Converging method and system for conversational remote control of a vehicle with progressive convergence of teaching

Abstract

The application provides a conversational remote control vehicle instruction progressive convergence method and system for fusing vehicle teaching, and relates to the technical field of human-computer interaction of Internet of Vehicles. The convergence method performs intention recognition on the current input of a user, and determines a target scene; in the vehicle control scene, the current vehicle atomic capability is used as the only constraint to cut, generate an intention, tools and a front-end component subset, and load or issue; a vehicle control request is received, objects, actions and parameters are extracted and matched with candidate capabilities, and the last round description is consistent and updated; the descriptor and the target capability mandatory parameter are used to determine the completeness state, if it is an explicit atomic state, a complete instruction is output, otherwise an interactive instruction is generated to guide the user to supplement the parameters, subsequent input is received and re-determined until the complete instruction is converged; when the target capability is unavailable, the teaching mode is switched to the control object as the key to retrieve the knowledge base to generate a manual guide. The application realizes the progressive convergence from a natural language request to an executable instruction.

Description

Technical Field

[0001] This invention relates to the field of human-computer interaction technology in the Internet of Vehicles (IoV), and more specifically, to a method and system for progressive convergence of conversational remote vehicle control commands that integrates vehicle teaching. Background Technology

[0002] With the development of vehicle-to-everything (V2X) technology and natural language interaction technology, users can send remote control requests to vehicles through natural language dialogue. However, the following technical problems exist in the existing conversational remote vehicle control process: Due to differences in optional hardware, software configurations, and remote control authorizations, different vehicles have different actual remote control actions and available parameters. However, existing solutions lack the constraint of accurately matching and judging based on the actual atomic vehicle control capabilities of the target vehicle. This can easily lead to matching requests to control functions that the vehicle does not have, or failing to identify capabilities that the vehicle actually possesses but are not covered by the template.

[0003] Meanwhile, the same control intent often appears in natural language forms with different degrees of completeness. Existing solutions lack a unified completeness judgment mechanism based on the actual set of required parameters for the target capability. Fixed templates are difficult to flexibly adapt to different parameter patterns, and large language models that directly generate instructions are prone to guessing or filling in incomplete parameters. All of these make it difficult to give objective and consistent judgment results for the same vehicle control request on different vehicles.

[0004] Furthermore, existing solutions lack differentiated interaction strategies when parameters are missing, typically using a uniform follow-up template for all incomplete requests; and during multi-round dialogues, there is a lack of a unified structured state representation between confirmed parameters, fields to be supplemented, and front-end presentation options, which can easily lead to problems such as repeated follow-up questions, incorrect parameter inheritance, and the continued use of old parameters after changing the control object. Summary of the Invention

[0005] The problem solved by this invention is one or more of the aforementioned related technical problems.

[0006] To address the aforementioned issues, this invention provides a method and system for progressive convergence of conversational remote vehicle control commands that integrates vehicle teaching.

[0007] In a first aspect, the present invention provides a progressive convergence method for conversational remote vehicle control commands that integrates vehicle teaching, comprising: Perform intent recognition on the user's current input to determine the corresponding target scenario; When the target scenario is a remote vehicle control scenario, obtain the set of available atomic vehicle control capabilities corresponding to the current vehicle; Using the set of available atomic vehicle capabilities as a constraint source, a subset of available intentions, a subset of available tools, and a subset of front-end interaction components are generated by matching from the pre-stored atomic vehicle capability mapping library. The subset of available intentions and the subset of available tools are loaded into the dialogue agent, and the subset of front-end interaction components is sent to the user terminal. Receive a user's natural language vehicle control request, extract the control object, control action and parameter fields from the natural language vehicle control request, and match them with the available atomic vehicle control capability set to obtain a candidate capability set; Obtain the structured vehicle control descriptor saved in the previous round of the current session, merge the control object, the control action and the parameter field extracted in this round into the structured vehicle control descriptor, and update the candidate capability information in the structured vehicle control descriptor with the candidate capability set; Based on the updated structured vehicle control descriptor and the set of required parameters for the current target capability, determine the current completeness status; If the current completeness state is a defined atomic state, the updated structured vehicle control descriptor will be used as the complete vehicle control instruction; If the current completeness state is an ambiguous state or a clearly non-atomic state, a structured interaction instruction is generated based on the current completeness state and the set of available atomic vehicle control capabilities. The structured interaction instruction is used to guide the user to supplement or correct the vehicle control parameters in order to conduct multi-round dialogue interaction. The system receives subsequent input from the user terminal based on the structured interaction command, updates the structured vehicle control descriptor, and re-determines the current completeness state until it converges to the complete vehicle control command.

[0008] Optionally, obtaining the set of available atomic vehicle control capabilities corresponding to the current vehicle includes: Obtain the vehicle identifier of the current vehicle; Using the vehicle identifier as the query key, query the vehicle capability library to obtain the set of available atomic control vehicle capabilities for the current vehicle and the corresponding capability snapshot identifier; The set of available atomic vehicle control capabilities is used as the sole source of capability constraints. When a change in the capability snapshot identifier is detected, the saved fields in the structured vehicle control descriptor that do not belong to the updated allowed parameter field are deleted, and the candidate capability set and the current completeness status are recalculated.

[0009] Optionally, the step of merging the controlled object, the controlled action, and the parameter fields extracted in this round into the structured vehicle control descriptor, and updating the candidate capability information in the structured vehicle control descriptor with the candidate capability set, includes: If the control object and control action extracted in this round are consistent with the control object and control action of the target capability in the previous round, the parameter fields extracted in this round will be written into the provided fields of the structured vehicle control descriptor. If the control object or control action extracted in this round is different from the control object or control action of the target capability in the previous round, the provided fields in the structured vehicle control descriptor are cleared, and the target capability information of the structured vehicle control descriptor is updated with the control object and control action extracted in this round to obtain the updated structured vehicle control descriptor.

[0010] Optionally, determining the current completeness status includes: When the number of candidate capabilities in the candidate capability set is not one, or when there is no uniquely determined control object and control action in the candidate capability set, the current completeness state is determined to be the fuzzy state. When the number of candidate capabilities in the candidate capability set is one, there is a unique target capability in the candidate capability set, and there are missing fields in the provided fields of the updated structured vehicle control descriptor relative to the set of required parameters of the unique target capability, or there are fields in the provided fields whose field values ​​do not belong to the corresponding allowed parameter domains, the current completeness state is determined to be the explicit non-atomic state. When the number of candidate capabilities in the candidate capability set is one, the unique target capability exists, the provided fields completely cover the set of required parameters, and the field values ​​of all the provided fields belong to the corresponding allowed parameter domains, the current completeness state is determined to be the explicit atomic state.

[0011] Optionally, the method for determining missing fields in the set of required parameters for the unique target capability includes: Obtain the set of required parameters for the unique target capability; Obtain the set of legal provided fields in the updated structured vehicle descriptor. The legal provided fields consist of provided fields whose field values ​​belong to the corresponding allowed parameter fields. The parameters in the set of required parameters that do not belong to the set of legally provided fields are designated as the missing fields.

[0012] Optionally, generating structured interaction instructions based on the current completeness state and the set of available atomic vehicle control capabilities includes: When the current completeness state is the fuzzy state, the control object and control action corresponding to each candidate capability are extracted from the candidate capability set to generate candidate selection items as the structured interaction instructions. The candidate selection items are subject to the pruning constraint of the available atomic vehicle control capability set. When the current completeness state is the explicitly non-atomic state, the current interaction field is determined from the missing fields, and the allowed parameter field and front-end component type corresponding to the current interaction field are determined according to the available atomic vehicle control capability set. A parameter supplementation instruction containing the current interaction field is generated as the structured interaction instruction.

[0013] Optionally, the progressive convergence method for conversational remote vehicle control commands that integrates vehicle teaching further includes: When the number of candidates in the candidate capability set exceeds a preset display threshold in the ambiguous state, the structured interaction instruction also includes a vehicle control entry instruction, which is used to guide the user to enter the vehicle control interface for selection.

[0014] Optionally, the structured vehicle control descriptor also includes the field source and confirmation round of each field, wherein the field source is used to indicate that the field originates from the current user input, historical user input, or the user's selection of interactive components in the structured interactive instruction.

[0015] Optionally, the progressive convergence method for conversational remote vehicle control commands that integrates vehicle teaching further includes: When the current target capability does not belong to the set of available atomic vehicle control capabilities of the current vehicle, the vehicle manual knowledge base is searched using the controlled object as the search key, a manual operation guide corresponding to the controlled object is generated, and the current dialogue is switched from the vehicle control command mode to the teaching question and answer mode. After switching to the teaching question-and-answer mode and generating the manual operation guidance, if subsequent input from the user is received, the subsequent input is used as the new current input to re-identify the target scenario corresponding to the intent. The target scenario also includes the vehicle teaching scenario. The vehicle teaching scenario includes teaching skills, and the remote vehicle control scenario includes vehicle control skills. The teaching skills and the vehicle control skills are isolated from each other and maintained independently. When the target scenario is the remote vehicle control scenario, the vehicle control skills are loaded and the loading of the teaching skills is prohibited. When the target scenario is the vehicle teaching scenario, the teaching skills are loaded and the loading of the vehicle control skills is prohibited.

[0016] Secondly, the present invention provides a conversational remote vehicle control command progressive convergence system integrating vehicle teaching, applied to the conversational remote vehicle control command progressive convergence method integrating vehicle teaching as described in the first aspect, the system comprising: The intent recognition module is used to recognize the user's current input intent and route it to either a vehicle teaching scenario or a remote vehicle control scenario. The two scenarios share the same dialogue context in the same session. The capability acquisition module is used to acquire the set of available atomic vehicle control capabilities corresponding to the current vehicle when the target scenario is a remote vehicle control scenario. The capability trimming and loading module is used to generate a subset of available intents, a subset of available tools, and a subset of front-end interaction components from a pre-stored atomic vehicle capability mapping library, using the set of available atomic vehicle capabilities as a constraint source. The module loads the subset of available intents and the subset of available tools into the dialogue agent and sends the subset of front-end interaction components to the user terminal. The semantic parsing and candidate matching module is used to extract the control object, control action and parameter fields from the natural language vehicle control request when receiving the user's natural language vehicle control request, and match them with the available atomic vehicle control capability set to obtain a candidate capability set; The state management and merging module is used to obtain the structured vehicle control descriptor saved in the previous round of the current session, merge the control object, the control action and the parameter field extracted in this round into the structured vehicle control descriptor, and update the candidate capability information in the structured vehicle control descriptor with the candidate capability set; The instruction generation and output module is used to determine the current completeness state based on the updated structured vehicle control descriptor and the set of required parameters for the current target capability; if the current completeness state is a definite atomic state, the updated structured vehicle control descriptor is used as a complete vehicle control instruction; if the current completeness state is an ambiguous state or a definite non-atomic state, a structured interaction instruction is generated according to the current completeness state and the set of available atomic vehicle control capabilities. The structured interaction instruction is used to guide the user to supplement or correct the vehicle control parameters for multi-turn dialogue interaction. The loop convergence module is used to receive subsequent inputs returned by the user terminal according to the structured interaction instructions, update the structured vehicle control descriptor and re-determine the current completeness state until it converges to the complete vehicle control instructions.

