Vehicle-mounted system modular framework and collaboration method thereof

Abstract

The invention discloses a vehicle-mounted system modular framework and a collaboration method thereof, the vehicle-mounted system modular framework comprises an agent module cluster, and the agent module cluster comprises an interaction agent and at least one function execution agent; a large language model is deployed in each agent, and all the large language models are coordinated through a standardized interface; the interaction agent is used for receiving a natural language instruction of a user and analyzing the natural language instruction by using a large language model deployed in the interaction agent to generate a task instruction; and the function execution agent is used for receiving the task instruction, analyzing and extracting key parameters by using a large language model deployed in the function execution agent, executing security rule check, and executing a corresponding specific task if the task instruction passes the security rule check. According to the method, the rule code quantity can be reduced, the interactive agent realizes cross-agent dynamic task distribution based on the big language model reasoning ability, new function access does not need to modify a bottom layer architecture, only peripheral agents need to be increased or decreased, the development period can be effectively shortened, and the development progress can be accelerated.

Description

Technical Field

[0001] The present application belongs to the technical field of vehicle-mounted system design, and specifically relates to a modular framework and collaborative method of a vehicle-mounted system. Background Art

[0002] Existing in-vehicle systems mostly use a "rules engine + independent functional modules" architecture. For example, voice recognition (e.g., based on keyword matching) and vehicle control (e.g., hard-coded logic) are implemented through handwritten code, and specific code must be written for each function (e.g., navigation and entertainment systems are developed independently). This model results in: First, code redundancy is high, requiring repeated development of identical logic, such as user interaction. Second, cross-module collaboration is inefficient, requiring additional interface adaptation code. Finally, intelligence is limited by pre-set rules, making it difficult to process complex user commands. For example, when a user requests something like "recommended coffee shops along the way based on real-time traffic conditions," traditional in-vehicle systems struggle to implement this because it involves integrating multi-source data. Summary of the Invention

[0003] The technical purpose of this application is to address the current problems of long vehicle system software development cycle, large amount of code and insufficient software intelligence, and to provide a modular framework and collaborative method for vehicle systems. By combining the natural language processing capabilities of the large language model (LLM), it can realize the "low-code" development and deployment of vehicle system functions, and solve the problems of long development cycle and inconvenient maintenance caused by the traditional vehicle system's reliance on a large amount of rule-based code; at the same time, through the multi-agent collaborative mechanism, it can improve the system's intelligence level in scenarios such as user intent recognition and proactive services.

[0004] In order to achieve the above technical objectives, this application adopts the following technical solutions.

[0005] In a first aspect, an embodiment of the present application provides a modular framework for an in-vehicle system, including: An intelligent agent module cluster is provided at the vehicle end, wherein the intelligent agent module cluster includes an interactive intelligent agent and at least one function execution intelligent agent; Large language model collaboration mechanism: A large language model is deployed in each intelligent body; The interactive agent is used to receive natural language instructions from users and parse them using the large language model deployed within it to generate task instructions; The function execution agent is used to receive the task instructions, and use the large language model deployed internally to parse them, extract key parameters, and perform security rule checks. If the security rule checks are passed, the corresponding specific tasks are executed.

[0006] Further, the intelligent agent module cluster further includes a search intelligent agent set in the cloud, which is used to obtain road conditions, meteorology, and Internet information through the network and feedback it to the interaction intelligent agent.

[0007] Further, the interaction intelligent agent supports multimodal input, including voice, text, and image input, and provides feedback through voice broadcast and graphic display.

[0008] Further, the function execution intelligent agent includes a vehicle control intelligent agent and a multimedia intelligent agent: The vehicle control intelligent agent is used to parse vehicle control task instructions, perform checks based on safety rules, and call vehicle atomic services according to the vehicle control task instructions after passing the checks; The multimedia intelligent agent is used to parse audio and video playback task instructions, call local multimedia atomic services or directly call cloud media resources, and support text-to-speech conversion functions.

