Three-dimensional computer-aided design model generation method based on multi-modal agent

Abstract

The application provides a three-dimensional computer-aided design model generation method based on a multi-modal intelligent agent, relates to the field of computer-aided design, and aims to solve the problems that in the prior art, multi-source design information is difficult to be stably converted into continuous and orderly parameterized modeling steps, and long-process generation lacks process explicit recording, resulting in inconsistent multi-round correction and unstable generation. The method constructs a main agent, a modeling planning agent and a plurality of modeling skill agents around the difference between the current model and the target geometric reference, and establishes structured state data for recording the model generation process; the main agent identifies the difference information between the current model and the target geometric reference and writes it into the structured state data, the modeling planning agent dynamically generates or corrects the modeling plan accordingly, the main agent schedules the modeling skill agent to execute parameterized feature generation, and continuously analyzes and updates the structured state data and the subsequent modeling plan according to the execution result until the preset generation condition is met.

Description

Technical Field

[0001] This application relates to the field of computer-aided design, and in particular to a method for generating three-dimensional computer-aided design models based on multimodal intelligent agents. Background Technology

[0002] The automated generation of 3D computer-aided design (CAD) models has wide applications in fields such as industrial design, intelligent manufacturing, and architectural modeling.

[0003] In existing technologies, during the mechanical manufacturing process, engineers need to quickly generate editable parametric CAD models based on design drawings or physical object scan data for subsequent simulation analysis, processing and manufacturing, or 3D (Three-Dimensional) printing. In the construction industry, designers need to generate building structure models that conform to specifications based on multi-view rendered images or user-provided sketches.

[0004] However, existing solutions often directly output meshes or geometric shells, which are difficult to reliably transform into an executable parametric modeling process. In multi-round interactions or long-process generation, they often rely on context to save intermediate information, which can easily lead to inconsistent geometric understanding, resulting in limited stability, editability, and engineering applicability of 3D CAD model generation. Summary of the Invention

[0005] This application provides a method for generating 3D computer-aided design models based on multimodal intelligent agents to solve the aforementioned technical problems. This method addresses the scenario of generating 3D computer-aided design models from multi-source design information. Through multi-agent collaborative decision-making and process control, it unifies the organization of target understanding, difference identification, task planning, generation execution, and result correction during the model generation process. This enables 3D model generation to move beyond simply outputting geometric shapes and instead continuously iterate towards a parametric modeling process, thereby improving the consistency, stability, editability, and engineering applicability of the model in long-process generation.

[0006] In a first aspect, this application provides a method for generating a three-dimensional computer-aided design model based on a multimodal intelligent agent, the method comprising:

[0007] Construct a main intelligent agent, a modeling and planning intelligent agent, and multiple modeling skill intelligent agents, and initialize structured state data for recording information about the model generation process;

[0008] The main intelligent agent obtains the current model, target geometric reference, and modeling skill set.

[0009] The main intelligent agent determines the difference information based on the current model and the target geometric reference, and writes the difference information into the structured state data.

[0010] By modeling and planning intelligent agents, a modeling plan containing ordered steps can be dynamically generated or modified based on the differences in structured state data.

[0011] The main agent schedules the corresponding modeling skill agents to execute the parameterized feature generation process according to the modeling plan and the set of modeling skills, and obtains the current execution result.

[0012] The main intelligent agent analyzes the current execution result to obtain the analysis result;

[0013] When the analysis results do not meet the preset generation conditions, the structured state data is updated according to the analysis results and the subsequent modeling plan is corrected. Then, the process is returned to the main agent to schedule the corresponding modeling skill agent, and the parameterized feature generation process is executed according to the modeling plan and the modeling skill set to obtain the current execution result.

[0014] The generation process ends when the analysis results meet the preset generation conditions.

[0015] In one possible implementation, the current model, the target geometric reference, and the modeling skills set are obtained, including:

[0016] Obtain the initial CAD model of the target 3D object to be processed as the current model;

[0017] Obtain multi-view renderings, 2D sketches, or text descriptions of the target 3D object as geometric references for the target;

[0018] Obtain a pre-set parametric modeling instruction library and extract multiple callable atomic modeling operations contained in the parametric modeling instruction library. Map the atomic modeling operations to a set of modeling skills. The set of modeling skills encapsulates the corresponding tool call interface and geometric verification rules.

[0019] In one possible implementation, a modeling plan containing ordered steps is dynamically generated or modified based on the difference information in the structured state data, including:

[0020] Analyze the discrepancies in structured state data to identify missing geometric features or existing geometric deviations in the current model;

[0021] Based on a pre-defined planning strategy, geometric features or geometric deviations are transformed into a sequence of modeling tasks containing multiple ordered steps.

[0022] Match the corresponding modeling skills to each step in the modeling task sequence, and generate a modeling plan that includes operation type, target parameters, and modeling skill dependencies.

[0023] In one possible implementation, based on the modeling plan and the set of modeling skills, a parametric feature generation process is executed to obtain the current execution result, including:

[0024] In accordance with the order of steps in the modeling plan, call the corresponding modeling skills in the modeling skill set one by one;

[0025] During the invocation process, the execution parameters required for each modeling skill are calculated based on the real-time geometric state of the current model. The calculation of the execution parameters includes: determining the sketch plane and normal vector, identifying the two-dimensional contour region, determining the stretching direction and termination surface, and performing geometric constraint verification.

[0026] Parametric features are generated based on execution parameters to update the topology of the current model, and the updated current model is used as the current execution result.

[0027] In one possible implementation, the current execution result is analyzed to obtain the analysis result, including:

[0028] Obtain the geometric feature data corresponding to the current execution result;

[0029] The geometric feature data is compared with the target geometric reference to calculate the geometric overlap or feature missing rate, and the current generation status of the model is determined based on the geometric overlap or feature missing rate.

[0030] When the geometric overlap is below a preset threshold and / or the feature missing rate indicates the presence of missing features, generate analysis results that include correction suggestions;

[0031] When the geometric overlap is higher than or equal to a preset threshold and the feature missing rate indicates no feature missing, the difference information in the structured state data is cleared, and an analysis result containing successful confirmation is generated.

[0032] In one possible implementation, the modeling skills set includes at least one of the following:

[0033] Sketching skills are used to generate two-dimensional outlines on a specified plane;

[0034] Feature manipulation skills are used to generate 3D solids based on 2D contours or to perform Boolean operations;

[0035] Transformation manipulation skills are used to perform spatial translation, rotation, or mirror symmetry processing on model features;

[0036] Chamfering and rounding techniques are used to smooth the edges of a model.

[0037] In one possible implementation, the preset generation conditions include:

[0038] The difference information in the structured state data is empty, and the matching degree between the current execution result and the target geometric reference reaches the preset standard;

[0039] Alternatively, the number of iterations of the modeling plan reaches a preset iteration threshold.