[0017] The beneficial effects of the present invention, including the dialogic remote vehicle control command progressive convergence method, system, and device for integrating vehicle teaching, are as follows: By acquiring the actual atomic vehicle control capability set of the current vehicle and using this capability set as the sole constraint source, the available intents, callable tools, and front-end interactive components of the dialogue agent are subject to homogeneous pruning. This ensures that subsequent intent recognition, tool invocation, and front-end interaction are all limited to the actual capability boundaries of the vehicle, avoiding presenting users with control functions or parameter options that the vehicle does not possess. Furthermore, this invention receives the user's natural language vehicle control request and extracts the control object, control action, and parameter fields, matching them with the available atomic vehicle control capability set to obtain a candidate capability set. Simultaneously, it retrieves the data saved from the previous round of the current session. The structured vehicle control descriptor merges the information extracted in the current round into the descriptor, ensuring that the parameters supplemented by the user during multi-round dialogue can be correctly inherited, and that the old parameters are not incorrectly used when the controlled object or action changes. Furthermore, the present invention determines the current completeness state based on the updated structured vehicle control descriptor and the set of required parameters of the current target capability. When the parameters are complete, a complete vehicle control command is obtained directly. When the parameters are incomplete or the target is unclear, a structured interactive command is generated to guide the user to supplement or correct the vehicle control parameters. After the user returns subsequent input, the descriptor is updated and the state is re-determined until it converges into a complete vehicle control command. By using the actual set of required parameters for the target capability, rather than a fixed sentence structure, as the criterion for atomicating instructions, the same sentence can obtain objective and repeatable completeness judgment results under different vehicles and capability configurations. At the same time, through the division and differentiated interaction strategies of three states—fuzzy state, explicit non-atomic state, and explicit atomic state—the system guides users to fill in the currently missing field when parameters are missing, rather than uniformly asking for all parameters. This improves the efficiency of multi-round interactions and reduces the user's operational burden. Furthermore, the entire convergence process of this invention is based on the same dialogue entry point and natural language interaction. Users do not need to know the precise function configuration or parameter format of the target vehicle in advance, nor do they need to switch between multiple systems or interfaces. In a unified and progressive manner, open natural language requests are gradually converged into complete vehicle control commands with complete parameters and executable functions. Ultimately, a coherent, efficient, and accurate remote vehicle control interaction experience is achieved in the same conversation context. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a progressive convergence method for conversational remote vehicle control commands that integrates vehicle teaching, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a vehicle control command confirmation, issuance, execution feedback, and downgrade teaching process according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a conversational remote vehicle control command progressive convergence system that integrates vehicle teaching, according to an embodiment of the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0020] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0021] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0022] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0023] The names of the messages or information exchanged between the multiple devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0024] like Figure 1 As shown in the figure, an embodiment of the present invention provides a progressive convergence method for conversational remote vehicle control commands that integrates vehicle teaching, comprising: Step S100: Perform intent recognition on the user's current input to determine the corresponding target scene.

[0025] Specifically, the system first receives the user's current input. This current input includes two forms: one is text spoken by the user in natural language or text converted from speech, such as the user saying "How do I turn on the air conditioner?" or "Turn on the air conditioner"; the other is callback data from the front-end interaction component, such as structured instruction data sent back to the dialogue agent after the user clicks the "Turn on the air conditioner" button on the interface. Both forms of input share the same dialogue entry point and enter the dialogue agent's processing flow through the same interface.

[0026] After acquiring the current input, the dialogue agent invokes the intent recognition module to parse the input. For natural language input, the intent recognition module performs semantic classification on the user input based on a pre-trained semantic understanding model or large language model to determine the user's intent in this statement. Specifically, the intent recognition module maps the user input to one of two target scenario labels: the first is a vehicle teaching scenario, corresponding to the intent of "asking about operation methods," such as expressions asking how to manually operate vehicle functions like "how to turn on the air conditioner" or "how to open the trunk"; the second is a remote vehicle control scenario, corresponding to the intent of "requesting to execute control," such as expressions requesting the vehicle to remotely execute specific control actions like "turn on the air conditioner" or "lower the driver's seat temperature." For structured input from front-end button callbacks, the intent recognition module does not perform natural language semantic understanding but directly classifies it as a remote vehicle control scenario based on the instruction identifier or scenario identifier carried in the callback data. During the intent recognition process, the current input itself is unique input data, and the output of the intent recognition module is the label of the target scenario.

[0027] After completing intent recognition and determining the target scenario, the dialogue agent executes the corresponding scenario routing based on the target scenario's label. If the target scenario is a vehicle teaching scenario, the dialogue agent loads the teaching skill and enters the vehicle usage teaching processing branch; if the target scenario is a remote vehicle control scenario, the dialogue agent loads the vehicle control skill and enters the subsequent vehicle control command convergence process. Both scenarios share the same session context, allowing users to naturally switch between teaching and vehicle control during multi-turn interactions. The dialogue agent can perform referential resolution and omission of responses based on the same session history.

[0028] Through the above process, two types of user input—teaching-related inquiries and vehicle control commands—are uniformly received within a single dialogue entry point. Intent recognition automatically routes input to the corresponding scenario, eliminating the need for users to pre-determine whether their request is "how to do it" or "make it do it," and reducing the need to manually switch between different function entry points, thus lowering the cognitive and operational costs for users. Simultaneously, by isolating and loading the skills corresponding to the teaching and vehicle control scenarios on demand, cross-contamination between teaching prompts and vehicle control tools is avoided, reducing the probability of misjudgment of intent and mis-adjustment of tools. This provides an architectural foundation for the independent iteration and canary release of the two scenarios. Furthermore, the method of directly assigning structured input to the vehicle control scenario for front-end button callbacks allows users to bypass natural language parsing and quickly enter the vehicle control process through graphical interface operations, improving interaction efficiency.

[0029] Step S200: When the target scenario is a remote vehicle control scenario, obtain the set of available atomic vehicle control capabilities corresponding to the current vehicle.

[0030] Step S300: Using the available atomic vehicle capability set as the constraint source, a subset of available intentions, a subset of available tools, and a subset of front-end interaction components are generated by matching from the pre-stored atomic vehicle capability mapping library. The subset of available intentions and the subset of available tools are loaded into the dialogue agent, and the subset of front-end interaction components is sent to the user terminal.

[0031] Specifically, once the target scenario is determined to be a remote vehicle control scenario, the dialogue agent obtains the set of available atomic vehicle control capabilities corresponding to the current vehicle. The set of available atomic vehicle control capabilities reflects the summary of remote control actions that the current vehicle can actually execute under its actual hardware configuration, software functions, and remote control authorization status. The specific content of this set varies from vehicle to vehicle due to differences in optional configurations, software versions, and authorizations.

[0032] The atomic vehicle control capability refers to the minimum executable capability that performs a specific action on a specific vehicle control object, and the range of values ​​for the parameters required to complete the action is already determined. Each atomic vehicle control capability record includes a capability identifier, a control object, a control action, and a parameter pattern. The capability identifier is a stable and unique code for the atomic vehicle control capability. For example, "turn on seat ventilation" can be represented as object=seat ventilation, action=turn on, parameterDomain.seat={driver's seat,passenger's seat}. The control object represents the target component of the capability, such as seat ventilation, air conditioning, or windows. The control action represents the operation performed on the object, such as "turn on," "turn off," or "adjust." The parameter pattern records all parameters required to execute the capability and their allowed value ranges. For example, for the atomic capability "turn on seat ventilation," its parameter pattern records the required parameter "seat," which can be either "driver's seat" or "passenger's seat."

[0033] The dialogue agent uses the acquired set of available atomic vehicle capabilities as the sole source of capability constraints. It matches and generates a subset of available intents, a subset of available tools, and a subset of front-end interaction components from a pre-stored atomic vehicle capability mapping library. The available intent and tool subsets are loaded into the dialogue agent, while the front-end interaction component subset is distributed to the user terminal. The atomic vehicle capability mapping library is a pre-established data storage structure that records the correspondence between atomic vehicle capabilities and dialogue intents, tool calls, and front-end components. Each mapping record in the library includes a capability identifier, an intent identifier, a tool identifier, a tool parameter structure definition, and a front-end component identifier. The intent identifier is a label used at the natural language understanding level to identify user intent; the tool identifier is a unique identifier for a backend function or application programming interface that the dialogue agent can call; the tool parameter structure definition records the name, type, and constraints of the parameters required to call the tool; and the front-end component identifier is a template identifier used to present interactive controls in the user interface.

[0034] The dialogue agent uses the capability identifiers in the set of available atomic vehicle control capabilities as matching keys to retrieve and extract corresponding mapping records from the atomic vehicle control capability mapping library, thereby generating three sets of corresponding subsets: the available intent subset, the available tool subset, and the front-end interaction component subset. The available intent subset is a set of all intent tags extracted from the mapping library corresponding to the vehicle control capabilities actually supported by the current vehicle, used to constrain the dialogue agent to only recognize the vehicle control intents that the current vehicle can execute during subsequent semantic understanding. The available tool subset is a set of all callable tools and their parameter structure definitions corresponding to the vehicle control capabilities actually supported by the current vehicle, extracted from the mapping library, used to constrain the dialogue agent to only call the vehicle control interfaces that the current vehicle actually possesses during subsequent execution. The front-end interaction component subset is a set of all front-end interaction control templates and their parameter options corresponding to the vehicle control capabilities actually supported by the current vehicle, extracted from the mapping library, used to constrain the user interface to only present the interaction entry points and parameter options that are actually available to the current vehicle. After generating the above three subsets, the dialogue agent loads the available intent subset and the available tool subset into its own runtime environment, making them boundary constraints for subsequent semantic parsing, intent matching, and tool invocation decisions; at the same time, it sends the front-end interaction component subset to the user terminal in the form of structured data, and the user terminal dynamically renders the interactive interface accordingly.

