM (Large Language Model) state preservation capabilities and central state management.

[0065] In some embodiments, the vehicle intelligent agent control method further includes: the central intelligent agent acquiring the status of the cockpit intelligent agent, the intelligent driving intelligent agent, and the central intelligent agent at a preset update frequency (e.g., 200ms), such as: executing, silent, etc.

[0066] As an example, such as Figure 5 As shown, after step S30, that is, after the central intelligent agent sends task execution requests corresponding to each sub-task to each of the executing intelligent agents according to the task execution plan, so that each executing intelligent agent can execute the corresponding sub-task according to the task execution request, the process further includes: S60, upon receiving a task anomaly warning message sent by the intelligent driving agent via the A2A protocol, the central intelligent agent determines the abnormal task from all the sub-tasks based on the task anomaly warning message; the task anomaly warning message is identified by the intelligent driving agent based on real-time vehicle information and surrounding environmental information; specifically, the intelligent driving agent monitors real-time vehicle information (such as steering wheel position, vehicle speed, speed exceeding limit, sudden change in yaw angle, path deviation, etc.) and surrounding environmental information (such as road conditions), and upon identifying task anomaly warning messages such as execution anomalies or environmental anomalies (such as road conditions indicating congestion ahead preventing progress), immediately sends a task anomaly warning message to the central intelligent agent via the A2A protocol to issue an alarm; the central intelligent agent determines the abnormal task corresponding to the task anomaly warning message.

[0067] S70, when the task anomaly warning information is a preset emergency message, the central intelligent agent sends an emergency braking instruction to the intelligent driving agent, so that the intelligent driving agent can generate a structured execution instruction corresponding to the emergency braking instruction based on a preset structured vocabulary, and output the structured execution instruction to the vehicle control module for execution; understandably, the preset emergency message is preset according to needs. If the task anomaly warning information is a preset emergency message such as a pedestrian suddenly staging an accident or a sudden vehicle malfunction, the central intelligent agent immediately outputs an emergency braking instruction to the intelligent driving agent, and the intelligent driving agent quickly generates a structured execution instruction corresponding to the emergency braking instruction based on a preset structured vocabulary, and outputs the structured execution instruction to the vehicle control module, thereby performing an emergency braking operation to ensure vehicle safety.

[0068] S80, when the task anomaly warning information is not a preset emergency message, the central agent sends a replanning request containing the abnormal task to the cockpit agent, so that the cockpit agent can respond to the replanning request and interact with the user to perform the replanning operation. That is, if the task anomaly warning information is not a preset emergency message, the central agent can notify the cockpit agent of the replanning request so as to instruct it to perform the replanning operation. Then, the cockpit agent interacts with the user again to resend the instruction or abandon the execution of the abnormal task. Subsequently, the cockpit agent feeds back the user instruction to the central agent, and the central agent adjusts the task execution plan according to the user instruction to ensure system robustness and interaction continuity.

[0069] In this embodiment, through the collaborative cooperation between the cockpit intelligent agent, the intelligent driving agent, and the central intelligent agent, vehicle safety and system recoverability can be ensured when anomalies occur during task execution. Therefore, the parameter design in this application needs to support distributed fault tolerance and semantic-level degradation feedback. In this embodiment, even if the cockpit intelligent agent crashes or misjudges semantics, it will not directly affect the control safety of the intelligent driving agent. Conversely, if the driving intelligent agent feeds back task anomaly warning information (such as an impassable environment), the central intelligent agent can block the emergency corresponding to the task anomaly warning information to ensure vehicle safety, or send the task anomaly warning information back to the cockpit intelligent agent to prompt the user to interact again to process it, thereby improving the overall fault tolerance of the system and enhancing the system's safety robustness and fault isolation capabilities.

[0070] In the above embodiments, if an abnormal state such as the central intelligent agent crashes or freezes occurs, the vehicle can automatically activate a backup central intelligent agent or roll back the intelligent agent's collaborative control task.