[0040] Secondly, this application provides a three-dimensional computer-aided design model generation device based on a multimodal intelligent agent, the device comprising:

[0041] The building module is used to build the main intelligent agent, the modeling and planning intelligent agent, and multiple modeling skill intelligent agents, and initialize the structured state data used to record information about the model generation process;

[0042] The acquisition module is used to acquire the current model, target geometric reference, and modeling skill set through the main intelligent agent;

[0043] The determination module is used by the main intelligent agent to determine the difference information based on the current model and the target geometric reference, and write the difference information into the structured state data;

[0044] The modeling module is used to plan the agent through modeling, and dynamically generate or modify the modeling plan containing ordered steps based on the difference information in the structured state data.

[0045] The execution module is used to schedule the corresponding modeling skill agent through the main agent, and execute the parameterized feature generation process according to the modeling plan and the modeling skill set to obtain the current execution result;

[0046] The analysis module is used to analyze the current execution results through the main intelligent agent and obtain the analysis results;

[0047] The update module is used to update the structured state data and correct the subsequent modeling plan when the analysis results do not meet the preset generation conditions. It then returns the steps of scheduling the corresponding modeling skill agent through the main agent, executing the parameterized feature generation process according to the modeling plan and the modeling skill set, and obtaining the current execution result.

[0048] The completion module is used to end the generation process when the analysis results meet the preset generation conditions.

[0049] In one possible implementation, the acquisition module is specifically used for:

[0050] Obtain the initial CAD model of the target 3D object to be processed as the current model;

[0051] Obtain multi-view renderings, 2D sketches, or text descriptions of the target 3D object as geometric references for the target;

[0052] Obtain a pre-set parametric modeling instruction library and extract multiple callable atomic modeling operations contained in the parametric modeling instruction library. Map the atomic modeling operations to a set of modeling skills. The set of modeling skills encapsulates the corresponding tool call interface and geometric verification rules.

[0053] In one possible implementation, the modeling module is specifically used for:

[0054] Analyze the discrepancies in structured state data to identify missing geometric features or existing geometric deviations in the current model;

[0055] Based on a pre-defined planning strategy, geometric features or geometric deviations are transformed into a sequence of modeling tasks containing multiple ordered steps.

[0056] Match the corresponding modeling skills to each step in the modeling task sequence, and generate a modeling plan that includes operation type, target parameters, and modeling skill dependencies.

[0057] In one possible implementation, the execution module is specifically used for:

[0058] In accordance with the order of steps in the modeling plan, call the corresponding modeling skills in the modeling skill set one by one;

[0059] During the invocation process, the execution parameters required for each modeling skill are calculated based on the real-time geometric state of the current model. The calculation of the execution parameters includes: determining the sketch plane and normal vector, identifying the two-dimensional contour region, determining the stretching direction and termination surface, and performing geometric constraint verification.

[0060] Parametric features are generated based on execution parameters to update the topology of the current model, and the updated current model is used as the current execution result.

[0061] In one possible implementation, the analysis module is specifically used for:

[0062] Obtain the geometric feature data corresponding to the current execution result;

[0063] The geometric feature data is compared with the target geometric reference to calculate the geometric overlap or feature missing rate, and the current generation status of the model is determined based on the geometric overlap or feature missing rate.

[0064] When the geometric overlap is below a preset threshold and / or the feature missing rate indicates the presence of missing features, generate analysis results that include correction suggestions;

[0065] When the geometric overlap is higher than or equal to a preset threshold and the feature missing rate indicates no feature missing, the difference information in the structured state data is cleared, and an analysis result containing successful confirmation is generated.

[0066] In one possible implementation, the set of modeling skills acquired by the acquisition module includes at least one of the following:

[0067] Sketching skills are used to generate two-dimensional outlines on a specified plane;

[0068] Feature manipulation skills are used to generate 3D solids based on 2D contours or to perform Boolean operations;

[0069] Transformation manipulation skills are used to perform spatial translation, rotation, or mirror symmetry processing on model features;

[0070] Chamfering and rounding techniques are used to smooth the edges of a model.

[0071] In one possible implementation, the preset generation conditions in the update module and the completion module include:

[0072] The difference information in the structured state data is empty, and the matching degree between the current execution result and the target geometric reference reaches the preset standard;

[0073] Alternatively, the number of iterations of the modeling plan reaches a preset iteration threshold.

[0074] Thirdly, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor.

[0075] The memory stores the instructions that the computer executes.

[0076] The processor executes computer execution instructions stored in memory to implement the method for generating a three-dimensional computer-aided design model based on a multimodal intelligent agent, as described in the first aspect of the invention.

[0077] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement a method for generating a three-dimensional computer-aided design model based on a multimodal intelligent agent, as described in the first aspect of the invention.

[0078] Fifthly, this application provides a computer program product, including a computer program, which, when executed by a processor, is used to implement the method for generating a three-dimensional computer-aided design model based on a multimodal intelligent agent, as described in the first aspect of the invention.

[0079] Based on the implementation methods provided in the above aspects, this application can be further combined to provide more implementation methods.

[0080] This application provides a method for generating 3D computer-aided design models based on multimodal intelligent agents. By constructing a main intelligent agent, a modeling planning intelligent agent, and multiple modeling skill intelligent agents, and combining structured state data to uniformly record, dynamically plan, and iteratively correct the current model, target geometric reference, difference information, modeling plan, and execution analysis results, it can gradually transform multi-source design information into an ordered parametric feature generation process. During the generation process, it continuously adjusts the subsequent modeling plan based on the analysis results, thereby improving the stability, consistency, and editability of the generated 3D computer-aided design model and enhancing the engineering application capability of the generated results. Attached Figure Description

[0081] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0082] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0083] Figure 1 A flowchart illustrating a method for generating a three-dimensional computer-aided design model based on a multimodal intelligent agent, provided in an embodiment of this application;

[0084] Figure 2 A schematic diagram of a three-dimensional computer-aided design model generation device based on a multimodal intelligent agent provided in this application embodiment;

[0085] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0086] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0087] In the embodiments of this application, the terms "first" and "second" are used to distinguish identical or similar items with substantially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that "first" and "second" do not necessarily imply difference. It should be noted that in the embodiments of this application, the words "exemplary" or "for example" are used to indicate that something is being used as an example, illustration, or description. Any embodiment or design scheme described as "exemplary" or "for example" in this application should not be construed as being better or more advantageous than other embodiments or design schemes. Specifically, the use of "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner. In the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more.

[0088] It should be noted that the phrase "at...time" in the embodiments of this application can refer to the instant at which a certain situation occurs, or to a period of time after the occurrence of a certain situation; the embodiments of this application do not specifically limit this. Furthermore, the method for generating a 3D computer-aided design model based on a multimodal intelligent agent provided in the embodiments of this application is merely an example; a method for generating a 3D computer-aided design model based on a multimodal intelligent agent may also include more or less content.