[0009] Further, the large language model supports differential deployment: The interaction intelligent agent uses a cloud-based multimodal large language model; the vehicle control intelligent agent uses a local lightweight large language model.

[0010] Further, each large language model collaborates through a standardized interface. The standardized interface is based on the RESTful protocol and supports the invocation of an atomic service list. The atomic service includes the description information of the atomic service and safety rules for the large language model to judge the invocation conditions and perform safety rule checks.

[0011] In a second aspect, an embodiment of the present application provides a collaboration method for the in-vehicle system modular framework provided in any possible implementation manner of the first aspect, including: The interaction intelligent agent receives a user's natural language instruction and parses it into a task instruction through an internal large language model; The interaction intelligent agent distributes the task instruction to the corresponding function execution intelligent agent according to the task type; The function execution intelligent agent extracts key parameters through an internal large language model and performs a safety rule check; after passing the safety check, the function execution intelligent agent calls an atomic service or cloud resource to execute the task and returns the result to the interaction intelligent agent, which summarizes and feeds back to the user.

[0012] Further, the method further includes: The interaction intelligent agent maintains the context by caching historical session records and optimizes the response speed of the large language model using a time forgetting mechanism.

[0013] Further, when the function execution intelligent agent includes a vehicle control intelligent agent, the method further includes: The vehicle service module is used to call vehicle atomic services to obtain real-time vehicle condition data and verify safety rules.

[0014] Further, when the function execution agents include a multimedia agent and a search agent; the method further includes: the multimedia agent calls the search agent to obtain media resources, and broadcasts the task result through a text-to-speech conversion tool.

[0015] Compared with the prior art, the in-vehicle system modular framework and its cooperation method provided by the embodiments of the present application have the following beneficial technical effects: Through the agent modular architecture and the processing logic judgment of the large language model, the in-vehicle system based on agents hands over a large number of logic judgments to the large language model for processing, greatly reducing the cumbersome rule code writing work required by the traditional rule engine and significantly reducing the amount of rule code. The interaction agent realizes cross-agent dynamic task distribution based on the reasoning ability of the large language model. When new functions are accessed, there is no need to modify the underlying architecture, and only the surrounding agents need to be added or reduced, which can effectively shorten the development cycle and speed up the development progress. The logical relationships of each functional module are clear, and maintenance personnel can quickly locate problems, reducing the maintenance workload and cost; at the same time, through the multi-agent cooperation mechanism, the intelligent level of the system in scenarios such as user intention recognition and proactive service is improved. Description of the Drawings

[0016] The drawings described herein are for illustrative purposes only and are not intended to limit the scope of the disclosure of the present application in any way. In addition, the shapes and proportional dimensions of the components in the drawings are only schematic and are used to assist in understanding the present application, rather than specifically limiting the shapes and proportional dimensions of the components of the present application. Those skilled in the art can, under the teaching of the present application, select various possible shapes and proportional dimensions according to specific circumstances to implement the present application. In the drawings: Figure 1 It is a schematic structural diagram of the in-vehicle system modular framework provided for the embodiment. Detailed Embodiments

[0017] In order to enable those skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0018] The embodiments of the present application provide an in-vehicle system modular framework, which includes an agent module cluster and a large language model cooperation mechanism arranged at the vehicle end. The agent module cluster includes an interaction agent and at least one function execution agent.

[0019] Large language model collaboration mechanism: Large language models are deployed in each intelligent body, and in the embodiments, the large language models collaborate through standardized interfaces.

[0020] The interaction intelligent body is used to receive natural language instructions from users and parse them using the large language model deployed inside it to generate task instructions; the function execution intelligent body is used to receive task instructions, parse them using the large language model deployed inside it, extract key parameters, and perform security rule checks. If the security rule check passes, it will execute the corresponding specific task.

[0021] In the embodiments, the interaction intelligent body, as the core intelligent body, can be deployed in the central coordination module of the vehicle-mounted computing unit (such as the domain controller).