[0071] As an example, in step S20, before the central intelligent agent receives the task parsing information sent by the cockpit intelligent agent, the following steps are included: After confirming that the central intelligent agent, the intelligent driving agent, and the cockpit intelligent agent have all started, the central intelligent agent receives the registration requests sent to it by the intelligent driving agent and the cockpit intelligent agent through the A2A protocol, and responds to the registration requests to establish an intelligent agent routing table. The intelligent agent routing table includes the task execution status and the intelligent agent capability map corresponding to the intelligent driving agent and the cockpit intelligent agent respectively.

[0072] This embodiment proposes an A2A protocol system for intelligent cockpits and autonomous driving, standardizing the semantic representation structure, control primitives, and task state feedback mechanisms among multiple agents. This enables the system to possess excellent scalability and module replacement capabilities, significantly outperforming traditional end-to-end black-box models. Specifically, after the system powers on, the cockpit agent, the intelligent driving agent, and the central agent start up respectively. The cockpit agent and the intelligent driving agent register with the central agent through the A2A protocol standard interface, declaring their own capabilities (such as semantic reasoning, path planning, control output, etc.) and resource status (such as the number of CPU cores, memory, inference latency, computing load, energy consumption budget, etc.). Then, the central agent establishes an agent routing table and maintains an agent capability graph (the agent capability graph includes, but is not limited to, agent state, task execution state, and vehicle environment state). The agent capability graph can be used for subsequent task scheduling decisions and maintaining load balancing.

[0073] As an example, after step S30, that is, after the central intelligent agent sends a task execution request corresponding to each of the execution intelligent agents according to the task execution plan, so that each of the execution intelligent agents can execute the corresponding sub-task according to the task execution request, the method further includes: After receiving the task completion information sent by the executing agent, the central agent updates the task execution status in the agent routing table according to the task completion information and performs a computing resource release operation.

[0074] In this embodiment, resources can be released promptly after the task is completed, thereby reducing system energy consumption and computing resource usage, and improving system throughput and maintenance efficiency. Specifically, after a subtask is completed, the executing agent corresponding to the subtask reports the task completion information to the central agent; after all subtasks corresponding to each agent collaborative control task are completed (each subtask has a pre-defined completion standard, such as reaching the destination or the end of broadcasting), each executing agent also needs to report the task completion information to the central agent. After receiving the task completion information sent by the executing agents, the central agent updates the task execution status in the agent routing table and releases the relevant computing resources. After all subtasks in the agent routing table are completed, all agents will return to the standby state, ready to handle the next agent collaborative control task.

[0075] In some embodiments, if the executing agent remains idle for a certain period of time (e.g., 30 seconds), it will automatically release relevant computing resources, for example, by switching its state from running to silent. Understandably, whether cached data during the execution of different subtasks is retained (e.g., route map cache, user preferences), and the retention period settings, can be predefined, set according to user preferences, or confirmed by the user in real time through the cockpit agent. By default, these settings are predefined or set according to user preferences. However, when the central agent detects that a task cannot be executed, or that user preferences have changed significantly, or when the user actively requests confirmation, it will trigger the operation of confirming with the user through the cockpit agent.

[0076] In the above embodiments of this application, complex task collaborative processing under multi-round human-machine interaction guidance between the cockpit intelligent agent and the user is supported. The cockpit intelligent agent can infer semantic compound tasks such as "giving someone a ride along the way," "navigating and adjusting the air conditioning," and "picking up a package without going through congestion," thereby realizing strategies orchestrated by the central intelligent agent and implemented by the intelligent driving agent, breaking through the passive response interaction mode centered on physical components in existing in-vehicle systems. The method proposed in this application is based on a multi-agent collaborative architecture, which decouples and links the intelligent cockpit intelligent agent, the autonomous intelligent driving agent, and the central coordination intelligent agent through the A2A protocol, so that they undertake task processing in different functional areas respectively, and realize dynamic collaboration, information exchange, and behavior scheduling at the task level.