[0089] 3D computer-aided design model generation technology is widely used in industrial design, intelligent manufacturing, and architectural modeling. In these scenarios, users typically propose modeling requirements based on multi-view images, sketches, text descriptions, or existing geometric models, and expect the system to output 3D CAD models that can be further edited, modified, and manufactured.

[0090] Related applications typically need to handle the collaborative relationship between the current model, the target geometric reference, and the modeling operation capabilities. They need to identify the differences between the target state and the current state, and transform the identification results into subsequent modeling actions. Therefore, they have high requirements for model understanding ability, process continuity, and the engineering usability of the generated results.

[0091] Existing solutions mostly employ image understanding models, text parsing models, or geometry generation models to directly generate 3D meshes and geometric shells based on input multi-view images, sketches, or text descriptions, or further output several CAD command sequences. The basic idea is usually to first extract shape features from the input information, and then construct the target geometric model based on the inference results. Some solutions will combine existing models for difference comparison during the generation process to gradually approximate the target structure.

[0092] However, the above-mentioned solutions still have significant shortcomings in practical applications. On the one hand, multi-source design information is often scattered across different perspectives and modalities, making it difficult for existing methods to consistently and reliably transform this discrepancy into ordered parametric modeling steps. This can easily lead to inaccurate judgments regarding sketch planes, feature orientations, or the order of operations. On the other hand, long-process generation typically relies on contextual storage of intermediate results, lacking explicit recording of information about the model generation process. This results in inconsistent geometric understanding during multiple rounds of revisions, leading to repeated changes in the plan, the spread of local errors, and even generation interruptions.

[0093] The aforementioned shortcomings further lead to a situation where, although the generated results may superficially resemble the target shape, it is difficult to form a stable, continuous, and correctable CAD modeling process, thus limiting its applicability in engineering modifications, manufacturing preparation, and iterative complex tasks. Maintaining consistency between difference identification, modeling plan correction, and feature generation during multiple rounds of analysis and execution has become a key issue that this type of technology needs to address.

[0094] Based on this, this application proposes a method for generating 3D computer-aided design models based on multimodal intelligent agents, applicable to the field of computer-aided design, aiming to solve the aforementioned technical problems of existing technologies. This method focuses on the differences between the current model and the target geometric reference. First, it constructs a master intelligent agent, a modeling planning intelligent agent, and multiple modeling skill intelligent agents, and initializes structured state data for recording information about the model generation process. Then, the master intelligent agent determines the difference information, the modeling planning intelligent agent dynamically generates or modifies a modeling plan containing ordered steps based on the structured state data, and the master intelligent agent schedules the corresponding modeling skill intelligent agents to execute the parametric feature generation process. During the generation process, the master intelligent agent can also analyze the current execution results; when the analysis results do not meet the preset generation conditions, it updates the structured state data based on the analysis results and modifies the subsequent modeling plan, continuing into the subsequent parametric feature generation process; when the analysis results meet the preset generation conditions, the generation ends. Through the above technical approach, the stability of the conversion process from multi-source design information to parametric CAD model generation can be improved, and the consistency and controllability in long-process modeling can be enhanced.

[0095] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0096] Figure 1 This is a flowchart illustrating a method for generating a 3D computer-aided design model based on a multimodal intelligent agent, as provided in an embodiment of this application. Figure 1 As shown, the method includes:

[0097] S101. Construct the main intelligent agent, the modeling planning intelligent agent, and multiple modeling skill intelligent agents, and initialize the structured state data used to record information about the model generation process.

[0098] In this embodiment, the master agent serves as the main control and execution unit for the entire 3D CAD model generation process, undertaking tasks such as task access, state maintenance, skill scheduling, result analysis, and iteration termination control. The modeling and planning agent, as the reasoning unit oriented towards the planning layer, is responsible for transforming discrepancies into executable, ordered modeling steps.

[0099] Multiple modeling skill agents serve as execution layer units, each corresponding to different parametric modeling capabilities. Structured state data is used to record information about the model generation process, supporting the continuous unfolding of subsequent difference identification, plan generation, and skill execution on the same process record.

[0100] In practice, the executing entity can be a modeling and control system deployed on a server, or a local processing module integrated into a computer-aided design platform plugin. The main agent can be instantiated through a unified task orchestration interface, the modeling and planning agent can be implemented using a planning module, and multiple modeling skill agents can each encapsulate calls to the underlying computer-aided design kernel.

[0101] During initialization, the main agent can create task sessions and establish structured state data objects to record relevant state information during task execution. Based on the above processing, this step establishes stable control relationships between agents at different layers and establishes a unified state baseline by initializing structured state data, enabling subsequent difference identification, modeling plan generation, and modeling skill execution to proceed continuously on the same process record.

[0102] S102. Obtain the current model, target geometric reference, and modeling skill set through the main intelligent agent.

[0103] In this embodiment, after the main agent obtains the current model, the target geometric reference, and the modeling skill set, it can uniformly attach the above-mentioned information to the same task context, so that the difference analysis and subsequent planning are based on the same version of input information.

[0104] In practice, the main agent can obtain the current model, target geometric reference, and modeling skill set from the corresponding data source and organize them uniformly for subsequent difference analysis, modeling planning, and skill scheduling. The current model refers to the current geometric model, and the target geometric reference can take forms including, but not limited to, multi-view renderings of the target CAD model, user-provided images, sketches, or text descriptions. Specifically, the user can input information such as the current CAD model, target geometric reference, initial viewpoint, allowed viewpoints, and allowed modeling skill set.

[0105] After acquiring the aforementioned information, the lead agent can write relevant information about the current model, target geometric reference, and modeling skill set into structured state data. Based on this step, the lead agent can unify input-side information and execution-side capabilities into the same data view, eliminating the need to repeatedly parse the source of input information when determining subsequent differences.

[0106] S103. The main intelligent agent determines the difference information based on the current model and the target geometric reference, and writes the difference information into the structured state data.

[0107] In this embodiment, the difference information is used to describe the differences between the current model and the target geometric reference. In this step, the main agent does not directly output the final model, but instead first expresses the differences between the target geometric reference and the current model as data items that can be processed subsequently.

[0108] In practice, the main agent can compare the current model with the target geometric reference to determine the differences between them and generate corresponding difference information. When writing structured state data, the main agent can store the difference information in the corresponding fields according to the current iteration round and establish a connection with the current model version. This step transforms the target shape requirement into a computable and traceable difference item, allowing subsequent modeling and planning agents to directly generate a modeling plan based on the difference information, thus maintaining semantic consistency across multiple processing rounds.

[0109] S104. By modeling and planning the intelligent agent, a modeling plan containing ordered steps is dynamically generated or modified based on the difference information in the structured state data.