[0035] Step S400: Receive the user's natural language vehicle control request, extract the control object, control action, and parameter fields from the natural language vehicle control request, and match them with the available atomic vehicle control capability set to obtain a candidate capability set.

[0036] Step S500: Obtain the structured vehicle control descriptor saved in the previous round of the current session, merge the control object, the control action and the parameter field extracted in this round into the structured vehicle control descriptor, and update the candidate capability information in the structured vehicle control descriptor with the candidate capability set.

[0037] Specifically, in step S400, the dialogue agent receives the user's natural language vehicle control request. This natural language request is a text expressing the user's intention to control the vehicle, input by speaking or typing, such as "turn on the seat ventilation," "cool down the car," or "turn on the driver's seat ventilation." The dialogue agent performs semantic parsing on the natural language request, extracting three core elements: the control object, the control action, and the parameter field. The control object represents the target vehicle component the user wishes to operate, such as "seat ventilation," "air conditioning," or "windows." The control action represents the type of operation the user wishes to perform on the control object, such as "open," "close," or "adjust." The parameter field represents the specific parameter value explicitly specified by the user; for example, in "turn on the driver's seat ventilation," "driver's seat" is the field value provided for the "seat" parameter. This extraction can be achieved through rule matching, natural language understanding models, or large language models, converting colloquial, open-ended expressions into structured element representations.

[0038] After element extraction, the dialogue agent matches the extracted control object, control action, and parameter fields with the set of available atomic vehicle control capabilities loaded in step S300. Each atomic vehicle control capability in the set of available atomic vehicle control capabilities includes a capability identifier, control object, control action, and parameter pattern. Therefore, by comparing the control object and control action in the user request with the object and action in the capability record, the dialogue agent identifies all atomic vehicle control capabilities that match the user's intent, thus obtaining a candidate capability set. The candidate capability set is a collection of all atomic vehicle control capabilities currently supported by the vehicle and semantically related to the user's current request. Each candidate capability includes a capability identifier and the fields extracted in this round. For example, when a user says "cool down the car," the matching result may simultaneously include two candidate capabilities: "turn on the air conditioning" and "open the windows." When a user says "turn on the seat ventilation," if the vehicle only supports seat ventilation, the candidate capability set will only include this single capability.

[0039] In step S500, the dialogue agent first obtains the structured vehicle control descriptor saved in the previous round of the current session. The structured vehicle control descriptor is a core data object that maintains the full state of a single vehicle control session, and at least includes a set of candidate capabilities, a target capability identifier, provided fields, and field values. This descriptor is saved to the session storage after each round of processing for retrieval in the next round, thus continuously accumulating confirmed information throughout the multi-round dialogue process.

[0040] After obtaining the structured vehicle control descriptor from the previous round, the dialogue agent merges the control object, control action, and parameter fields extracted in this round into the descriptor, updates the candidate capability information in the structured vehicle control descriptor with the obtained candidate capability set, writes the candidate capability set matched in this round into the descriptor, and replaces the candidate capability record from the previous round, thereby obtaining the updated structured vehicle control descriptor for subsequent steps to determine its completeness.

[0041] It should be noted that the current input received by the dialogue agent in step S100 is raw input data, containing both natural language text and structured data from the front-end button callback, used for intent recognition to determine the target scenario. In step S400, the user's natural language vehicle control request is only the portion of the input that is routed to the remote vehicle control scenario and is in natural language text form, used for semantic parsing to extract the control object, control action, and parameter fields. When the user inputs "how to turn on the air conditioner" and is routed to the teaching scenario, this input belongs to the current input but not to the natural language vehicle control request; when the user inputs "turn on the air conditioner" and is routed to the vehicle control scenario, this input belongs to both the current input and the natural language vehicle control request; and when the user clicks the front-end button, the current input is structured callback data, not a natural language vehicle control request, and the dialogue agent directly reads the structured elements from the callback data without going through the semantic extraction process of step S400.

[0042] By using a dialogue agent, the system transforms users' open-ended natural language requests into structured vehicle control elements, which are then matched with the vehicle's actual capabilities to obtain a set of candidate capabilities. Simultaneously, structured vehicle control descriptors are used to achieve unified maintenance and updating of multi-turn dialogue states. This allows users to gradually supplement vehicle control parameters through natural language during multiple rounds of dialogue, without needing to provide all information at once. The system can continuously accumulate confirmed information based on the descriptors saved from the previous round, laying a unified data foundation for subsequent differentiated interactive guidance based on completeness states.

[0043] Step S600: Based on the updated structured vehicle control descriptor and the set of required parameters for the current target capability, determine the current completeness status.

[0044] Step S700: If the current completeness state is a defined atomic state, the updated structured vehicle control descriptor is used as the complete vehicle control instruction.

[0045] Step S800: If the current completeness state is an ambiguous state or a clearly non-atomic state, a structured interaction instruction is generated based on the current completeness state and the set of available atomic vehicle control capabilities. The structured interaction instruction is used to guide the user to supplement or correct the vehicle control parameters in order to conduct multi-round dialogue interaction.

[0046] Specifically, in step S600, the dialogue agent determines the completeness state of the current vehicle control request based on the structured vehicle control descriptor updated in step S500 and the required parameter set of the current target capability. The updated structured vehicle control descriptor includes at least a candidate capability set and provided fields. The provided fields record all parameters and corresponding field values ​​provided by the user in the current and historical rounds. The current target capability refers to the atomic vehicle control capability locked in the candidate capability set for subsequent execution. The required parameter set is the set of all parameters recorded in the parameter pattern of the target capability that must be provided to execute the capability. The dialogue agent determines the current completeness state based on the coverage of the provided fields relative to the required parameter set. The current completeness state includes three types: explicit atomic state, ambiguous state, and explicit non-atomic state.

[0047] In step S700, the dialogue agent directly outputs the updated structured vehicle control descriptor as a complete vehicle control command. This complete vehicle control command is a structured data object with complete parameters, usable for subsequent verification and execution. It includes a uniquely identified target capability identifier and valid field values ​​for all required parameters. For example, if the user has explicitly stated that the driver's seat ventilation is enabled, and the vehicle's ability to enable seat ventilation is missing only one required parameter, and the user has provided the driver's seat, the dialogue agent will output a complete vehicle control command containing the target capability "enable seat ventilation" and the parameter "seat = driver's seat".

[0048] In step S800, when the determined current completeness state is an ambiguous state or a definite non-atomic state, the dialogue agent generates a structured interaction instruction based on the current state and the set of available atomic vehicle control capabilities, and returns the structured interaction instruction to the user terminal. The structured interaction instruction is a structured data object output by the dialogue agent to the user terminal to guide the user in supplementing or correcting vehicle control parameters. The user terminal renders the corresponding interactive control based on the instruction. After the user selects or inputs through the interactive control, the user terminal encapsulates the user's operation result as subsequent input and returns it to the dialogue agent. After the user terminal returns subsequent input based on the structured interaction instruction, the dialogue agent re-executes steps S400 to S800, that is, performs semantic parsing, merging and updating, and completeness determination again, until convergence to the complete vehicle control instruction.

[0049] By using the actual set of required parameters for the current target capability, rather than a fixed sentence structure, as the judgment criterion for the dialogue agent, the same natural language expression can obtain objective and repeatable completeness judgment results under different vehicles and capability configurations. This avoids the problem of fixed templates not being able to adapt to different parameter modes, and also avoids the risk of erroneous execution caused by the model guessing or filling in parameters when they are incomplete. Simultaneously, by distinguishing between ambiguous states and clear non-atomic states and providing corresponding differentiated interactive guidance, when the target is unclear, the user is guided to select candidates to eliminate ambiguity; when the target is clear but parameters are missing, the user is only guided to fill in the currently missing fields, rather than asking for all parameters or using a uniform template. This improves the efficiency of multi-turn interactions and reduces the user's operational burden. Furthermore, structured interaction commands serve as a standardized data interface between the dialogue agent and the user end, ensuring that the rendering of front-end interactive components is entirely driven by the judgment results of the dialogue agent. The user end is only responsible for presentation and feedback, avoiding the problem of misalignment caused by inconsistent state representations between confirmed parameters, fields to be filled, and front-end options.

[0050] Step S900: Receive subsequent input from the user terminal based on the structured interaction instruction, update the structured vehicle control descriptor and re-determine the current completeness state until convergence to the complete vehicle control instruction.

[0051] Specifically, in step S900, the dialogue agent receives subsequent input returned by the user terminal based on the structured interaction instruction. This subsequent input is data that the user terminal encapsulates and sends back to the dialogue agent after the user selects or inputs data through interactive controls following the receipt of the structured interaction instruction generated in step S800. The data format of the subsequent input varies depending on the triggering method: when the user responds using natural language, the subsequent input is natural language text; when the user selects data by clicking on a front-end interactive control, the subsequent input is structured callback data.

[0052] After receiving the subsequent input, the dialogue agent uses it as the new input for this round and re-executes the complete processing flow from steps S400 to S800. Specifically, the dialogue agent first extracts the control object, control action, and parameter fields from the subsequent input. If the subsequent input is structured callback data, it directly reads the already structured elements. Then, it obtains the structured vehicle control descriptor saved in the previous round, merges the elements extracted in this round into the descriptor, and updates the candidate capability information in the descriptor with the candidate capability set obtained from the rematch. Subsequently, based on the updated structured vehicle control descriptor and the required parameter set of the current target capability, it re-determines the current completeness state. Finally, based on the re-determined completeness state, if it is a definite atomic state, it outputs a complete vehicle control command; if it is still an ambiguous state or a definite non-atomic state, it generates a structured interaction command again and returns it to the user.