[0077] In this application, the intelligent driving agent adopts a compression strategy based on a pre-defined structured small vocabulary. Its decoding space contains only a small number of vehicle control commands (such as "turn left," "decelerate," and "pull over") to ensure high real-time performance and high reliability in vehicle control. Furthermore, the intelligent driving agent can directly generate safe driving behavior decisions based on sensory inputs (such as camera, radar, and map information) and intent signals transmitted from the central processing unit, and possesses high-frequency response capabilities. Correspondingly, the cockpit agent, based on a large vocabulary language model, undertakes tasks such as natural language understanding, semantic reasoning, multi-turn dialogue, content retrieval, and preference management, supporting user voice / visual input interaction. The cockpit agent obtains the user's intent through natural language interaction and performs multi-turn semantic reasoning and context preservation processing to ultimately form a structured semantic task package, achieving a structured representation of task intent. The central processing unit is the core scheduling node of the system, performing task semantic mediation, state synchronization, and policy routing between the two heterogeneous agents, the intelligent driving agent and the cockpit agent, based on the A2A protocol. This accomplishes the following key tasks: Intent mapping: The central agent converts the user intent corresponding to the natural language generated by the cockpit agent into a task execution plan that can be received by the intelligent driving agent.

[0078] State perception: The central intelligent agent obtains data such as the current driving state, feasible paths, and environmental perception results from the intelligent driving agent to support the feasibility judgment of user intentions and task avoidance.

[0079] Control Feedback: During task execution, the central intelligent agent promptly transmits feedback information from the intelligent driving agent (such as inability to complete or path conflict) back to the cockpit intelligent agent, guiding it to conduct secondary confirmation or intention correction with the user.

[0080] Task orchestration: Coordinate the timing and priority of sub-tasks that the cockpit intelligent agent needs to perform (such as playing music and announcing routes) with the sub-tasks that the intelligent driving agent needs to perform, so as to achieve seamless integration of behaviors.

[0081] Throughout the process, the A2A protocol not only defines the semantic protocol format for communication between agents (including intent structure, state vector, control primitives, etc.), but also specifies behavior negotiation mechanisms (such as priority judgment, task preemption, and exception recovery) to ensure stable operation of the system under multi-task and multi-modal input. For example, a user in the car says, "Navigate me to the nearest coffee shop and pick up a package along the way." The cockpit agent recognizes the user's intent and calls the map service for route planning, uploading the multi-stage user intent to the central agent. The central agent calls the intelligent driving agent to execute dynamic path control and coordinate task progress (such as prioritizing navigation and requesting stops along the way). The intelligent driving agent continuously provides feedback on the path execution status. If it finds that the navigation path needs to turn around and the current road does not support it, the central agent can notify the user through the cockpit agent to adjust the strategy.

[0082] This application, through agent decoupling and A2A protocol collaboration, retains the high requirements of real-time performance and safety for autonomous driving control while fully releasing the cockpit's semantic understanding capabilities. This achieves true semantic integration and system-level collaboration between the cockpit and driver, providing a new architectural paradigm for intelligent vehicle systems. Furthermore, this application balances generalization ability with control robustness, overcoming the structural contradiction of strong semantic understanding but weak control precision in existing multimodal large models. Through responsibility separation and semantic collaboration, control tasks possess structured, rigid outputs while maintaining open semantic expression capabilities for interactive tasks, thus improving the overall system's generalization ability, safety robustness, and user experience. Moreover, this application achieves structural improvements to existing intelligent vehicle system solutions at the architecture, communication protocol, and model responsibility division layers, providing a new technical path to achieve true "cockpit-driver integration, semantic collaboration, and multi-agent fusion."

[0083] In this application, all communication between intelligent agents is based on the A2A protocol, which supports structured intent representation, context state synchronization, and anomaly handling identifiers. Each intelligent agent can be deployed locally (e.g., on an automotive-grade MCU), at an edge computing module, or connected to a cloud-based intelligent agent system via in-vehicle 5G, depending on the actual computing power available.