[0110] In this embodiment, the modeling plan is an execution scheme generated based on parameterized features, consisting of multiple steps with sequential relationships. The modeling planning agent can read difference information, current model-related information, and execution status from structured state data, transform the differences into a sequence of modeling tasks with sequential relationships, and correct any unexecuted or incomplete steps when subsequent analysis results change. Dynamically generating or correcting the modeling plan specifically includes adding steps, deleting failed steps, adjusting step content, and reordering steps.

[0111] In practical implementation, the modeling and planning agent can generate candidate steps based on the discrepancies in the structured state data, and combine this with the current model state and modeling skill set to produce a modeling plan. The modeling plan can be organized according to the sequential relationships between steps, and subsequent steps can be locally modified when analysis results indicate that some steps require adjustment. When the modeling plan is written back to the structured state data, the corresponding plan information can be recorded. Based on the above processing, it can be seen that this step organizes the discrepancies into an executable parameterized modeling sequence, which can support stable corrections in subsequent iterations.

[0112] S105. The main agent schedules the corresponding modeling skill agent to execute the parameterized feature generation process according to the modeling plan and the modeling skill set, and obtains the current execution result.

[0113] In this embodiment, the main agent can retrieve a matching modeling skill agent from the modeling skill set based on the current step to be executed in the modeling plan, and issue an execution request to that agent. The modeling skill agent can execute a parameterized feature generation process according to the step requirements in the modeling plan and the current model state, and generate an updated model as the current execution result. The current execution result may also include state information related to the execution of this step.

[0114] In practice, the main agent can read the current plan and current step information from the structured state data and select the corresponding modeling skill agent for scheduling. The modeling skill agent can execute corresponding modeling operations based on the current model state and planned steps, and return the execution results. During execution, the main agent can synchronously record the step execution status and write relevant information into the structured state data. For successful execution, the main agent can set the updated model as the new current model version and advance the planned steps. For failed execution, the main agent can retain the corresponding execution status information for subsequent analysis and modeling plan correction. This step completes the transformation from abstract planning to concrete computer-aided design operations, enabling multimodal inference results to fall into an editable and modifiable parametric model history.

[0115] S106. The main intelligent agent analyzes the current execution result and obtains the analysis result.

[0116] In this embodiment, the analysis results are used to characterize whether the current execution result meets the preset generation conditions, and if not, what content still needs to be corrected. After each round of modeling skill execution, the main agent can compare the updated current model with the target geometric reference, and determine whether the current plan is progressing as expected based on the execution status. The analysis results may include at least a satisfaction flag, content to be corrected, and a basis for subsequent processing.

[0117] In practice, the main agent can read the current execution result and call the analysis module to check the result, determining whether there is still a difference between the current execution result and the target geometric reference, and whether the current execution state meets the conditions for continuing execution or ending the generation process. After the analysis result is generated, it can be written into the analysis field in the structured state data. Through this step, the next action can be determined based on the explicit result after each feature generation, rather than based on a simple execution success signal, thus providing a basis for subsequent corrections.

[0118] S107. When the analysis results do not meet the preset generation conditions, update the structured state data according to the analysis results and correct the subsequent modeling plan. Then, return to the main agent to schedule the corresponding modeling skill agent, and execute the parameterized feature generation process according to the modeling plan and the modeling skill set to obtain the current execution result.

[0119] In this embodiment, when the analysis results indicate that the current model still differs from the target geometric reference, or the current execution step fails to achieve the expected generation effect, the main agent may not terminate the task. Instead, it may continue to write the remaining differences, execution status, and affected information obtained from the analysis back to the structured state data, and trigger the modeling planning agent to revise the subsequent modeling plan. After the revision is completed, the main agent re-enters the modeling skill scheduling and parameterized feature generation process, forming a closed loop of analysis, update, revision, and re-execution.

[0120] In practice, the main agent can update the current model version, difference information, and step status in the structured state data based on the analysis results. Subsequently, the modeling planning agent reads the updated structured state data, performs a local replanning of the subsequent modeling plan, and generates a revised modeling plan. The main agent can then reschedule the corresponding modeling skill agent to continue the parametric feature generation process based on the revised modeling plan, generating a new current execution result and continuing into the analysis step. By explicitly updating the structured state data and replanning the modeling plan when preset generation conditions are not met, the geometric understanding before and after the 3D CAD model generation process can continuously converge around the same state record, preventing repeated drift of the modeling plan due to loss of intermediate context.

[0121] In some embodiments, the core iterative process can be represented by the following formula:

[0122]

[0123] Among them, M t+1 Let C represent the structured state data corresponding to the (t+1)th execution of the parameterized feature generation process, and M represent the data preparation function. t For the structured state data corresponding to the t-th execution of the parameterized feature generation process, Ot R represents the results of this round of observations, specifically the execution results after the t-th execution of the parameterized feature generation process. t The reflection result refers to the analysis result corresponding to the t-th execution of the parameterized feature generation process. t+1 Let S be the current generated state of the model after the (t+1)th execution of the parameterized feature generation process, where T is the state transition function and S is the parameterized feature generation function. t Let A be the current generation state of the model after the t-th execution of the parameterized feature generation process. t For the modeling operation corresponding to the t-th execution of the parametric feature generation process, Y t This represents the execution result for the t-th execution of the modeling operation.

[0124] S108. When the analysis results meet the preset generation conditions, the generation process ends.

[0125] In this embodiment, when the main agent confirms that the current execution result meets the preset generation conditions based on the analysis results, it can stop the generation and execution of subsequent modeling plans and output the final three-dimensional computer-aided design model. Upon completion of the generation process, the main agent can also perform finishing processing on the structured state data to ensure that the final output model remains consistent with the process record.

[0126] In practice, after confirming that the analysis results meet the preset generation conditions, the main agent can update the task status to "complete" and use the current model as the output model. Structured state data can retain relevant records from the task execution process at the end, facilitating subsequent editing, backtracking of generation logic, or incremental modifications. Based on the above process, it is clear that the generation process does not end upon successful execution of a single step. The generation process only terminates when the main agent, based on the analysis results, confirms that the target geometric requirements have been accurately implemented into the parametric CAD model through continuous modeling planning and skills. This mechanism ensures consistency between the output results and the target shape while preserving an editable modeling history for easy modification in subsequent engineering projects.