[0053] The above process constitutes a complete iterative loop. Each time the dialogue agent receives subsequent input from the user, it executes a round of "parsing-merging-determining-responding" processing. After each round, the updated structured vehicle control descriptor is saved to the session storage for retrieval in the next round. During this iteration, the structured vehicle control descriptor continuously accumulates the fields provided by the user and dynamically updates candidate capability information until a round determines that the current completeness state is a definite atomic state. At this point, the dialogue agent no longer generates structured interaction instructions but outputs the updated structured vehicle control descriptor as a complete vehicle control instruction. This complete vehicle control instruction is the final result of asymptotic convergence. Convergence refers to the process by which a vehicle control instruction, initially an open natural language expression, gradually clarifies the target capability and completes all required parameters through one or more rounds of dialogue interaction, ultimately becoming a parameter-complete and executable structured instruction. The phrase "until convergence" indicates that the iterative loop continues, with the achievement of a definite atomic state as the loop termination condition, rather than being limited to a fixed number of interaction rounds.

[0054] For example, if the user's first input is "turn on seat ventilation," the dialogue agent determines it to be a definite non-atomic state (missing seat parameter), and generates a structured interaction command containing seat selection controls. If the user returns with the subsequent input "driver's seat," the dialogue agent writes "seat = driver's seat" into the provided fields of the structured vehicle control descriptor, re-determines it to be a definite atomic state, and outputs a complete vehicle control command containing the target capability "turn on seat ventilation" and the parameter "seat = driver's seat," then the iteration terminates. If the user's first input is "cool down the car," the dialogue agent determines it to be a fuzzy state (matching two candidates: turn on the air conditioning and open the windows), and generates a structured interaction command containing candidate options. If the user returns with the subsequent input "turn on the air conditioning," the dialogue agent locks the target capability as "turn on the air conditioning," re-determines and finds that the temperature parameter is missing, and generates another structured interaction command containing temperature adjustment controls. If the user returns with the subsequent input "26 degrees," the dialogue agent writes "temperature = 26 degrees" into the provided fields, re-determines it to be a definite atomic state, outputs a complete vehicle control command, and the iteration terminates.

[0055] Based on a dialogue agent, structured vehicle control descriptors enable unified maintenance and continuous accumulation of multi-turn conversation states. This allows users to gradually supplement parameters through natural language or front-end interaction during multiple rounds of dialogue, rather than providing a complete vehicle control command all at once. After each supplement, the system reassesses the completeness and provides targeted guidance for the next round until the command is complete and executable. This progressive convergence mechanism avoids execution failures caused by incomplete parameters or guesswork in traditional solutions. Users do not need to understand the specific parameter format or functional boundaries of the target vehicle, nor do they need to switch between multiple interfaces. They can complete the construction of complete remote vehicle control commands simply through natural language interaction at the same dialogue entry point. This significantly reduces the user's operational threshold and cognitive burden, and improves the interactive coherence and task completion rate of multi-turn dialogue-based vehicle control.

[0056] In this embodiment, by acquiring the actual atomic vehicle control capability set of the current vehicle and using this capability set as the sole constraint source, the available intents, callable tools, and front-end interactive components of the dialogue agent are subject to homogeneous pruning. This ensures that subsequent intent recognition, tool invocation, and front-end interaction are all limited by the actual capability boundaries of the vehicle, avoiding presenting users with control functions or parameter options that the vehicle does not possess. Furthermore, the invention receives the user's natural language vehicle control request and extracts the control object, control action, and parameter fields, matching them with the available atomic vehicle control capability set to obtain a candidate capability set. Simultaneously, it acquires the previous round of the current session. The saved structured vehicle control descriptor incorporates the information extracted in the current round, ensuring that parameters added by the user during multi-round dialogues are correctly inherited and that old parameters are not incorrectly used when the controlled object or action changes. Furthermore, this invention determines the current completeness state based on the updated structured vehicle control descriptor and the set of required parameters for the current target capability. When the parameters are complete, a complete vehicle control command is obtained directly. When the parameters are incomplete or the target is unclear, a structured interactive command is generated to guide the user to supplement or correct the vehicle control parameters. After the user returns subsequent input, the descriptor is updated and the state is re-determined until it converges to a complete vehicle control command.

[0057] Optionally, obtaining the set of available atomic vehicle control capabilities corresponding to the current vehicle includes: Obtain the vehicle identifier of the current vehicle; Using the vehicle identifier as the query key, query the vehicle capability library to obtain the set of available atomic control vehicle capabilities for the current vehicle and the corresponding capability snapshot identifier; The set of available atomic vehicle control capabilities is used as the sole source of capability constraints. When a change in the capability snapshot identifier is detected, the saved fields in the structured vehicle control descriptor that do not belong to the updated allowed parameter field are deleted, and the candidate capability set and the current completeness status are recalculated.

[0058] In some optional embodiments, the dialogue agent first obtains the vehicle identifier of the current vehicle. The vehicle identifier is a unique identification number for the vehicle, such as the vehicle identification number (VIN), used to uniquely identify a specific vehicle entity in the vehicle-to-everything (V2X) system. After obtaining the vehicle identifier, the dialogue agent uses the vehicle identifier as the query key to initiate a query request to the vehicle capability library. The vehicle capability library is a pre-built data storage system that stores the actual available vehicle control capabilities of each vehicle. Its data sources include vehicle factory configuration information, optional hardware information, software version information, V2X access status, and remote control authorization status. The vehicle capability library performs a search based on the vehicle identifier, returning the set of available atomic vehicle control capabilities corresponding to the current vehicle, and simultaneously returning the corresponding capability snapshot identifier. The capability snapshot identifier is a version identifier generated by the vehicle capability library for each set of capabilities returned in each query, used to mark the version status of the capability set; when the vehicle's hardware configuration, software version, or remote control authorization changes, the capability record of that vehicle in the vehicle capability library is updated accordingly, and the capability snapshot identifier changes to reflect the version evolution of the capability set.

[0059] After acquiring the set of available atomic vehicle control capabilities, the dialogue agent uses this set as the sole source of capability constraints for all subsequent processing stages. Subsequent intent recognition, tool invocation, and front-end interaction all adhere to this set as their boundary, ensuring they do not exceed the capability range defined by it. Simultaneously, the dialogue agent associates the capability snapshot identifier with the session state for use in subsequent processing to detect whether the vehicle's capabilities have changed.

[0060] During subsequent vehicle control dialogues, the dialogue agent checks whether the current capability snapshot identifier matches the previously saved capability snapshot identifier each time it enters a vehicle control scenario or at a key decision node. When a change in the capability snapshot identifier is detected, it indicates that the set of available vehicle control capabilities has changed. For example, the vehicle may have acquired new vehicle control functions through an OTA upgrade, the remote control authorization may have been adjusted, or the available parameter fields for certain functions may have changed. In this case, the dialogue agent deletes saved fields in the structured vehicle control descriptor that do not belong to the updated allowed parameter fields. The allowed parameter fields are the allowed value ranges for each parameter recorded in the parameter pattern of the atomic vehicle control capabilities. For example, the allowed parameter field for the "seat" parameter is {driver, passenger}. When a change in the capability snapshot identifier causes a change in the allowed parameter field of a certain parameter, the previously saved field values, if no longer belonging to the new allowed parameter field, are no longer valid and should be deleted to avoid incorrect inheritance. After deletion, the dialogue agent recalculates the candidate capability set and the current completeness state based on the updated set of available atomic vehicle control capabilities and the cleaned structured vehicle control descriptors, ensuring that subsequent vehicle control command convergence is always based on the actual capability state of the current vehicle.

[0061] Through the above process, this implementation achieves automatic perception of vehicle capability changes and coordinated clearing of session states. When the available vehicle control capabilities change due to software upgrades, authorization adjustments, or configuration changes, the dialogue agent can promptly perceive the change through the capability snapshot identifier and automatically delete no longer valid saved fields in the structured vehicle control descriptor. This avoids the problem of historical parameters not matching current capabilities due to vehicle capability changes. For example, if the vehicle previously supported passenger seat ventilation and the user provided the "passenger" parameter but the session had not yet converged, and subsequently the vehicle no longer supports passenger seat ventilation due to authorization adjustments, without clearing this field, subsequent judgments may incorrectly treat "passenger" as a valid parameter and output unexecutable commands. At the same time, recalculating the candidate capability set and completeness state ensures that subsequent interactions are always based on the latest capability state, without requiring the user to manually refresh or re-initiate requests, thus guaranteeing the accuracy and reliability of vehicle control command convergence in long-term dialogue scenarios.

[0062] Optionally, the step of merging the controlled object, the controlled action, and the parameter fields extracted in this round into the structured vehicle control descriptor, and updating the candidate capability information in the structured vehicle control descriptor with the candidate capability set, includes: If the control object and control action extracted in this round are consistent with the control object and control action of the target capability in the previous round, the parameter fields extracted in this round will be written into the provided fields of the structured vehicle control descriptor. If the control object or control action extracted in this round is different from the control object or control action of the target capability in the previous round, the provided fields in the structured vehicle control descriptor are cleared, and the target capability information of the structured vehicle control descriptor is updated with the control object and control action extracted in this round to obtain the updated structured vehicle control descriptor.

[0063] Optionally, the structured vehicle control descriptor also includes the field source and confirmation round of each field, wherein the field source is used to indicate that the field originates from the current user input, historical user input, or the user's selection of interactive components in the structured interactive instruction.

[0064] In some embodiments, when the dialogue agent merges the control object, control action, and parameter fields extracted in the current round into the structured vehicle control descriptor, it first compares whether the control object and control action extracted in the current round are consistent with the control object and control action of the target capability in the previous round. The target capability in the previous round refers to the atomic vehicle control capability corresponding to the target capability identifier stored in the structured vehicle control descriptor after the previous round of processing, that is, the vehicle control capability that was locked in the previous round of dialogue; its control object and control action record the target component that the capability acts on and the corresponding operation type. For example, if the user was configuring "turn on seat ventilation" in the previous round, then the control object of the target capability in the previous round is "seat ventilation" and the control action is "turn on".

[0065] If the comparison results are consistent, meaning the user is still supplementing information for the same target capability in this round, the dialogue agent will write the parameter fields extracted in this round into the provided fields of the structured vehicle control descriptor. The provided fields are data record areas in the structured vehicle control descriptor used to store parameter key-value pairs that the user has legally confirmed. For example, if the user inputs "driver" in this round, the dialogue agent will write "seat = driver" into the provided fields, which, together with other fields saved in the previous round, constitute the complete set of parameters currently provided.