[0084] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0085] This application also provides a vehicle intelligent agent control device. Please refer to... Figure 6 ,include: The receiving module 11 is used for the central intelligent agent to receive task parsing information sent by the cockpit intelligent agent. The task parsing information is obtained by the cockpit intelligent agent after receiving the intelligent agent collaborative control task and performing semantic parsing on the intelligent agent collaborative control task. The plan generation module 12 is used by the central intelligent agent to divide the intelligent agent collaborative control task into multiple sub-tasks according to the task parsing information, assign an execution intelligent agent to each sub-task, and generate a task execution plan; the execution intelligent agents include the intelligent driving intelligent agent and the cockpit intelligent agent, and the task execution plan is used to instruct each execution intelligent agent to execute the corresponding sub-task; The execution module 13 is used to send task execution requests corresponding to each subtask to each of the execution agents according to the task execution plan, so that each of the execution agents can execute the corresponding subtask according to the task execution request.

[0086] This application also provides an electronic device 100, please refer to... Figure 7 The device includes a memory 1001 and a controller 1002. The memory 1001 stores computer programs, and the controller 1002 executes the programs stored in the memory 1001 to implement the vehicle intelligent agent control method described in any embodiment of this application. The electronic device can be the central intelligent agent in the above embodiments. In some embodiments, the electronic device can also be the aforementioned intelligent driving agent or the cockpit intelligent agent.

[0087] In one embodiment, this application provides a vehicle including the aforementioned electronic device 100.

[0088] This application also provides a computer-readable storage medium storing a computer program that, when executed by a controller, implements the vehicle intelligent agent control method described in any embodiment of this application.

[0089] In this application, "multiple" refers to two or more.

[0090] The terms “first,” “second,” “third,” “fourth,” etc., used in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0091] In this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, in this application, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0092] Unless otherwise specified, all steps in this application may be performed sequentially or randomly. For example, if the method includes steps A and B, it means that the method may include steps A and B performed sequentially, or it may include steps B and A performed sequentially. For example, if the method may also include step C, it means that step C may be added to the method in any order. For example, the method may include steps A, B, and C, or it may include steps A, C, and B, or it may include steps C, A, and B, etc.

[0093] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A vehicle agent control method, characterized by, The method comprises the following steps: The central intelligent agent receives task analysis information sent by the cockpit intelligent agent, wherein the task analysis information is obtained by performing semantic analysis on the intelligent agent cooperative control task received by the cockpit intelligent agent; The central intelligent agent splits the intelligent agent cooperative control task into multiple subtasks according to the task analysis information, allocates an execution intelligent agent to each subtask, and generates a task execution plan; the execution intelligent agent comprises a smart driving intelligent agent and the cockpit intelligent agent, and the task execution plan is used to instruct each execution intelligent agent to execute the corresponding subtask; The central intelligent agent sends a task execution request corresponding to each subtask to each execution intelligent agent according to the task execution plan, so that each execution intelligent agent executes the corresponding subtask according to the task execution request.

2. The vehicle agent control method of claim 1, wherein, The method further comprises the following steps: After the cockpit intelligent agent performs semantic analysis on the received intelligent agent cooperative control task, extracts the user intention, task entity and constraint condition in the intelligent agent cooperative control task, generates task analysis information according to the user intention, task entity and constraint condition, and encapsulates the task analysis information into a structured semantic task package, the cockpit intelligent agent sends the structured semantic task package containing the task analysis information.

3. The vehicle agent control method of claim 1, wherein, The method further comprises the following steps: According to the task analysis information, vehicle surrounding environment information and intelligent agent capability graph associated with each execution intelligent agent, the intelligent agent cooperative control task is split into multiple subtasks, and the intelligent agent capability graph comprises executable task types corresponding to each execution intelligent agent; According to the task type of each subtask, an execution intelligent agent corresponding to the subtask is allocated; According to the task analysis information, a task execution sequence and a task dependency relationship corresponding to each subtask are determined, and a task execution plan of all subtasks is generated according to the task execution sequence and the task dependency relationship.

4. The vehicle agent control method of claim 3, wherein, The method further comprises the following steps: When the task type of the subtask is path planning and control, the execution intelligent agent corresponding to the subtask is determined to be a smart driving intelligent agent; When the task type of the subtask is information broadcasting and interaction, the execution intelligent agent corresponding to the subtask is determined to be a cockpit intelligent agent.