[0022] In some embodiments, the interaction intelligent body can support receiving multi-modal inputs from users. For example, users can interact with the system in various ways such as voice, text, and image, and the system can give feedback in the form of voice announcements, graphical displays, etc.

[0023] In the embodiments, the interaction intelligent body can obtain vehicle body data through the vehicle service module, and the vehicle control intelligent body can call vehicle atomic services through the vehicle service module to obtain relevant data such as the real-time operating status of the vehicle, and provide information such as the current vehicle speed and vehicle fault information for users.

[0024] In the embodiments, the interaction intelligent body can also obtain road conditions, meteorological, and other Internet information through the search intelligent body set in the cloud: connect to the cloud intelligent body through the network to obtain meteorological information such as weather forecasts and inform users of weather changes in advance; it can also obtain other Internet information such as news.

[0025] The interaction intelligent body can implement function routing. Based on the reasoning ability of the large model, it can identify user instructions and perform function routing (such as 'window control' corresponding to the vehicle control intelligent body, 'play music' corresponding to the multimedia intelligent body). It can achieve dynamic task distribution across intelligent bodies. Compared with traditional hard-coded interfaces, this mechanism can improve the efficiency of new function access by more than 70%.

[0026] In some embodiments, the interaction intelligent body can be implemented through the following examples.

[0027] 1. Provide an input interface based on the restful protocol, including the interface path and parameter format. The parameter format is an array, and each element in the array is a JSON object containing the following fields: type: Represents the input type, with values of text (text), audio (audio), or image (image).

[0028] data: Stores data of the corresponding type. The text is in string form, and the audio and images are Base64-encoded strings.

[0029] Interface characteristics: Supports receiving multiple different types of inputs at once. For example, it can receive image and audio data simultaneously.

[0030] For example: The API is ux_agent / input, and the parameter is: [{"type": "text / audio / image", "data": "text string / audio base64 string / image base64 string"}]. The interface can receive multiple different types of inputs at once. For example, the user captures an image on the central control screen and then uses voice to input what's in the picture. The parameter example is [{"type": "image", "data": "captured image base64 string"}, {"type": "audio", "data": "voice base64 string"}].

[0031] 2. The intelligent agent tools include: The tool for calling the vehicle control intelligent agent, control_vehicle, with the parameter command. The tool description is: This tool is very useful when you want to control the vehicle. Parameter: command is the vehicle control instruction refined from the user input.

[0032] The tool for calling the multimedia intelligent agent, control_media, with the parameter command. The tool description is: This tool is very useful when you want to control the multimedia. Parameter: command is the multimedia control instruction refined from the user input.

[0033] The tool for calling the search intelligent agent, search_info, with the parameter content. The tool description is: This tool is very useful when you want to search for Internet information. The searched information includes but is not limited to news, weather, road conditions, points of interest, music, videos, etc. Parameter: content is the search keyword refined from the user input. For example: When the user asks about the current weather information, the interaction intelligent agent can directly search and then give feedback.

[0034] Large models are integrated inside the intelligent agents. Therefore, compared with traditional applications, the input does not need to be very strict, and the input information can be provided in the form of natural language descriptions.

[0035] 3. Intelligent agent system prompt: "You are a vehicle assistant, and complete the corresponding tasks according to the existing tools and the user's instructions. And announce the final execution result by voice when the task is completed."

[0036] 4. Large model access: The interactive agent requires higher intelligence and supports multi-modal capabilities. It adopts the method of accessing cloud models. In the embodiment, the cloud can use the Tongyi Qianwen - Omni-Turbo multi-modal model compatible with OPENAI. A multi-modal model is an artificial intelligence model that can process and integrate multiple different types of data (modalities). These data modalities include, but are not limited to, text, images, speech, videos, sensor signals, etc. Its core goal is to comprehensively understand complex scenarios by integrating information from different modalities and generate richer and more human-cognition-approaching outputs.