[0066] If the comparison results are inconsistent, meaning the control object or control action extracted in this round is different from the target capability in the previous round, it indicates that the user has changed the control target; for example, in the previous round, the user said "turn on the seat ventilation," but in this round, they say "then let's turn on the air conditioning." At this point, the dialogue agent clears the provided fields in the structured vehicle control descriptor and updates the target capability information of the structured vehicle control descriptor with the control object and control action extracted in this round, switching the target capability to the new atomic vehicle control capability, resulting in an updated structured vehicle control descriptor. This clearing operation ensures that the parameters collected in the previous round for the old target capability are not incorrectly inherited to the new target capability; for example, "seat = driver's seat" will not be incorrectly passed to the new target "turn on the air conditioning."

[0067] More preferably, the structured vehicle control descriptor further includes the field source and confirmation round for each field. The field source indicates the generation path of each provided field, specifically including three source types: from current user input, indicating that the field was directly provided by the user in this round of natural language request; from historical user input, indicating that the field was provided by the user in a previous round; and from the user's selection of an interactive component in the structured interaction command, indicating that the field was generated by the user clicking a front-end interactive control (such as selecting the "Driver's Seat" button). The confirmation round records which round of dialogue each provided field is finally confirmed. Whenever a field is legally provided, the dialogue agent records the current round number as the confirmation round for that field. Recording the source of fields and the confirmation rounds provides precise traceability for subsequent state management. That is, when it is necessary to determine whether a field should be retained, whether it has been updated, or whether it needs to be cleaned up due to changes in the capability snapshot, the dialogue agent can make accurate decisions based on the source of the field and the confirmation rounds. This avoids the uncertainty caused by judging solely based on the field value itself, prevents confusion between the values ​​automatically returned by the front end and the values ​​actively spoken by the user, and avoids the problem of out-of-order delivery caused by network latency.

[0068] Through the above merging and updating process, accurate inheritance and automatic clearing of vehicle control parameters in multi-turn dialogues are achieved. When a user continuously adds parameters under the same vehicle control intention, previously provided fields are retained, and the user only needs to state the new parameter without repeating previously confirmed information. When the user changes the control object or control action, the old parameters are automatically cleared, fundamentally avoiding the problem of incorrect parameter cross-use in traditional solutions. For example, the "driver's seat" parameter provided when "turn on seat ventilation" will not be incorrectly passed to the subsequent "turn on air conditioning" request. Meanwhile, the recording of field sources and confirmation rounds provides traceable metadata for each parameter, enabling the dialogue agent to accurately identify the generation path and confirmation sequence of each parameter in complex scenarios such as capability changes, session resumption, or anomaly handling. This provides reliable traceability assurance for the continuous maintenance and accurate convergence of the structured vehicle control descriptor. Thus, even in complex interaction scenarios such as repeated user modifications to intents, mid-process switching of control targets, changes in vehicle capabilities, or mixed parameter sources during multi-round dialogues, the dialogue agent can still accurately maintain the attribution and timing of each field, avoiding misjudgments and invalid follow-up questions caused by state synchronization failures or incorrect parameter inheritance. Users do not need to pay attention to the state management details behind the dialogue; they only need to continuously express their vehicle control needs in a natural and coherent manner to obtain stable and predictable interactive feedback and accurate vehicle control results.

[0069] Optionally, determining the current completeness status includes: When the number of candidate capabilities in the candidate capability set is not one, or when there is no uniquely determined control object and control action in the candidate capability set, the current completeness state is determined to be the fuzzy state. When the number of candidate capabilities in the candidate capability set is one, there is a unique target capability in the candidate capability set, and there are missing fields in the provided fields of the updated structured vehicle control descriptor relative to the set of required parameters of the unique target capability, or there are fields in the provided fields whose field values ​​do not belong to the corresponding allowed parameter domains, the current completeness state is determined to be the explicit non-atomic state. When the number of candidate capabilities in the candidate capability set is one, the unique target capability exists, the provided fields completely cover the set of required parameters, and the field values ​​of all the provided fields belong to the corresponding allowed parameter domains, the current completeness state is determined to be the explicit atomic state.

[0070] Specifically, after updating the structured vehicle control descriptor, the dialogue agent determines the completeness status of the current vehicle control request based on the updated structured vehicle control descriptor and the set of required parameters for the current target capability. The determination of the completeness status is based on the composition of the candidate capability set, the locking status of the unique target capability, the coverage of the provided fields relative to the set of required parameters, and the validity of the provided field values.

[0071] The dialogue agent analyzes the candidate capability set: if the number of candidate capabilities in the set is not one, it indicates that the current matching result is empty or multiple candidates are simultaneously related to the user's request. In this case, the dialogue agent cannot uniquely determine which specific vehicle control capability the user's intent points to, and therefore determines the current completeness state as fuzzy. Similarly, even if the number of candidate capabilities is one, if the control object or control action of the unique candidate still has uncertainties at the semantic level, the dialogue agent also determines the current state as fuzzy. The fuzzy state indicates that the user's intent is not yet clear and needs further clarification before subsequent vehicle control command convergence.

[0072] When the number of candidate capabilities in the candidate capability set is one, and this unique candidate capability is locked as the unique target capability, the dialogue agent further examines the coverage of the provided fields in the updated structured vehicle control descriptor relative to the required parameter set of the unique target capability. The unique target capability refers to the atomic vehicle control capability that uniquely exists in the candidate capability set and is determined by the dialogue agent to be the target of the user's intent, i.e., the target object for subsequent vehicle control commands. The required parameter set is the set of all parameters recorded in the parameter pattern of the unique target capability that must be provided to execute the capability. The dialogue agent compares each parameter in the provided fields with the required parameter set one by one. If a parameter in the required parameter set does not appear in the provided fields, that parameter is identified as a missing field; if a field in the provided fields has a value that does not belong to the corresponding allowed parameter domain, that field is identified as a field to be corrected. The missing fields refer to parameters that are not yet available before the target capability is executed, such as the "seat" parameter missing for "turn on seat ventilation". The fields to be corrected refer to fields that have been provided by the user but whose values ​​exceed the allowed range of the parameter, such as the user providing "rear seats" but the vehicle's "seat" parameter only allowing "driver's seat" or "passenger's seat". At this point, the dialogue agent determines that the current completeness state is a definite non-atomic state, indicating that the user's intent has clearly pointed to a specific atomic vehicle control capability, but there are still missing or illegal parameters, which need to be supplemented or corrected by the user through subsequent interactions.

[0073] When the number of candidate capabilities in the candidate capability set is one, the unique target capability exists, and the provided fields completely cover all parameters in the required parameter set, and the field values ​​of all provided fields belong to the corresponding allowed parameter domains, the dialogue agent determines the current completeness state as a defined atomic state. The defined atomic state indicates that the vehicle control request has met all execution conditions—the user intent is uniquely determined, all required parameters have been legally provided, and the dialogue agent can directly output the structured vehicle control descriptor as a complete vehicle control command without further interaction.

[0074] For example, when a user says "turn on seat ventilation," the dialogue agent matches the user's request with the set of available atomic vehicle control capabilities. The candidate capability set only contains "turn on seat ventilation," and this capability is locked as the sole target capability. The dialogue agent then checks the provided fields: at this point, the user has not yet provided the "seat" parameter, and the required parameter set for this target capability includes "seat," therefore a missing field is determined, and the current state is a declarative non-atomic state. If the user then says "driver's seat," the dialogue agent writes "seat = driver's seat" into the provided fields. Upon further checking, the provided fields completely cover the required parameter set, and the field values ​​are valid, transitioning the state to a declarative atomic state, and outputting the complete vehicle control command. If the user directly says "turn on driver's seat ventilation," the dialogue agent immediately determines it to be a declarative atomic state and directly outputs the command.

[0075] By classifying and judging these three states, the dialogue agent uses the actual set of required parameters for the target capability, rather than a fixed sentence structure, as the judgment criterion. This allows the same sentence to obtain objective and repeatable judgment results under different vehicles and capability configurations. For example, the sentence "turn on the air conditioner" is directly judged as a clear atomic state on a vehicle where the air conditioner can be remotely turned on without specifying a temperature, while on another vehicle where the air conditioner needs to be remotely turned on with a specified temperature, it is judged as a clear non-atomic state, guiding the user to fill in the temperature value. Thus, natural language requests with different levels of completeness are all included in a unified judgment framework. The dialogue agent determines the subsequent interaction strategy based on the current state rather than a preset template. When the target is unclear, it guides the user to select candidates to eliminate ambiguity. When the target is clear but the parameters are missing, it guides the user to fill in the missing fields. This avoids recognition failures, mis-executions, or inefficient follow-up questions caused by different levels of completeness in user requests, improving the adaptability and interaction efficiency of conversational vehicle control under heterogeneous vehicles and diverse user expressions.

[0076] Optionally, the method for determining missing fields in the set of required parameters for the unique target capability includes: Obtain the set of required parameters for the unique target capability; Obtain the set of legal provided fields in the updated structured vehicle descriptor. The legal provided fields consist of provided fields whose field values ​​belong to the corresponding allowed parameter fields. The parameters in the set of required parameters that do not belong to the set of legally provided fields are designated as the missing fields.

[0077] Specifically, before determining the current completeness state, the dialogue agent first identifies the specific composition of the missing fields in the required parameter set of the unique target capability. The unique target capability is the only existing atomic vehicle control capability in the candidate capability set, which the dialogue agent identifies as the target of the user's intent. The dialogue agent obtains the required parameter set of this unique target capability. This required parameter set is the set of all parameters recorded in the parameter pattern of the target capability that must be provided to execute the capability. For example, the required parameter set for "turn on seat ventilation" includes "seat," and the required parameter set for "turn on air conditioning" may include one or more of "temperature" and "fan speed." Simultaneously, the dialogue agent obtains the set of legally provided fields in the updated structured vehicle control descriptor. These legally provided fields are subsets of those fields whose values ​​belong to the corresponding allowed parameter domains. The provided fields record all parameters and corresponding field values ​​provided by the user in the current and historical rounds; for example, seat = driver's seat, temperature = 26 degrees Celsius.