[0127] Based on the above analysis, this application provides a method for generating 3D computer-aided design models based on multimodal intelligent agents. This method includes constructing a main intelligent agent, a modeling planning intelligent agent, and multiple modeling skill intelligent agents, and initializing structured state data for recording information about the model generation process. The main intelligent agent acquires the current model, the target geometric reference, and the modeling skill set. The main intelligent agent determines the difference information based on the current model and the target geometric reference and writes this difference information into the structured state data. The modeling planning intelligent agent dynamically generates or modifies a modeling plan containing ordered steps based on the difference information in the structured state data. The main intelligent agent schedules the corresponding modeling skill intelligent agents to execute a parametric feature generation process according to the modeling plan and the modeling skill set, obtaining the current execution result. The main intelligent agent analyzes the current execution result to obtain an analysis result. If the analysis result does not meet the preset generation conditions, the structured state data is updated based on the analysis result, the subsequent modeling plan is modified, and the process returns to the step of executing the parametric feature generation process. When the analysis result meets the preset generation conditions, the generation process ends. In this embodiment, through hierarchical collaboration among the main agent, the modeling planning agent, and the modeling skill agent, the target requirements are first transformed into structured difference information, then into a parametric modeling plan with a sequential order. After each round of execution, the analysis results are used to write back the structured state data, driving subsequent modeling plan corrections and modeling skill rescheduling. The generation process continuously advances around the same explicit state. Based on this implementation, multi-source design information is not limited to a one-time understanding result but is continuously mapped to a traceable modeling plan and feature history, giving the final output 3D computer-aided design model a foundation for continuous parametric editing.

[0128] In one possible implementation, the process involves acquiring the current model, the target geometric reference, and a set of modeling skills. This includes: acquiring the initial CAD model of the target 3D object as the current model; acquiring multi-view renderings, 2D sketches, or text descriptions of the target 3D object as the target geometric reference; acquiring a pre-built parametric modeling instruction library and extracting multiple callable atomic modeling operations from the library; mapping these atomic modeling operations to a set of modeling skills, which encapsulates the corresponding tool call interfaces and geometric verification rules.

[0129] Specifically, after acquiring the parametric modeling instruction library, interfaces for each atomic modeling operation can be encapsulated, allowing each modeling skill to correspond to a callable tool entry point, bound to geometric verification rules. The tool call interface can point to the feature creation function, sketch editing function, or constraint solving function of the computer-aided design kernel. Geometric verification rules are used to check entity closure, topological consistency, parameter validity, and dependency integrity after feature generation. The modeling skill set is stored in association with the current model and target geometric references for direct retrieval and invocation during modeling plan generation.

[0130] During the process, the initial CAD model can be used as the input for the current model. Then, multi-view renderings, 2D sketches, or text descriptions can be used as the target geometric reference input, and the modeling skill set can be loaded simultaneously. This ensures that subsequent difference identification is based on a unified model state and skill capabilities. Subsequently, the main agent can extract missing features, dimensional deviations, and orientation differences based on the current model and the target geometric reference. Then, it can combine this with the tool call interface of the encapsulated modeling skill set to complete parametric feature generation. Since atomic modeling operations are mapped to skills with verification rules, the execution results can be directly fed back into the current model state and used as the basis for subsequent corrections.

[0131] After adopting the above method, a direct correspondence is formed between the current model, the target geometric reference and the modeling skill set. Atomic-level modeling capabilities are uniformly encapsulated into executable skills, thereby making the connection between model input, target understanding and feature generation more stable, and the generated CAD model structure can maintain good editability and geometric consistency.

[0132] In one possible implementation, a modeling plan containing ordered steps is dynamically generated or modified based on the discrepancies in the structured state data. This includes: parsing the discrepancies in the structured state data to identify missing geometric features or existing geometric deviations in the current model; transforming the geometric features or deviations into a sequence of modeling tasks containing multiple ordered steps based on a preset planning strategy; matching each step in the modeling task sequence with a corresponding modeling skill to generate a modeling plan that includes operation type, target parameters, and modeling skill dependencies.

[0133] In its implementation, the modeling and planning agent can read the difference information field from the structured state data, first normalize the difference content by type, and then determine the attachment surface, generation direction, and size range of the missing features, or determine the correction object and correction magnitude of the deviation features, based on the current model's topological context. Subsequently, based on a preset planning strategy, the features to be completed are decomposed into multiple consecutive modeling actions, forming a modeling task sequence in the order of sketch creation, feature generation, and local refinement. For each task, a skill corresponding to the geometric semantics of the task can be selected from the modeling skill set, and the operation type, target parameters, and dependencies between preceding and subsequent tasks can be written to form a modeling plan that can be scheduled for execution.

[0134] During runtime, the modeling plan serves as the basis for the main agent to schedule modeling skill agents. The operation type indicates the invocation of sketch, feature, or editing capabilities; the target parameters define the sketch plane, extrusion direction, extrusion length, rotation axis, rotation angle, edge set, and dimension parameters; and the modeling skill dependencies describe whether the generation of a particular feature depends on a preceding reference plane, contour, or entity boundary. In this way, discrepancy information is stably transformed into an ordered and executable parametric modeling plan, maintaining consistency with subsequent feature generation.

[0135] After adopting this implementation method, the differences in the structured state data can be directly mapped to ordered modeling tasks and further refined into modeling plans with dependencies. This ensures that the modeling plan generation process is consistent with the model geometric context, reduces the probability of feature order and parameter assignment errors, and enhances the callability and continuity of subsequent CAD kernel execution.

[0136] In one possible implementation, a parametric feature generation process is executed based on the modeling plan and the modeling skill set to obtain the current execution result. This includes: sequentially invoking the corresponding modeling skills in the modeling skill set according to the order of steps in the modeling plan. During the invocation process, the execution parameters required for each modeling skill are calculated based on the real-time geometric state of the current model. The calculation of execution parameters includes: determining the sketch plane and normal vector, identifying the 2D contour region, determining the stretching direction and termination surface, and performing geometric constraint verification. Parametric feature generation is performed based on the execution parameters to update the topology of the current model, and the updated current model is used as the current execution result.

[0137] In practical implementation, the corresponding modeling skill identifiers can be read sequentially according to the modeling plan, and the latest face, edge, and vertex relationships can be extracted from the current model to form a real-time geometric state. For features requiring sketch-driven operation, a sketch plane can be calculated based on the reference face in the current model, and the normal vector can be obtained from the geometric direction of the plane. Then, the 2D contour region can be identified by combining the closed boundary. For features requiring volume extension, the stretching direction can be determined based on the contour region and the target geometry, and the termination face can be determined by matching with adjacent faces or preset depth conditions. Subsequently, the sketch plane, contour closure, direction consistency, and termination face validity can be constrained and verified. When the verification passes, the corresponding modeling skill is called to generate parametric features, and the newly added faces, edges, and topological connection relationships are written into the current model.

[0138] During operation, the modeling plan provides sequential control, the modeling skill set provides function execution, real-time geometric state provides the basis for parameter calculation, and parametric feature generation completes the local reconstruction of the model topology, ensuring that the current execution result always corresponds to the model state after feature updates. This approach enables feature invocation to remain synchronized with the geometric state and allows subsequent modeling skills to continue execution based on the latest model.

[0139] Through the above processing, the operation sequence in the modeling plan can be stably implemented, the execution parameters can be obtained and verified in real time based on the current model, and the topology update results after parameterized feature generation can also be written back to the current model in a timely manner, thereby maintaining the continuity and consistency of the long-process modeling process and reducing feature positioning deviation and subsequent call mismatch.