5. The vehicle agent control method of claim 1, wherein, The method further comprises the following steps: Gained The target intelligent agent is instructed to execute the target task, and task associated information corresponding to the target task is transmitted to the target intelligent agent, so that the target intelligent agent executes the target task according to the task associated information. The task execution sequence and the task dependency relationship in the task execution plan determine a target task from all the sub-tasks, and determine an execution agent assigned for the target task as a target execution agent; ​ 6. The vehicle agent control method of claim 5, wherein, The task execution request of the target task is issued to the target agent, and the task association information corresponding to the target task is transmitted to the target agent, so that the target agent executes the target task according to the task association information, and the method comprises the following steps: When the target agent is an intelligent driving agent, the task execution request of the target task is issued to the intelligent driving agent through an A2A protocol, and the task association information corresponding to the target task is transmitted to the intelligent driving agent, so that the intelligent driving agent determines the path planning control data corresponding to the target task according to the task association information, and then generates at least one structured execution instruction according to the path planning control data and a preset structured small vocabulary table, and outputs the structured execution instruction to a vehicle control module for execution; wherein the preset structured small vocabulary table comprises a plurality of structured vehicle control instructions, and each structured execution instruction is one of the vehicle control instructions.

7. The vehicle agent control method of claim 5, wherein, The task execution request of the target task is issued to the target agent, and the task association information corresponding to the target task is transmitted to the target agent, so that the target agent executes the target task according to the task association information, and the method comprises the following steps: When the target agent is a cockpit agent, the task execution request of the target task is issued to the cockpit agent through an A2A protocol, and the task association information corresponding to the target task is transmitted to the cockpit agent, so that the cockpit agent determines and executes the voice broadcast operation or / and the cockpit feedback operation corresponding to the target task according to the task association information.

8. The vehicle agent control method of claim 1, wherein, The central agent sends the task execution request corresponding to each subtask to each execution agent according to the task execution plan, so that each execution agent executes the corresponding subtask according to the task execution request, and the method further comprises the following steps: The central agent receives the task execution state fed back by each execution agent through an A2A protocol, and updates the task execution plan according to the task execution state.

9. The vehicle agent control method of claim 1, wherein, The central agent sends the task execution request corresponding to each subtask to each execution agent according to the task execution plan, so that each execution agent executes the corresponding subtask according to the task execution request, and the method further comprises the following steps: When the task change request sent by the cockpit agent through the A2A protocol is received, the central agent obtains the change intention in the task change request; the task change request is generated according to the change intention obtained by the cockpit agent after semantic analysis of the received change information; The central agent determines the change task from all the subtasks according to the change intention, and determines the associated task from all the subtasks according to the task execution order and the task dependency relationship in the task execution plan, and then updates the change task and the associated task in the task execution plan. The task execution request of the target task is issued to the target agent, and the task association information corresponding to the target task is transmitted to the target agent, so that the target agent executes the target task according to the task association information, and the method comprises the following steps: When the target agent is an intelligent driving agent, the task execution request of the target task is issued to the intelligent driving agent through an A2A protocol, and the task association information corresponding to the target task is transmitted to the intelligent driving agent, so that the intelligent driving agent determines the path planning control data corresponding to the target task according to the task association information, and then generates at least one structured execution instruction according to the path planning control data and a preset structured small vocabulary table, and outputs the structured execution instruction to a vehicle control module for execution; wherein the preset structured small vocabulary table comprises a plurality of structured vehicle control instructions, and each structured execution instruction is one of the vehicle control instructions. The task execution request of the target task is issued to the target agent, and the task association information corresponding to the target task is transmitted to the target agent, so that the target agent executes the target task according to the task association information, and the method comprises the following steps: When the target agent is a cockpit agent, the task execution request of the target task is issued to the cockpit agent through an A2A protocol, and the task association information corresponding to the target task is transmitted to the cockpit agent, so that the cockpit agent determines and executes the voice broadcast operation or / and the cockpit feedback operation corresponding to the target task according to the task association information. The central agent sends the task execution request corresponding to each subtask to each execution agent according to the task execution plan, so that each execution agent executes the corresponding subtask according to the task execution request, and the method further comprises the following steps: The central agent receives the task execution state fed back by each execution agent through an A2A protocol, and updates the task execution plan according to the task execution state. The central agent sends the task execution request corresponding to each subtask to each execution agent according to the task execution plan, so that each execution agent executes the corresponding subtask according to the task execution request, and the method further comprises the following steps: When the task change request sent by the cockpit agent through the A2A protocol is received, the central agent obtains the change intention in the task change request; the task change request is generated according to the change intention obtained by the cockpit agent after semantic analysis of the received change information; The central agent determines the change task from all the subtasks according to the change intention, and determines the associated task from all the subtasks according to the task execution order and the task dependency relationship in the task execution plan, and then updates the change task and the associated task in the task execution plan.