[0078] The allowed parameter domains are the ranges of allowed values ​​for each parameter recorded in the parameter pattern of the atomic vehicle capability. For example, the allowed parameter domain for "seat" is {driver's seat, passenger's seat}, and the allowed parameter domain for "temperature" is 16 to 32 degrees Celsius. If the value of a provided field falls within the corresponding allowed parameter domain, the field is included in the set of valid provided fields; if the field value exceeds the allowed parameter domain, the field is not included in the set of valid provided fields but is classified as a field to be corrected. The dialogue agent identifies parameters in the set of required parameters that do not belong to the set of valid provided fields as the missing fields. The missing fields refer to those parameters in the set of required parameters for the unique target capability that have not yet been validly provided. For example, when the set of required parameters for "turn on seat ventilation" is {seat} and the set of valid provided fields is empty, the missing field is {seat}; when the set of required parameters is {temperature, wind speed} and the set of valid provided fields is {temperature = 26 degrees}, the missing field is {wind speed}. By filtering the provided fields as described above, provided fields with invalid values ​​are excluded from the set of provided fields and thus identified as fields to be corrected rather than required parameters that have already been met, ensuring the accuracy of subsequent interactive guidance.

[0079] By employing the aforementioned method for identifying missing fields, the dialogue agent can accurately identify currently missing parameters from the difference between the required parameter set and the set of legally provided fields. This provides precise data for clarifying differentiated interaction guidance in non-atomic states. Specifically, the dialogue agent only needs to guide the user to supplement missing fields and correct fields requiring correction, without needing to inquire about legally provided fields or uniformly query all required parameters, thus improving interaction efficiency. Simultaneously, the filtering of legally provided fields ensures that parameters with invalid values ​​are correctly identified as fields requiring correction rather than being incorrectly considered satisfied. This avoids misjudgments caused by invalid field values ​​being included in the comparison of complete coverage, improving the accuracy of state determination and the effectiveness of subsequent interaction guidance.

[0080] Optionally, generating structured interaction instructions based on the current completeness state and the set of available atomic vehicle control capabilities includes: When the current completeness state is the fuzzy state, the control object and control action corresponding to each candidate capability are extracted from the candidate capability set to generate candidate selection items as the structured interaction instructions. The candidate selection items are subject to the pruning constraint of the available atomic vehicle control capability set. When the current completeness state is the explicitly non-atomic state, the current interaction field is determined from the missing fields, and the allowed parameter field and front-end component type corresponding to the current interaction field are determined according to the available atomic vehicle control capability set. A parameter supplementation instruction containing the current interaction field is generated as the structured interaction instruction.

[0081] Optionally, the progressive convergence method for conversational remote vehicle control commands that integrates vehicle teaching further includes: When the number of candidates in the candidate capability set exceeds a preset display threshold in the ambiguous state, the structured interaction instruction also includes a vehicle control entry instruction, which is used to guide the user to enter the vehicle control interface for selection.

[0082] Specifically, the dialogue agent generates structured interaction instructions based on the current completeness state and the set of available atomic control capabilities. These structured interaction instructions are structured data objects output by the dialogue agent to the user terminal, used to drive the user terminal to render the corresponding interactive interface and guide the user to complete the next input operation. Their specific form varies depending on the current completeness state.

[0083] When the current completeness state is in the fuzzy state, it indicates that there are multiple candidate capabilities in the candidate capability set or the user's intent is unclear. The dialogue agent needs to guide the user to select one from multiple candidates to eliminate ambiguity. At this time, the dialogue agent extracts the control object and control action corresponding to each candidate capability from the candidate capability set, generating candidate options as the structured interaction instruction. The candidate options are a set of option data contained in the structured interaction instruction, allowing the user to select one from multiple candidates. Each option corresponds to a combination description of the control object and control action of an atomic vehicle control capability. For example, when the user says "cool down the car", the dialogue agent matches two candidate capabilities: "turn on the air conditioner" and "open the car window". The candidate options include two options: "turn on the air conditioner" and "open the car window". The user terminal renders these as two clickable option buttons for the user to choose from.

[0084] The candidate options are subject to a pruning constraint based on the available atomic vehicle control capabilities set. That is, each option in the candidate options comes from the subset of available intents and front-end interaction components generated in step S300 by pruning the available atomic vehicle control capabilities set, and only includes options corresponding to the capabilities actually supported by the current vehicle. Therefore, the candidate options presented to the user are all atomic vehicle control capabilities that the current vehicle can execute, and no functional options that the vehicle does not possess will appear.

[0085] When the current completeness state is the explicitly non-atomic state, it indicates that the unique target capability has been locked, but there are still missing or uncorrected fields. The dialogue agent needs to guide the user to supplement or correct specific parameters. At this time, the dialogue agent determines the current interaction field from the missing fields, and determines the allowed parameter domain and front-end component type corresponding to the current interaction field according to the available atomic vehicle control capability set. It then generates a parameter supplementation instruction containing the current interaction field as the structured interaction instruction. The current interaction field is the parameter that needs to be guided to operate in the current round among multiple missing or uncorrected fields, determined according to a preset priority. For example, when "turn on the air conditioner" is missing the two required parameters "temperature" and "fan speed", the dialogue agent first determines "temperature" as the current interaction field according to the preset priority. The allowed parameter domain is the range of values ​​that the parameter is allowed to take. For example, the allowed parameter domain for "temperature" is 16 to 32 degrees Celsius, and the allowed parameter domain for "seat" is either the driver's seat or the passenger seat. The front-end component type is the interactive control form that the parameter should take on the user's end. For example, "temperature" corresponds to a slider component, "seat" corresponds to a radio button component, and "wind speed" corresponds to a stepper component. The dialogue agent determines the allowed parameter domain and front-end component type corresponding to the current interactive field based on the available atomic vehicle control capability set. The allowed parameter domain and front-end component type are derived from the subset of front-end interactive components generated by trimming the available atomic vehicle control capability set in step S300. Therefore, the supplementary parameter controls and their options rendered on the user's end are all parameter ranges and component forms actually supported by the current vehicle. For example, when a user says "turn on seat ventilation" and the car only supports driver and front passenger seats, the supplementary parameter instruction includes a "seat" field, allowing the parameter fields to be driver and front passenger, and the front-end component type to be radio button. Based on this, the client only renders the "driver" and "front passenger" radio buttons for the user to choose from, and does not present the "rear seat" option that the car does not support.

[0086] Preferably, when the number of candidates in the candidate capability set exceeds a preset display threshold in the fuzzy state, the structured interaction instruction further includes a vehicle master control entry instruction. The preset display threshold is a pre-set critical value used to determine whether candidate capabilities are presented in an item-by-item listing manner or as a master control entry manner; for example, it is set to 3. When the number of candidates exceeds 3, the master control entry mode is triggered. The vehicle master control entry instruction guides the user to enter the vehicle master control interface for selection. When there are too many candidate capabilities, listing all candidate options item by item in the interaction interface may lead to interface congestion and reduced user selection efficiency. At this time, the dialogue agent generates a structured interaction instruction containing the vehicle master control entry instruction. The user terminal renders an entry control for "Enter Vehicle Master Control" accordingly. After the user clicks, they enter the complete vehicle control interface, where they can complete the selection and parameter configuration of the target vehicle control capability.

[0087] By generating differentiated interaction commands as described above, the dialogue agent adopts corresponding interaction guidance strategies based on the current completeness state, achieving precise control over "what the user should select and what the user should fill in" during multi-turn dialogues. This avoids the inefficient interaction caused by using a uniform follow-up question template in all incomplete states. In ambiguous states, the agent guides the user to select from actually available candidates; in clear, non-atomic states, it only guides the user to supplement the currently missing single parameter rather than listing all unfilled fields, reducing the user's reading and operational burden and improving the interaction efficiency of multi-turn dialogues. Simultaneously, by setting a preset display threshold to degrade the display of the central control entry when there are too many candidates, the information presentation and operation path of the user interface are further optimized. Both candidate selections and parameter supplementation commands are tailored based on the available atomic vehicle control capability set as a constraint source, ensuring that the candidate options and parameter value ranges presented at the front end are all capabilities and parameters actually supported by the vehicle, avoiding presenting users with functional options or parameter values ​​that the vehicle does not possess.

[0088] Optionally, the progressive convergence method for conversational remote vehicle control commands that integrates vehicle teaching further includes: When the current target capability does not belong to the set of available atomic vehicle control capabilities of the current vehicle, the vehicle manual knowledge base is searched using the controlled object as the search key, a manual operation guide corresponding to the controlled object is generated, and the current dialogue is switched from the vehicle control command mode to the teaching question and answer mode. After switching to the teaching question-and-answer mode and generating the manual operation guidance, if subsequent input from the user is received, the subsequent input is used as the new current input to re-identify the target scenario corresponding to the intent. The target scenario also includes the vehicle teaching scenario. The vehicle teaching scenario includes teaching skills, and the remote vehicle control scenario includes vehicle control skills. The teaching skills and the vehicle control skills are isolated from each other and maintained independently. When the target scenario is the remote vehicle control scenario, the vehicle control skills are loaded and the loading of the teaching skills is prohibited. When the target scenario is the vehicle teaching scenario, the teaching skills are loaded and the loading of the vehicle control skills is prohibited.

[0089] In some implementations, the progressive convergence method for conversational remote vehicle control commands that integrates vehicle instruction also includes a downgraded instruction process when the vehicle control capability is unavailable. When the current target capability does not belong to the set of available atomic vehicle control capabilities for the current vehicle, the dialogue agent triggers a downgrade process. The current target capability is the atomic vehicle control capability that the dialogue agent has locked and is ready to execute in the main vehicle control process. Its absence from the set of available atomic vehicle control capabilities indicates that the control item is unavailable on the current vehicle. For example, if a user requests "open the car door," but the current vehicle does not have a remote door opening function, then the control item "remote door opening" does not belong to the set of available atomic vehicle control capabilities for that vehicle. In this case, the dialogue agent uses the controlled object as the search key to search the vehicle manual knowledge base.