[0037] 5. Large model integration: (1)Historical records: The interactive agent needs to maintain the conversation context. It needs to create a new cache to save the historical conversation records. When calling the large model next time, the historical records are used as part of the prompt words. And a time-based forgetting mechanism is adopted to dynamically delete historical records outside the configured time, thereby reducing the number of tokens input to the large model and increasing the response speed of the large model.

[0038] (2)Prompt words: "The prompt words include system prompt words, full tool description information, historical records, and the user information of this session."

[0039] (3)Tool execution: Identify the tool call identifier tool_calls returned by the large model. After traversing the name and arguments of the function of each toolcall and calling the corresponding tool, the result is returned to the large model as a prompt word for judgment.

[0040] As Figure 1 shown, in some embodiments, the function execution agent includes a vehicle control agent and a multimedia agent.

[0041] As an example, the vehicle control agent is used to parse vehicle control task instructions and perform checks based on safety rules. After passing the checks, it calls the vehicle atomic service according to the vehicle control task instructions; the multimedia agent is used to parse audio and video playback task instructions, call the local multimedia atomic service through the multimedia service module or directly call cloud media resources, and support text-to-speech conversion functions.

[0042] In the embodiment, the vehicle control agent can implement instruction parsing, that is, parse the vehicle control task instructions from the vehicle-side core interactive agent to clarify the specific operation content of the instructions, such as determining whether it is an instruction to start the vehicle or adjust the window, etc.

[0043] In some embodiments, the vehicle control agent can perform security checks based on the safety rules of the atomic service to ensure that executing the instruction will not pose a safety risk to the vehicle and its occupants. For example, it checks whether certain vehicle function adjustments are allowed when the vehicle speed is too high.

[0044] The vehicle control agent can call vehicle atomic services through the vehicle service module, convert the instructions that have been parsed and passed the security check into actual vehicle control operations, and achieve precise control of the vehicle.

[0045] The vehicle control agent can be implemented through the following examples: 1. Provide an input interface based on the restful protocol, with the api being vc_agent / input and the parameter being: command; 2. The intelligent agent tool includes: vehicle atomic service call, call_vehicle_atomic_service, with the parameters being api and params. Tool description: Use this tool to obtain vehicle data and vehicle control. Parameters: api is the name of the atomic service, and params are the call parameters.

[0046] 3. System prompt: "You are a vehicle controller. You obtain vehicle data and perform control through atomic services. The current atomic services include: {list of atomic services}". Note: When using atomic services for vehicle control, be sure to check the security rules! The system prompt needs to include the current complete list of atomic services so that the large language model can use the atomic services to obtain vehicle data and perform control.

[0047] 4. Connection to the large language model: The vehicle control agent requires a higher response speed and low latency. It adopts the local access point method. For example, use the orinx chip on the vehicle to deploy the qwen2.5 - 7B compatible with OPENAI. Small parameter models have a faster response speed and can reduce the latency by an average of 30% compared to cloud access.

[0048] 5. Integration of the large language model: (1) Historical records: Each time the vehicle control agent is regarded as a new task, so historical session management is abandoned, thereby reducing the number of tokens input to the large language model to achieve the purpose of accelerating the response of the large language model.

[0049] (2) Prompt: The prompt includes the system prompt, the full tool description information, and the user information of the current session.

[0050] (3) Tool execution: Identify the tool call identifier tool_calls returned by the large language model, traverse the name and arguments of the function of each toolcall, call the corresponding tool, and then return the result as a prompt to the large language model for judgment.

[0051] In the embodiment, the multimedia agent is responsible for multimedia information processing. For example, after receiving the audio and video playback instructions from the user, it obtains the corresponding audio and video resources from the local storage or network for playback; for speech-to-text conversion, it receives the text playback requirements of other agents, etc.

[0052] The multimedia agent can be implemented through the following examples: 1. Provide an input interface based on the restful protocol, with the api being media_agent / input and the parameter being: command.