[0140] When the master agent schedules the modeling skill agent to execute parametric feature generation, it needs to calculate the execution parameters based on the real-time geometric state of the current model. For example, for extrusion features, the system first determines the sketch plane and its normal vector (such as the Z-axis direction of the front reference plane), identifies the two-dimensional contour region (such as a rectangular sketch), and then calculates the extrusion direction and the termination surface (such as the positive Z-axis direction and the top surface of the model). During this process, geometric constraint verification rules (such as the extrusion direction needing to be parallel to the sketch plane normal vector) are executed synchronously. If the verification fails (such as a direction conflict), the master agent will record the reason for the failure and recommend a correction strategy (such as adjusting the termination surface selection). This logic ensures that parametric feature generation conforms to geometric constraints and avoids feature generation failures due to unreasonable parameters.

[0141] In some embodiments, it may further include a visual observation agent, a geometric selection agent, a sketch area selection agent, a CAD execution tool, and a reflection module.

[0142] The visual observation agent is responsible for rendering the current model and the target geometric reference from a specified viewpoint, comparing their differences from that viewpoint, and outputting a difference summary, key geometric facts, unresolved issues, and recommended observation views. This visual observation agent can be responsible only for generating local visual evidence and not for directly deciding the final modeling operation, thus avoiding misjudgments of local viewpoints from directly affecting CAD execution.

[0143] The modeling planning agent can be used to transform geometric facts obtained from multiple rounds of observations into a feature generation plan. Each planning step can include feature objectives, operation types, evidence perspectives, sketch planar cues, contour cues, termination conditions, and dependencies. For complex models, a basic body can be generated first, followed by the generation of additional bosses, fillets, or array structures.

[0144] Modeling skill agents can be used to convert planning steps into candidate parameters for a specific type of CAD operation. Taking the Extrude skill as an example, its output includes candidate operations, sketch plane view, sketch plane hints, termination face view, termination face hints, and candidate confidence. Taking the Revolve skill as an example, its output includes revolve profile, candidate revolve axis, candidate revolve angle, and generation / cut type. Different modeling skills can share the structured state data, observation, and reflection mechanisms of the main workflow, but each has its own different parameterization rules.

[0145] The geometry selection agent is responsible for selecting faces, edges, axes, or paths that meet certain conditions within the realistic topology of a CAD model. This agent enables visual cues to fall on geometric entities recognizable by the CAD kernel. For example, the Extrude skill requires selecting sketch faces and termination faces; the Revolve skill requires selecting the axis of rotation; the Chamfer skill requires selecting the set of edges; and the Hole Feature skill requires selecting the hole plane and hole axis direction.

[0146] The sketch region selection agent is responsible for defining a 2D contour region on a selected sketch plane. It can select a suitable contour based on occluded and unoccluded wireframe diagrams and fit the contour to CAD sketch line segments. If the selection result does not meet the target feature description, the main agent can call the tool again with correction prompts, such as adding a region, removing a hole, or correcting the boundary range.

[0147] The CAD execution tool is responsible for submitting verified parametric features to the CAD kernel for execution. Successful execution outputs a new CAD model file; failure returns error messages for the reflection module to determine whether the issue stems from incorrect geometry selection, incomplete parameters, topological conflicts, or an unreasonable modeling order.

[0148] The reflection module is responsible for performing process analysis after each tool call. Its output includes an understanding summary, tool result evaluation, structured state data editing, planning guidance, recommended next actions, and identifying blocking issues. This reflection module enables the system not only to call tools but also to adjust subsequent strategies based on tool results, avoiding continuous execution on incorrect paths.

[0149] The division of labor between the tools and the intelligent agent described above improves upon the instability of directly outputting CAD commands from large models in existing technologies. By breaking down complex tasks into sub-tasks such as observation, planning, skill parameterization, geometric selection, region selection, and execution, each sub-task has clearly defined inputs, outputs, and verification conditions, thereby improving the accuracy and interpretability of the generation process.

[0150] In some embodiments, the process of generating parameterized features can be represented by the following formula:

[0151]

[0152] Among them, F i For the generated parameterized features, OP i For the operation type of the i-th feature, G i P represents a geometric reference, such as a plane, axis, edge, or path. i Indicates operational parameters, such as length, angle, radius, quantity, or spacing; C i Represents a sketch outline, a selected region, or a topological set.

[0153] In other words, this embodiment employs a cyclical generation process. First, the generation task is initialized, establishing a task scene based on the current model and target geometric reference, and rendering the current and target geometric states from one or more initial perspectives. The initial images can be written into structured state data and used as visual evidence for subsequent observation, modeling planning, and modeling skill invocation.

[0154] Subsequently, the main agent enters a cyclical workflow. Each cycle can include five stages: understanding and reasoning, tool selection, tool execution, result reflection, and state update.

[0155] During the understanding and reasoning phase, the agent can read the current dialogue history, structured state data, observed images, and the current generated state to form an understanding of the target geometry and current differences. For example, it can determine that the current model is missing a certain boss, hole, groove, solid of revolution, or chamfer structure, and record the evidence perspective and confidence level of this judgment.

[0156] During the tool selection phase, the main agent can choose the next action based on its current understanding. If information is insufficient, it can choose an observation tool to gain more perspectives; if the overall structure is clear, it can choose a modeling and planning tool (i.e., a modeling and planning agent) to generate or revise the modeling plan; if a certain planning step already has the conditions for execution, it can choose the corresponding modeling skill. If complete parametric features are available, it can choose a CAD execution tool; if the target state is met, it can choose a termination tool.

[0157] During the tool execution phase, the selected modeling tool or skill can be invoked. For example, in the Extrude skill, extrusion candidates can be proposed first, then the sketch plane and termination plane can be selected, followed by the sketch contour region, and finally, an executable extrusion feature can be generated. Let the sketch plane be:

[0158]

[0159] in, Let be the sketch plane, o be the origin of the plane, n be the plane normal vector, and u and v be coordinate axes in the plane, satisfying: , and .

[0160] Let the stretching direction be d, the stretching length be L, and the sketch profile be S, then the stretching feature can be represented as: , of which F ext For the extrusion feature, `op` is for either adding or cutting the extrusion. If the extrusion termination face needs to remain parallel to the sketch plane, then verification can be performed. , where n s Let n be the normal vector of the sketch plane. t Let ε be the normal vector of the termination surface, and ε be the allowable error. The stretching length can be calculated from the distance between the starting surface and the termination surface along the stretching direction. The specific calculation formula is as follows:

[0161]

[0162] Among them, P t P represents the coordinates of a reference point on the starting plane (the plane on which the sketch is located). s Let L be the coordinates of the reference point on the termination surface, and L > 0 is required.

[0163] In the rotation skill, you can select the sketch plane, rotation profile, rotation axis, and rotation angle. Let the rotation axis be represented by a point a on the axis and the unit direction vector k, the sketch profile point be p, and the rotation angle be θ. Then the position of the point after rotation can be represented as:

[0164]

[0165] in, Rotation about a unit vector k The rotation matrix can be used to call the CAD kernel to generate revolved solids or revolved cut bodies.