[0090] The controlled object is the target component that the user wants to operate, extracted from the user's request by the dialogue agent. For example, when the user requests "open the car door," the controlled object is "car door." The vehicle manual knowledge base is a pre-built data system that stores vehicle user manual content. The content is indexed with tags based on dimensions such as controlled object, function name, and operation steps, supporting both precise tag-based retrieval and combined retrieval based on text and images. The dialogue agent uses the controlled object as the search key to search the vehicle manual knowledge base, recalling vehicle usage instructional content related to that controlled object and generating manual operation instructions corresponding to that controlled object. The manual operation instructions are structured instructional information generated by the dialogue agent based on the content recalled from the knowledge base, including text descriptions and illustrative instructions. For example, when "open the car door" is unavailable, the dialogue agent generates illustrative instructions on "how to manually unlock the car door." If no valid content is recalled in the vehicle manual knowledge base, the dialogue agent further triggers an online search to obtain supplementary information. Simultaneously with generating the manual operation instructions, the dialogue agent switches the current dialogue from a vehicle control command mode to an instructional question-and-answer mode. The vehicle control command modality refers to the current session's operational state, which revolves around parsing and executing vehicle control commands. The teaching question-and-answer modality refers to the current session's operational state, which revolves around retrieving and presenting teaching knowledge. The vehicle control command modality and the teaching question-and-answer modality are different operational states of the same session; switching between these states does not result in the loss of session history, and the user's dialogue context is fully preserved.

[0091] After switching to the teaching question-and-answer modality and generating the manual operation guidance, if subsequent input from the user is received, the dialogue agent will use this subsequent input as the new current input to re-perform intent recognition for the corresponding target scenario. At this time, the target scenario also includes vehicle teaching scenarios; that is, the intent recognition module is no longer limited to recognizing remote vehicle control intents but also supports recognizing vehicle teaching intents. The subsequent input is the content entered by the user after seeing the manual operation guidance, and its form may be natural language text or front-end interactive data. The dialogue agent will use this subsequent input as the new current input and re-execute intent recognition and scenario routing. For example, if the user first requests "open the car door" (routed to the remote vehicle control scenario), and because the car does not have remote door opening capability, the instruction guide "how to manually unlock the car door" is generated, then the user continues to enter "What about the trunk?" in the same dialogue. The dialogue agent will then use this input as the new current input to re-perform intent recognition, route it to the vehicle teaching scenario, and retrieve relevant teaching content for "trunk".

[0092] The vehicle teaching scenario includes teaching skills, and the remote vehicle control scenario includes vehicle control skills. The teaching skills and vehicle control skills are isolated and maintained independently. The teaching skills are a set of prompts and tools encapsulated within the dialogue agent for the vehicle usage teaching scenario, including teaching-related system prompts, manual knowledge base retrieval tools, and image / text generation tools. The vehicle control skills are a set of prompts and tools encapsulated within the dialogue agent for the vehicle control scenario, including vehicle control-related system prompts, atomic vehicle control capability matching tools, and state transition engine tools. The teaching skills and vehicle control skills are completely isolated in code and runtime environment, supporting independent maintenance. Each has its own independent prompt configuration, toolset, and version iteration branches; modifications or canary releases of one will not affect the other. When the target scenario is the remote vehicle control scenario, the dialogue agent loads the vehicle control skills and disables the loading of the teaching skills; when the target scenario is the vehicle teaching scenario, the dialogue agent loads the teaching skills and disables the loading of the vehicle control skills. The term "loading" refers to loading the system prompts and tool registration information for the corresponding skills into the runtime environment of the dialogue agent, making them effective for the current dialogue; the term "disallowing loading" refers to not loading the system prompts and tools for the corresponding skills, making them ineffective for the current dialogue. Through this on-demand loading and mutually exclusive loading mechanism, the dialogue agent exposes only vehicle-related tools when executing vehicle control logic, avoiding accidental invocation of teaching retrieval tools; and exposes only teaching-related tools when executing teaching logic, avoiding accidental invocation of vehicle control distribution tools.

[0093] Through the aforementioned downgraded teaching and dual-scenario bidirectional switching mechanism, the dialogue agent does not report a fatal error to the user when the target vehicle control capability is lacking. Instead, it smoothly transitions to the teaching question-and-answer modality, reusing existing teaching capabilities to generate manual operation guidance. This ensures that users still receive actionable instructions when remote vehicle control is unavailable, preventing experience interruption. The isolated loading and mutual exclusion loading mechanism between teaching skills and vehicle control skills ensures that the two sets of prompts and tools do not interfere with each other while sharing the same dialogue entry point, reducing the probability of misjudgment of intent and mis-adjustment of tools. Simultaneously, a mapping relationship is established between the vehicle control command modality and the teaching question-and-answer modality for the same controlled object. Downgraded switching when vehicle control capability is lacking uses this controlled object as a bridge, allowing users to obtain targeted teaching guidance without having to repeatedly describe their needs. If a user continues to speak after being downgraded to a teaching session, that speech becomes new input and re-enters the main intent recognition process. This allows teaching and vehicle control to switch and connect repeatedly within the same conversation. After watching the teaching guide on "how to manually open the door," a user can ask "then how do I lock the car?" (continuing the teaching session) or simply say "never mind, open the window for me" (switching back to vehicle control). The two scenarios flow naturally within the same dialogue context, forming a complete closed loop for vehicle services. This avoids the cognitive and operational burden that users have to switch between different entry points or systems due to scenario fragmentation in traditional solutions.

[0094] In some specific embodiments, such as Figure 2 As shown, after obtaining the complete vehicle control command through the aforementioned progressive convergence process, the dialogue agent outputs the command to a separate vehicle control verification and confirmation module. This module performs a final verification of parameter integrity and vehicle status. After successful verification, the vehicle control verification and confirmation module generates a confirmation request and returns it to the user terminal. The user terminal then presents a confirmation interface containing the control action to be executed and its corresponding parameters.

[0095] When the user clicks the confirmation button, the user's client encapsulates the confirmation action and user information (such as the order ID) into a confirmation command and sends it back to the vehicle control verification and confirmation delivery module. The user information is used to uniquely identify the session binding relationship between this vehicle control operation and the current user in concurrent request scenarios. Upon receiving the confirmation command, the vehicle control verification and confirmation delivery module sends the corresponding control command to the target vehicle and generates execution result feedback based on the response returned by the vehicle.

[0096] When the command is successfully issued, the dialogue agent generates a success feedback message and returns it to the user's terminal. The user's terminal displays the success message, such as "The command to turn on the driver's seat ventilation has been issued for you," and this round of vehicle control process ends.

[0097] When a command fails to be issued, the dialogue agent does not directly return a simple error message to the user. Instead, it determines that the current vehicle control command cannot be executed remotely and triggers a degradation process. This degradation process uses the currently controlled object as the search key to retrieve the vehicle manual knowledge base, recalls vehicle usage instructional content related to that controlled object, generates manual operation guidance corresponding to the controlled object, and switches the current dialogue from a vehicle control command mode to an instructional question-and-answer mode. For example, when the "open the car door" command fails to be issued, the dialogue agent uses "car door" as the search key to retrieve the vehicle manual knowledge base, generates a graphic and textual operation guide on "how to manually unlock the car door," and returns it to the user, informing the user how to complete the operation manually. After the user receives this graphic and textual operation guide, they can manually operate according to the guide without facing an unrecoverable error message.

[0098] If, after receiving manual operation instructions, the user does not click on the front-end control but continues speaking in natural language, such as "How do I open the trunk?", the dialogue agent receives this subsequent input and treats it as the new current input for intent recognition. This subsequent input carries user information (e.g., order ID) and system information (e.g., the current dialogue channel), enabling the dialogue agent to identify the user's identity and terminal source within the same conversational context, ensuring correct state binding and channel adaptation during multi-turn interactions. The dialogue agent performs intent recognition on this subsequent input: if the recognition result is a vehicle teaching scenario, it enters the teaching question-and-answer modality, using "trunk" as the search key to retrieve knowledge from the knowledge base and generate corresponding instructions; if the recognition result is a remote vehicle control scenario, it re-enters the vehicle control command progressive convergence process, executing the processing logic from steps S400 to S900, that is, re-executing the complete process from "receiving natural language vehicle control requests, extracting control objects and parameter fields" to "multi-turn interactions until convergence into a complete vehicle control command".

[0099] Through the above process, complete vehicle control commands are only issued after user confirmation following output, avoiding the security risks associated with unconfirmed automatic execution. When remote vehicle control fails, the dialogue agent automatically downgrades to a teaching question-and-answer mode, generating manual operation guidance by retrieving the vehicle manual knowledge base. This transforms a failed vehicle control operation into an effective teaching service, providing users with an immediate and actionable remedial solution, avoiding the negative experience and user confusion caused by simple error reports in traditional solutions. Furthermore, if the user continues to speak after downgrading to the teaching mode, whether asking a question or re-initiating a vehicle control request, the dialogue agent re-identifies the intent and guides them to the correct scenario branch. This creates a complete two-way flow loop between teaching and vehicle control within the same conversation, eliminating the need for users to reopen the conversation or switch entry points, achieving a natural transition between different service modes. The transmission of user and system information ensures the consistency and traceability of dialogue state management across rounds and channels, providing a reliable session binding foundation for long-term, multi-round interactions.

[0100] like Figure 3 As shown in the figure, an embodiment of the present invention provides a progressive convergence system for conversational remote vehicle control commands that integrates vehicle teaching, used to implement the aforementioned progressive convergence method for conversational remote vehicle control commands that integrates vehicle teaching. The system includes: The intent recognition module is used to recognize the user's current input and determine the corresponding target scenario. The capability acquisition module is used to acquire the set of available atomic vehicle control capabilities corresponding to the current vehicle when the target scenario is a remote vehicle control scenario. The capability trimming and loading module is used to generate a subset of available intents, a subset of available tools, and a subset of front-end interaction components from a pre-stored atomic vehicle capability mapping library, using the set of available atomic vehicle capabilities as a constraint source. The module loads the subset of available intents and the subset of available tools into the dialogue agent and sends the subset of front-end interaction components to the user terminal. The semantic parsing and candidate matching module is used to extract the control object, control action and parameter fields from the natural language vehicle control request when receiving the user's natural language vehicle control request, and match them with the available atomic vehicle control capability set to obtain a candidate capability set; The state management and merging module is used to obtain the structured vehicle control descriptor saved in the previous round of the current session, merge the control object, the control action and the parameter field extracted in this round into the structured vehicle control descriptor, and update the candidate capability information in the structured vehicle control descriptor with the candidate capability set; The instruction generation and output module is used to determine the current completeness state based on the updated structured vehicle control descriptor and the set of required parameters for the current target capability; if the current completeness state is a definite atomic state, the updated structured vehicle control descriptor is used as a complete vehicle control instruction; if the current completeness state is an ambiguous state or a definite non-atomic state, a structured interaction instruction is generated according to the current completeness state and the set of available atomic vehicle control capabilities. The structured interaction instruction is used to guide the user to supplement or correct the vehicle control parameters for multi-turn dialogue interaction. The loop convergence module is used to receive subsequent inputs returned by the user terminal according to the structured interaction instructions, update the structured vehicle control descriptor and re-determine the current completeness state until it converges to the complete vehicle control instructions.