[0053] 2. The agent tools include: Text-to-speech conversion tool, text_to_speech, with parameter: text. Description: This tool can be used for text-to-speech conversion. The parameter: text is the text to be converted, and it returns the base64 string after speech conversion.

[0054] Multimedia playback tool: media_display, with parameters: type, data. Description: This tool can be used for multimedia playback. The parameter: type is the media type, including image / audio / video, data_type: base64 / file / uri, and data: when data_type is base64, it is the converted base string, when data_type is file, it is the local file path, and when data_type is uri, it is the network resource address.

[0055] File system tool: local_files, with parameter: file_extension. Description: This tool can be used to traverse the files in the current file system. The parameter: file_extension is the file extension to be traversed.

[0056] Tool for calling the search agent, search_info, with parameter content. Tool description: This tool is very useful when you want to search for Internet media information. The searched information includes music, videos, etc. The parameter: content is the search keyword refined from the user input.

[0057] 3. Agent system prompt: "You are a multimedia manager, and complete multimedia-related tasks according to the provided tools and user input."

[0058] 4. Access to the large language model: For better interaction effects, the multimedia agent adopts the method of accessing the cloud model. The cloud uses the Tongyi Qianwen - Omni-Turbo multimodal model compatible with OPENAI.

[0059] 5. Integration of the large language model: (1) Historical record: Each time the multimedia agent is regarded as a new task, so the historical conversation management is abandoned, thereby reducing the number of tokens input to the large language model to accelerate the response of the large language model.

[0060] (2) Prompt: The prompt includes the system prompt, the full tool description information, and the user information of the current conversation.

[0061] (3) Tool execution: Identify the tool call identifier tool_calls returned by the large language model, traverse the name and arguments of the function of each toolcall, call the corresponding tool after that, and then return the result as a prompt to the large language model for judgment.

[0062] As Figure 1 shown, in some embodiments, the intelligent agent module cluster further includes a cloud search intelligent agent configured to obtain road conditions, meteorology, and Internet information through the network and feedback it to the interactive intelligent agent.

[0063] The search intelligent agent can provide the ability to search for road conditions, meteorology, and other Internet information. It can interact with the cloud server, accurately search for relevant content in a vast amount of information according to the user's needs, such as real-time road condition information, weather forecasts, and various Internet information, and feedback the search results to the in-vehicle core interactive intelligent agent.

[0064] The search intelligent agent can be implemented through the following examples: 1. Provide an input interface based on the restful protocol, with the api being searching_agent / input and the parameter being: command.

[0065] 2. The intelligent agent tools include: Aggregated audio and video search tool: media_search, parameter: key_word, description: This tool is used for multimedia information search, including mainstream music search and video search. Parameter: key_word is the search keyword.

[0066] Map reverse coding tool: map_reverse, parameter: location, description: This tool is used to encode the location information described in natural language into longitude and latitude information. Parameter: location: is the location to be reverse-coded.

[0067] Meteorological tool: meteorological_search, parameter: location, days, description: This tool is used for meteorological information search, including meteorological information for the next few days. Parameter location is the search location, and days is the number of days to query.

[0068] Point of Interest (POI) Search Tool: poi_search, with parameters query, location, and radius. It is described as a tool for searching for surrounding POIs. The parameter query is the search content, location is the longitude and latitude information of the center point, and radius is the search radius.

[0069] 3. Agent System Prompt: "You are an information searcher. Search according to the existing tools and user input. Please return a response in JSON format."

[0070] 4. Integration with Large Language Model: Connect to the cloud-based large language model. The search function does not require multi-modal support. Use the Qwen2.5-32B base model compatible with OPENAI.

[0071] 5. Integration of Large Language Model: (1) History Record: Each time the information agent treats it as a new task, so historical conversation management is abandoned, thereby reducing the number of tokens input to the large language model to accelerate the response of the large language model and increasing the concurrency of cloud services.

[0072] (2) Prompt: The prompt includes the system prompt, the full description information of the tools, and the user information of the current conversation.