[0166] During the results reflection phase, the tool execution results, generated images, error messages, current steps, and existing structured state data can be input into the reflection module. The reflection module determines whether the tool call was effective, whether additional observations are needed, whether the modeling plan needs to be revised, and whether a new region or geometric benchmark needs to be selected.

[0167] During the state update phase, the structured state data can be updated based on the analysis results, and subsequent modeling plans can be revised. If the CAD feature execution is successful, the current model path is updated to the newly generated model, the current planned step is marked as completed, and the process proceeds to the next executable step. If the tool fails, the reason for the failure is recorded, the current state is retained, and the main agent selects a new remedial action.

[0168] This cyclical process improves upon the shortcomings of existing technologies that can only generate data once and lack error correction, enabling the 3D CAD model generation process to have the capabilities of step-by-step verification, dynamic correction, and resumable execution.

[0169] In one possible implementation, the current execution result is analyzed to obtain the analysis result, including: obtaining the geometric feature data corresponding to the current execution result; comparing the geometric feature data with the target geometric reference, calculating the geometric overlap or feature missing rate, and determining the current model generation status based on the geometric overlap or feature missing rate; when the geometric overlap is lower than a preset threshold and / or the feature missing rate indicates the presence of missing features, generating an analysis result containing correction suggestions; when the geometric overlap is higher than or equal to a preset threshold and the feature missing rate indicates no feature missing features, setting the difference information in the structured state data to null, and generating an analysis result containing successful confirmation.

[0170] In the specific implementation, after the main agent completes the parameterized feature generation, it can first extract geometric feature data from the current execution result and write the extraction result to the analysis cache. Subsequently, it can establish a corresponding reference feature set based on the target geometric reference, and match the two from various dimensions such as size, position, contour closure, and feature existence to calculate the geometric overlap or feature missing rate. The geometric overlap can be obtained from the proportion of overlapping areas between the current model and the reference model, the proportion of key feature matching, or the consistency of the projected contour. The feature missing rate is used to characterize the proportion of features that exist in the target geometric reference but have not yet appeared in the current execution result. When the geometric overlap is lower than a preset threshold or the feature missing rate indicates that there are missing features, an analysis result containing correction suggestions can be generated. The correction suggestions correspond to the reconstruction of missing features, relocation of biased features, or size correction, and are written back to the subsequent modeling plan.

[0171] When the geometric overlap is higher than or equal to a preset threshold and the feature missing rate indicates no feature missing, the difference information in the structured state data can be cleared, and an analysis result containing successful confirmation can be output to indicate that the current execution result meets the target geometric reference. This analysis mechanism forms a closed-loop verification during the generation process, ensuring that subsequent modeling plans are only corrected when differences exist, and generation is terminated directly after the differences are eliminated.

[0172] When the analysis results indicate that the matching degree between the current execution result and the target geometric reference meets the standard (e.g., the geometric overlap is higher than the preset threshold and the feature missing rate is 0), the main agent can empty the difference information field in the structured state data and mark the model version as "converged". If the number of iterations in the modeling plan reaches the preset iteration threshold (e.g., 10 rounds), the generation process can also be terminated. This condition setting ensures that the update of the structured state data is consistent with the preset generation conditions, avoiding infinite loops caused by context loss.

[0173] Through the above implementation, the consistency of the current execution result can be determined, and the geometric comparison result can be directly converted into correction suggestions or success confirmation information, thereby keeping the structured state data synchronized with the generated state and improving the state continuity and result controllability in multi-round parametric modeling.

[0174] Building upon the foregoing embodiments, the modeling skill set further includes at least one of the following: sketching skills for generating two-dimensional contours on a specified plane; feature manipulation skills for generating three-dimensional solids or performing Boolean operations based on the two-dimensional contours; transformation manipulation skills for performing spatial translation, rotation, or mirror symmetry processing on model features; and chamfering and rounding skills for smoothing model edges.

[0175] In the actual modeling process, after receiving the current model and the target geometric reference, the main agent can write the difference information into the structured state data and use this information to call the modeling planning agent to generate an ordered modeling plan. When the modeling plan needs to form a new two-dimensional cross-section, the sketching skill can be used to complete the contour construction. When it is necessary to form a solid or supplement local volume, the feature operation skill can be used to complete parametric generation or Boolean processing. When it is necessary to adjust the position, orientation, or symmetry relationship of features, the transformation operation skill can be used to perform spatial transformation. When it is necessary to transition sharp edges, the chamfering and rounding skills can be used to complete edge correction. The results of each modeling skill execution can be written back into the structured state data as a basis for subsequent modeling plan correction and execution result analysis.

[0176] By unifying 2D contour generation, 3D feature construction, spatial transformation, and edge smoothing capabilities into a unified set of modeling skills, the difference identification results can be stably mapped into callable parametric modeling actions, maintaining the continuity of the model generation process. Thus, the executed results, remaining differences, and subsequent skill selections in the structured state data can form a consistent correlation, giving the generated CAD model a clear feature hierarchy and subsequent editability.

[0177] Based on the aforementioned embodiments, the preset generation conditions further include: the difference information in the structured state data is empty, and the matching degree between the current execution result and the target geometric reference reaches a preset standard; or, the number of iterations of the modeling plan reaches a preset iteration threshold.

[0178] In the specific implementation, after each round of execution, the main agent can obtain the geometric feature data of the current execution result and perform matching calculations with the feature set corresponding to the target geometric reference. The matching degree can be determined by a combination of factors such as the proportion of overlapping areas, the feature missing rate, or the key size deviation. When the matching degree reaches a preset standard and there are no unresolved issues in the structured state data, the difference information can be written blank, and a generation completion signal can be output. If there are still unmet requirements in the current execution result, the modeling and planning agent can generate subsequent modeling plans based on the updated difference information, and the main agent will continue to schedule the corresponding modeling skill agents to execute, thus forming an iterative generation.

[0179] If the modeling plan fails to meet the preset generation conditions after multiple rounds of execution, the number of repeated executions of the same modeling plan can be accumulated and recorded. This accumulated record can be stored in structured state data and compared with a preset iteration threshold. When the preset iteration threshold is reached, it indicates that the current difference cannot converge within the predetermined number of rounds, and the execution of subsequent modeling plans can be stopped, and an end command can be output to prevent the program from entering an infinite loop. In practical applications, the preset iteration threshold can be set according to the model complexity, the number of features, and the task tolerance; this embodiment does not impose specific limitations on this.

[0180] With the above settings, the model generation process can end naturally when the differences have been eliminated and the matching degree meets the standard, or it can be terminated in time when the number of iterations is limited. This gives the generation process a clear termination criterion and maintains the consistency between the structured state data and the current execution result, reducing invalid repeated execution.