[0101] An embodiment of the present invention provides a vehicle, including a memory and a processor; the memory is used to store a computer program; the processor is used to implement, when executing the computer program, the progressive convergence method of conversational remote vehicle control commands for integrated vehicle teaching as described above.

[0102] This invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the dialogic remote vehicle control command progressive convergence method for integrated vehicle teaching as described above.

[0103] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A progressive convergence method for conversational remote vehicle control commands that integrates vehicle teaching, characterized in that, include: Perform intent recognition on the user's current input to determine the corresponding target scenario; When the target scenario is a remote vehicle control scenario, obtain the set of available atomic vehicle control capabilities corresponding to the current vehicle; Using the set of available atomic vehicle capabilities as a constraint source, a subset of available intentions, a subset of available tools, and a subset of front-end interaction components are generated by matching from the pre-stored atomic vehicle capability mapping library. The subset of available intentions and the subset of available tools are loaded into the dialogue agent, and the subset of front-end interaction components is sent to the user terminal. Receive a user's natural language vehicle control request, extract the control object, control action and parameter fields from the natural language vehicle control request, and match them with the available atomic vehicle control capability set to obtain a candidate capability set; Obtain the structured vehicle control descriptor saved in the previous round of the current session, merge the control object, the control action and the parameter field extracted in this round into the structured vehicle control descriptor, and update the candidate capability information in the structured vehicle control descriptor with the candidate capability set; Based on the updated structured vehicle control descriptor and the set of required parameters for the current target capability, determine the current completeness status; If the current completeness state is a defined atomic state, the updated structured vehicle control descriptor will be used as the complete vehicle control instruction; If the current completeness state is an ambiguous state or a clearly non-atomic state, a structured interaction instruction is generated based on the current completeness state and the set of available atomic vehicle control capabilities. The structured interaction instruction is used to guide the user to supplement or correct the vehicle control parameters in order to conduct multi-round dialogue interaction. The system receives subsequent input from the user terminal based on the structured interaction command, updates the structured vehicle control descriptor, and re-determines the current completeness state until it converges to the complete vehicle control command.

2. The method for progressive convergence of dialogic remote vehicle control commands integrating vehicle teaching as described in claim 1, characterized in that, The acquisition of the set of available atomic vehicle control capabilities corresponding to the current vehicle includes: Obtain the vehicle identifier of the current vehicle; Using the vehicle identifier as the query key, query the vehicle capability library to obtain the set of available atomic control vehicle capabilities for the current vehicle and the corresponding capability snapshot identifier; The set of available atomic vehicle control capabilities is used as the sole source of capability constraints. When a change in the capability snapshot identifier is detected, the saved fields in the structured vehicle control descriptor that do not belong to the updated allowed parameter field are deleted, and the candidate capability set and the current completeness status are recalculated.

3. The method for progressive convergence of dialogic remote vehicle control commands integrating vehicle teaching as described in claim 1, characterized in that, The step of merging the controlled object, the controlled action, and the parameter fields extracted in this round into the structured vehicle control descriptor, and updating the candidate capability information in the structured vehicle control descriptor with the candidate capability set, includes: If the control object and control action extracted in this round are consistent with the control object and control action of the target capability in the previous round, the parameter fields extracted in this round will be written into the provided fields of the structured vehicle control descriptor. If the control object or control action extracted in this round is different from the control object or control action of the target capability in the previous round, the provided fields in the structured vehicle control descriptor are cleared, and the target capability information of the structured vehicle control descriptor is updated with the control object and control action extracted in this round to obtain the updated structured vehicle control descriptor.

4. The method for progressive convergence of dialogic remote vehicle control commands integrating vehicle teaching as described in claim 1, characterized in that, The determination of the current completeness status includes: When the number of candidate capabilities in the candidate capability set is not one, or when there is no uniquely determined control object and control action in the candidate capability set, the current completeness state is determined to be the fuzzy state. When the number of candidate capabilities in the candidate capability set is one, there is a unique target capability in the candidate capability set, and there are missing fields in the provided fields of the updated structured vehicle control descriptor relative to the set of required parameters of the unique target capability, or there are fields in the provided fields whose field values ​​do not belong to the corresponding allowed parameter domains, the current completeness state is determined to be the explicit non-atomic state. When the number of candidate capabilities in the candidate capability set is one, the unique target capability exists, the provided fields completely cover the set of required parameters, and the field values ​​of all the provided fields belong to the corresponding allowed parameter domains, the current completeness state is determined to be the explicit atomic state.

5. The method for progressive convergence of dialogic remote vehicle control commands integrating vehicle teaching according to claim 4, characterized in that, The methods for determining missing fields in the required parameter set of the unique target capability include: Obtain the set of required parameters for the unique target capability; Obtain the set of legal provided fields in the updated structured vehicle descriptor. The legal provided fields consist of provided fields whose field values ​​belong to the corresponding allowed parameter fields. The parameters in the set of required parameters that do not belong to the set of legally provided fields are designated as the missing fields.

6. The method for progressive convergence of dialogic remote vehicle control commands integrating vehicle teaching as described in claim 5, characterized in that, The step of generating structured interaction instructions based on the current completeness state and the set of available atomic vehicle control capabilities includes: When the current completeness state is the fuzzy state, the control object and control action corresponding to each candidate capability are extracted from the candidate capability set to generate candidate selection items as the structured interaction instructions. The candidate selection items are subject to the pruning constraint of the available atomic vehicle control capability set. When the current completeness state is the explicitly non-atomic state, the current interaction field is determined from the missing fields, and the allowed parameter field and front-end component type corresponding to the current interaction field are determined according to the available atomic vehicle control capability set. A parameter supplementation instruction containing the current interaction field is generated as the structured interaction instruction.

7. The method for progressive convergence of dialogic remote vehicle control commands integrating vehicle teaching as described in claim 6, characterized in that, The progressive convergence method for conversational remote vehicle control commands that integrates vehicle teaching also includes: When the number of candidates in the candidate capability set exceeds a preset display threshold in the ambiguous state, the structured interaction instruction also includes a vehicle control entry instruction, which is used to guide the user to enter the vehicle control interface for selection.

8. The method for progressive convergence of dialogic remote vehicle control commands integrating vehicle teaching according to claim 1, characterized in that, The structured vehicle control descriptor also includes the field source and confirmation round for each field. The field source is used to indicate that the field originates from the current user input, historical user input, or the user's selection of interactive components in the structured interaction command.

9. The method for progressive convergence of dialogic remote vehicle control commands integrating vehicle teaching as described in claim 1, characterized in that, The progressive convergence method for conversational remote vehicle control commands that integrates vehicle teaching also includes: When the current target capability does not belong to the set of available atomic vehicle control capabilities of the current vehicle, the vehicle manual knowledge base is searched using the controlled object as the search key, a manual operation guide corresponding to the controlled object is generated, and the current dialogue is switched from the vehicle control command mode to the teaching question and answer mode. After switching to the teaching question-and-answer mode and generating the manual operation guidance, if subsequent input from the user is received, the subsequent input is used as the new current input to re-identify the target scenario corresponding to the intent. The target scenario also includes the vehicle teaching scenario. The vehicle teaching scenario includes teaching skills, and the remote vehicle control scenario includes vehicle control skills. The teaching skills and the vehicle control skills are isolated from each other and maintained independently. When the target scenario is the remote vehicle control scenario, the vehicle control skills are loaded and the loading of the teaching skills is prohibited. When the target scenario is the vehicle teaching scenario, the teaching skills are loaded and the loading of the vehicle control skills is prohibited.

10. A conversational remote vehicle control command progressive convergence system integrating vehicle teaching, characterized in that, The system is used to implement the conversational remote vehicle control command progressive convergence method for integrated vehicle teaching as described in any one of claims 1 to 9, the system comprising: The intent recognition module is used to recognize the user's current input and determine the corresponding target scenario. The capability acquisition module is used to acquire the set of available atomic vehicle control capabilities corresponding to the current vehicle when the target scenario is a remote vehicle control scenario. The capability trimming and loading module is used to generate a subset of available intents, a subset of available tools, and a subset of front-end interaction components from a pre-stored atomic vehicle capability mapping library, using the set of available atomic vehicle capabilities as a constraint source. The module loads the subset of available intents and the subset of available tools into the dialogue agent and sends the subset of front-end interaction components to the user terminal. The semantic parsing and candidate matching module is used to extract the control object, control action and parameter fields from the natural language vehicle control request when receiving the user's natural language vehicle control request, and match them with the available atomic vehicle control capability set to obtain a candidate capability set; The state management and merging module is used to obtain the structured vehicle control descriptor saved in the previous round of the current session, merge the control object, the control action and the parameter field extracted in this round into the structured vehicle control descriptor, and update the candidate capability information in the structured vehicle control descriptor with the candidate capability set; The instruction generation and output module is used to determine the current completeness state based on the updated structured vehicle control descriptor and the set of required parameters for the current target capability; if the current completeness state is a definite atomic state, the updated structured vehicle control descriptor is used as a complete vehicle control instruction; if the current completeness state is an ambiguous state or a definite non-atomic state, a structured interaction instruction is generated according to the current completeness state and the set of available atomic vehicle control capabilities. The structured interaction instruction is used to guide the user to supplement or correct the vehicle control parameters for multi-turn dialogue interaction. The loop convergence module is used to receive subsequent inputs returned by the user terminal according to the structured interaction instructions, update the structured vehicle control descriptor and re-determine the current completeness state until it converges to the complete vehicle control instructions.

Images