[0073] (3) Tool Execution: Identify the tool call identifier tool_calls returned by the large language model, traverse the name and arguments of the function of each toolcall, call the corresponding tool, and then return the result as a prompt to the large language model for judgment.

[0074] The embodiment of this application also provides a collaboration method based on the in-vehicle system modular framework provided in the above embodiment, including: the interaction agent receives the user's natural language instruction, parses it into a task instruction through the internal large language model; the interaction agent distributes the task instruction to the corresponding function execution agent according to the task type; the function execution agent extracts key parameters through the internal large language model and performs a security rule check; after the security check passes, the function execution agent calls the atomic service or cloud resource to execute the task and returns the result to the interaction agent, which is summarized by the interaction agent and then fed back to the user.

[0075] The following takes the user instruction "Open the driver's window and play soothing music" as an example for further illustration.

[0076] In the embodiment, the interactive agent receives the user's instruction through voice recognition. The large model infers that the user needs to perform vehicle control and multimedia control based on the currently provided tool capabilities. The vehicle control agent and the multimedia control agent are respectively called, with the parameters being command: open the driver's window for the vehicle control agent and command: play soothing music for the multimedia control agent.

[0077] In the embodiment, the vehicle control agent parses the instruction and infers that the vehicle atomic service to be called is window / control, with the parameter being {"windowId":0, "percentage":100}. The security rule described by this atomic service is: when the vehicle speed is less than 120 km / h, first call the get_vehicle_speed atomic service to obtain the current vehicle speed information to determine whether the security rule passes, and then decide whether to call the window / control atomic service according to the result.

[0078] As an example, the multimedia control agent parses the instruction, infers that music is to be played and no specific track information is specified, starts to call cloud search, and starts playing music based on the music information returned by the cloud search. The core interactive agent determines whether the corresponding operation is successful based on the operation results returned by the vehicle control agent and the multimedia control agent, and generates a description text to call the multimedia control agent for broadcasting.

[0079] In the embodiment, the vehicle service module has the functions of vehicle data acquisition and control. It can interact with the vehicle's underlying hardware system to obtain various types of vehicle operation data in real time, such as vehicle speed, fuel level, tire pressure, etc., and control the relevant functions of the vehicle according to system instructions, such as starting and stopping the engine, adjusting devices such as windows and seats.

[0080] The vehicle service module can be implemented through the following example: Provide vehicle service calls externally based on protocols such as restful.

[0081] 1. Follow the sdv atomic service specification, abstract all vehicle atomic services, and form a complete list of vehicle atomic services. For example, for window control, the api is window / control, and the parameter is {"windowId": window number, "percentage": opening percentage}. To obtain the current window opening information, the api is window / status, and the return value is {"fl": driver's window opening, "fr": front passenger window opening, "rl": left rear window opening, "rr ": right rear window opening}.

[0082] 2. Provide a list of atomic services. The API is api / list, which returns all atomic service operations supported by the current vehicle service. Each item in the list is {"api": "api name", "description": "The description information and security rules of this atomic service", "params": [{"name": "parameter name", "type": "data type", "description": "The description information for this parameter"}], "return": {"type": "the type of the returned data", "description": "The description of the return value"}}. Example: {"api": "window / control", "description": "Used for window control. This service can only be called when the vehicle speed is less than 120 km / s.", "params": [{"name": "windowId", "type": "number", "description": "The number of the window. 0 is the driver's side, 1 is the front passenger side, 2 is the left rear, and 3 is the right rear."}], "return": {"type": "bool", "description": "Indicates whether the operation is successful"}}.

[0083] All the description information in the service list is mainly provided for the large model so that the large model can select the appropriate API for call according to the inference result and fill in the appropriate parameters. Providing security rules allows the large model to judge whether the current conditions meet the security requirements before making a call.