10. The vehicle agent control method of claim 1, wherein, The central agent sends a task execution request corresponding to each subtask to each execution agent according to the task execution plan, so that each execution agent executes the corresponding subtask according to the task execution request, and the method further comprises the following steps of: When receiving the task exception warning information sent by the intelligent driving agent through the A2A protocol, the central agent determines an exception task from all the subtasks according to the task exception warning information; the task exception warning information is obtained by the intelligent driving agent according to the vehicle information and the vehicle surrounding environment information collected in real time by the vehicle; When the task exception warning information is preset emergency information, the central agent sends an emergency braking instruction to the intelligent driving agent, so that the intelligent driving agent generates a structured execution instruction corresponding to the emergency braking instruction according to a preset structured small vocabulary table, and outputs the structured execution instruction to a vehicle control module for execution; When the task exception warning information is not preset emergency information, the central agent sends a re-planning request containing the exception task to the cockpit agent, so that the cockpit agent interacts with the user to perform a re-planning operation in response to the re-planning request.

11. The vehicle agent control method of claim 1, wherein, Before the central agent receives the task analysis information sent by the cockpit agent, the method comprises the following steps of: After confirming that the central agent, the intelligent driving agent and the cockpit agent are all started, the central agent receives a registration request sent by the intelligent driving agent and the cockpit agent to the central agent through the A2A protocol, and establishes an agent routing table in response to the registration request, wherein the agent routing table comprises a task execution state and an agent capability graph corresponding to the intelligent driving agent and the cockpit agent, respectively.

12. The vehicle agent control method of claim 11, wherein, The central agent sends a task execution request corresponding to each subtask to each execution agent according to the task execution plan, so that each execution agent executes the corresponding subtask according to the task execution request, and the method further comprises the following steps of: After receiving the task completion information sent by the execution agent, the central agent updates the task execution state in the agent routing table according to the task completion information, and performs a computing resource release operation.

13. A vehicle agent control device characterized by comprising: Comprise: The receiving module is configured to receive, by the central agent, task analysis information sent by the cockpit agent, wherein the task analysis information is obtained by the cockpit agent by performing semantic analysis on the agent cooperative control task after receiving the agent cooperative control task; The plan generation module is configured to split, by the central agent, the agent cooperative control task into a plurality of subtasks according to the task analysis information, assign an execution agent to each subtask, and generate a task execution plan; the execution agent comprises an intelligent driving agent and the cockpit agent, and the task execution plan is used to instruct each execution agent to execute the corresponding subtask; The execution module is configured to send, by the central agent, a task execution request corresponding to each subtask to each execution agent according to the task execution plan, so that each execution agent executes the corresponding subtask according to the task execution request.

14. An electronic device, comprising: Comprise a controller and a memory, wherein, a memory for storing a computer program; a controller for executing the program stored on the memory to implement the vehicle agent control method of any one of claims 1 to 12.

15. A vehicle characterized by comprising: The electronic device of claim 14 is included.

16. A computer-readable storage medium, characterized in that, The computer program stored in the computer readable storage medium is executed by the controller to implement the vehicle agent control method of any one of claims 1 to 12.

Images