[0084] In the embodiment, the multimedia service module can provide basic image / audio / video input and output capabilities. It can connect to input devices such as in-vehicle cameras and microphones to obtain image and audio information, and can also output the processed audio and video content to devices such as in-vehicle displays and speakers to implement functions such as navigation map display, music playback, and video entertainment.

[0085] In the embodiment, the multimedia service module can be implemented through the following example: Provide basic multimedia playback interfaces based on the restful protocol, including: Image display. The API is display / image, and the parameter is [{"url": "base64 / file / uri: / / data or path"}], and it is displayed using the default image program.

[0086] Audio playback. The API is diaplay / audio, and the parameter is [{"url": "base64 / file / uri: / / data or path"}]. It is played using the default audio player.

[0087] Video playback, the API is diaplay / video, and the parameter is [{"url":"base64 / file / uri: / / data or path"}]. The default video player is used for playback.

[0088] The Url protocol is divided into base64 embedded data, file local file path, and uri network address.

[0089] Four-channel camera image acquisition: the API is camera / capture, and the parameter is {"number":"camera number"}.

[0090] In the embodiment, each intelligent agent interacts through the RESTful protocol standardized interface. The vehicle service and the multimedia service can be abstracted into an atomic service list (such as the window / control interface). Developers do not need to pay attention to the details of the underlying hardware interaction. They only need to call the interface to complete the function development. At the same time, the core interaction intelligent agent automatically realizes cross-module task distribution (such as routing the "play music" instruction to the multimedia intelligent agent), avoiding the complex work of cross-module interface adaptation in the traditional architecture.

[0091] This application can achieve the following beneficial technical effects: 1. Reduce the development and maintenance difficulty: The vehicle-mounted system based on intelligent agents entrusts a large amount of logical judgment to the large language model for processing, which can greatly reduce the regular code. This makes the architecture system of the vehicle-mounted system more unified, and the logical relationship between each functional module is clearer. Developers no longer need to write a large amount of complex and cumbersome rule codes, greatly reducing the difficulty of system function development. At the same time, the unified architecture system also makes the later system maintenance more convenient. Maintenance personnel can locate and solve problems more quickly, reducing the maintenance cost and workload. Through the intelligent agent modular architecture, the typical function development cycle is shortened from 20 person-days in the traditional solution to 5 person-days, and the code volume is reduced by more than 60% (compared with the traditional rule engine solution).

[0092] 2. Flexible expansion of vehicle capabilities: The functions of the vehicle can be flexibly expanded by adding or removing peripheral Agents. For example, if you want to add a schedule assistant function, you only need to add a schedule agent to achieve it. This schedule agent can actively help users complete schedule arrangements according to the schedule information pre-entered by users, such as reminding users of upcoming meetings and combining real-time traffic conditions to plan the best departure time to the meeting place. This flexible expansion method can quickly meet the diverse functional requirements of users without large-scale changes to the underlying architecture of the entire in-vehicle system, enabling the in-vehicle system to be upgraded and expanded in function in a timely manner as user needs change and technology develops. The average access time of the newly added function agent is <2 hours, and it supports OTA remote dynamic loading without the need to upgrade the vehicle software version.

[0093] 3. Enhancement of user interaction experience: The core interaction agent on the vehicle side supports multi-modal user interaction methods. Users can interact with the in-vehicle system through various natural methods such as voice, gesture, and touch, changing the relatively single interaction mode of traditional in-vehicle systems and making operations more convenient and natural. At the same time, its powerful user command intention recognition ability can more accurately understand the real needs of users, avoiding incorrect operations caused by command understanding deviations, and providing users with a more intelligent and efficient interaction experience. The accuracy rate of user command intention recognition based on LLM is over 98%, which is significantly improved compared with the traditional keyword matching algorithm (accuracy rate 85%), especially performing better in complex command scenarios (such as multi-intention combinations).

[0094] 4. Integration of information acquisition channels: The information query function integrates vehicle local services and cloud agents. Users can not only obtain real-time vehicle data but also obtain traffic conditions, meteorology, and other Internet information, realizing one-stop acquisition of information. This provides users with comprehensive and rich information resources. For example, before traveling, users can simultaneously understand vehicle status, real-time traffic conditions, and destination weather, etc., so as to better plan the journey.

[0095] The modular framework of the in-vehicle system and its cooperation method provided in this application have been introduced in detail above. Specific examples are used in this article to elaborate on the principle and implementation method of this application. The description of the above embodiments is only used to help understand the concept of this application and should not be construed as a limitation on the protection scope of this application.

Claims

1. A modular framework for a vehicle-mounted system, characterized in that, It includes: An intelligent agent module cluster set at the vehicle end, where the intelligent agent module cluster includes an interaction intelligent agent and at least one function execution intelligent agent; A large language model collaboration mechanism: large language models are deployed in each intelligent agent; The interaction intelligent agent is used to receive user natural language instructions and parse them using the large language model deployed inside it to generate task instructions; The function execution intelligent agent is used to receive the task instructions, parse them using the large language model deployed inside it, extract key parameters, and perform a security rule check. If the security rule check passes, it executes the corresponding specific task.

2. The modular framework of the vehicle-mounted system according to claim 1, characterized in that The intelligent agent module cluster further includes a search intelligent agent set in the cloud, which is used to obtain road conditions, meteorology, and Internet information through the network and feedback it to the interaction intelligent agent.

3. The modular framework of the vehicle-mounted system according to claim 1, characterized in that The interaction intelligent agent supports multimodal input, including voice, text, and image input, and provides feedback through voice broadcast and graphic display.

4. The modular framework of the vehicle-mounted system according to claim 1, characterized in that The function execution intelligent agent includes a vehicle control intelligent agent and a multimedia intelligent agent: The vehicle control intelligent agent is used to parse vehicle control task instructions, perform a check based on security rules, and call vehicle atomic services according to the vehicle control task instructions if the check passes; The multimedia intelligent agent is used to parse audio and video playback task instructions, call local multimedia atomic services or directly call cloud media resources, and support text-to-speech conversion functions.

5. The modular framework of the vehicle-mounted system according to claim 4, wherein, The large language model supports differential deployment: The interaction intelligent agent uses a cloud multimodal large language model; the vehicle control intelligent agent uses a local lightweight large language model.

6. The modular framework of the vehicle-mounted system according to claim 1, characterized in that Each of the large language models collaborates through a standardized interface. The standardized interface is based on the RESTful protocol and supports the invocation of an atomic service list. The atomic service includes the description information of the atomic service and security rules for the large language model to judge the invocation conditions and perform security rule checks.

7. The collaborative method of the in-vehicle system modular framework according to any one of claims 1 to 6, characterized in that It includes: The interaction intelligent agent receives user natural language instructions and parses them into task instructions through the internal large language model; The interaction intelligent agent distributes the task instructions to the corresponding function execution intelligent agents according to the task type; The function execution intelligent agent extracts key parameters through the internal large language model and performs a security rule check; after the security check passes, the function execution intelligent agent calls atomic services or cloud resources to execute tasks and returns the results to the interaction intelligent agent, which summarizes and feedbacks to the user.

8. The collaborative method of the in-vehicle system modular framework according to claim 7, characterized in that The method further includes: the interaction intelligent agent maintains the context by caching historical session records and optimizes the response speed of the large language model using a time forgetting mechanism.

9. The collaborative method of the in-vehicle system modular framework according to claim 7, characterized in that, When the function execution intelligent agent includes a vehicle control intelligent agent, the method further includes: calling vehicle atomic services through the vehicle service module to obtain real-time vehicle condition data and verifying security rules.

10. The collaborative method of the in-vehicle system modular framework according to claim 7, characterized in that, When the function execution intelligent agent includes a multimedia intelligent agent and a search intelligent agent; the method further includes: the multimedia intelligent agent calls the search intelligent agent to obtain media resources and broadcasts the task results through a text-to-speech conversion tool.

Images