[0181] In this embodiment of the application, for example in the field of mechanical manufacturing, engineers can generate editable parametric CAD models based on scanned images of physical objects or multi-view renderings provided by users, for subsequent simulation analysis, processing and manufacturing, or 3D printing. To address situations where multi-view information integration is difficult, the generated model is not editable, and the scalability of modeling skills is poor, a main intelligent agent, a modeling planning intelligent agent, and multiple modeling skill intelligent agents can be constructed first, and structured state data for recording information about the model generation process can be initialized. In the example of generating a parametric CAD model containing bosses, through holes, and chamfers from multi-view images of a mechanical part, the target geometric reference can be the multi-view rendering of the part, and the current model can be an empty shell base. The main intelligent agent reads the current model, the target geometric reference, the allowed view set, and the modeling skill set, where the allowed view set includes the front view, side view, and top view, and the modeling skill set includes extrusion, rotation, and chamfering. The main agent simultaneously acquires a pre-set parametric modeling instruction library, extracts multiple callable atomic modeling operations from the library, and maps these atomic modeling operations to a set of modeling skills. The modeling skill set encapsulates corresponding tool call interfaces and geometric verification rules. The modeling skill set includes at least one or more of the following: sketching skills, feature manipulation skills, transformation manipulation skills, and chamfering and filleting skills. The system renders the initial view of the current model and the target geometric reference. The initial view can be a front view or a top view, and a structured memory is established to explicitly record observed views, visual difference summaries, and unresolved issues. The visual difference summary can record "The current model lacks a boss and a through hole," and the unresolved issue can record "The boss sketch plane needs to be determined." The task scenario description is explicitly stated as "Generate a boss in plane A, generate a through hole in plane B, and add a chamfer."

[0182] The main agent can perform difference detection based on the current model and the target geometric reference, and write the detected difference information into the structured state data. The visual observation agent can render multi-view images of the current model and the target geometric reference based on the front view, side view, and top view from the allowed viewpoint set. After comparing the differences, it outputs a difference summary: "The current model lacks a boss and a through-hole," and a recommended observation view: "Observe the side view to confirm the boss height." The structured memory in the structured state data can be updated accordingly to: "The boss needs to be generated in plane A, with a height of 10 mm; the through-hole needs to be generated in plane B, with a diameter of 8 mm." The modeling planning agent reads the difference information in the structured state data, analyzes the missing geometric features or existing geometric deviations in the current model, and transforms the geometric features or geometric deviations into a modeling task sequence containing multiple ordered actions according to a preset planning strategy. It then matches corresponding modeling skills to each action in the modeling task sequence, forming a modeling plan that includes operation type, target parameters, and modeling skill dependencies. The modeling plan can adopt a rolling modeling plan and be organized in conjunction with an evidence perspective, which can include a "side view" to confirm the boss height. The modeling and planning agent can generate a rolling modeling plan based on difference summaries and structured memory. This includes: extrude a boss into plane A (operation type: extrude add), with plane A as the sketch plane, the extrude direction along the Z-axis, and a length of 10 mm; generate a through hole into plane B (operation type: extrude cut), with plane B as the sketch plane, the extrude direction along the Z-axis, a length of 20 mm, and a diameter of 8 mm; and add a chamfer to the boss edge (operation type: chamfer), with the edge set being the top edge of the boss and a size of 1.5 mm. When the structured state data is updated in subsequent iterations, the modeling and planning agent can dynamically correct the modeling plan according to the updated difference information. For example, if subsequent observation reveals that the boss height needs to be adjusted to 12 mm, the modeling and planning agent automatically corrects the extrude length in the corresponding extrude add action in plane A to 12 mm.

[0183] The main agent can schedule corresponding modeling skill agents according to the modeling plan, and sequentially call the corresponding modeling skills in the modeling skill set according to the order in the modeling plan. During the calling process, the execution parameters of each modeling skill are calculated in combination with the real-time geometric state of the current model. The execution parameter calculation process includes determining the sketch plane and normal vector, identifying the 2D contour region, determining the extrusion direction and termination surface, and performing geometric constraint verification. For extrusion addition, the modeling skill agent can extract plane A of the sketch plane, the Z-axis of the extrusion direction, and a length of 10 mm, and verify that "the extrusion direction must be parallel to the normal vector of plane A"; when it is necessary to cooperate with the sketch drawing skill, a 2D contour is generated on the specified plane A, and then a 3D solid is generated based on the 2D contour through the feature operation skill. For extrusion cut, the modeling skill agent can extract plane B of the sketch plane, the Z-axis of the extrusion direction, a length of 20 mm, and a diameter of 8 mm, and verify that the termination surface is parallel to the sketch plane; after generating the corresponding through hole 2D contour on the specified plane B, Boolean cut is performed through the feature operation skill. For chamfering, the modeling skill agent can extract the top edge of the edge set boss and its size of 1.5 mm, verify the continuity of the edge set, and perform smoothing processing on the model edges using chamfering and rounding skills. For transformation operation skills in the modeling skill set, spatial translation, rotation, or mirror symmetry processing can also be performed on the model features, where rotation belongs to the currently read modeling skill set content. After each parametric feature generation operation is completed, the topology of the current model is updated, and the updated model is recorded as the current execution result.

[0184] The CAD execution tool can call the CAD kernel to generate new models. The CAD kernel is the underlying geometric modeling engine and can return execution results, which can be either successful or failed. The main agent obtains the geometric feature data corresponding to the current execution result and compares the geometric feature data with the target geometric reference to calculate the geometric overlap or feature missing rate, thereby determining the current generation status of the model. The reflection module evaluates the execution result and updates the structured memory in the structured state data. In one type of update scenario, it records "the boss has been generated, and the diameter of the through hole needs to be adjusted" and suggests the next action "correct the diameter of the through hole to 8 mm". When the geometric overlap is lower than a preset threshold and / or the feature missing rate indicates that there are missing features, an analysis result containing correction suggestions is generated. The main agent updates the structured state data based on the analysis result and submits the corrected difference information to the modeling planning agent to correct the subsequent modeling plan. Then, the corresponding modeling skill agent is re-scheduled to execute the parametric feature generation process. If the through hole diameter fails due to parameter errors, the structured state data records the reason for the failure and recommends correcting the through hole diameter to 8 mm to avoid repeatedly using incorrect parameters. When the geometric overlap is higher than or equal to a preset threshold and the feature missing rate indicates no feature missing, the difference information in the structured state data is cleared, and an analysis result containing successful confirmation is generated. The generation process ends when the difference information in the structured state data is cleared and the matching degree between the current execution result and the target geometric reference reaches a preset standard, or when the number of iterations in the modeling plan reaches a preset iteration threshold. Throughout the process, the structured memory in the structured state data continuously and explicitly stores key geometric facts such as "the current model lacks a boss and a through hole," "the boss needs to be generated in plane A with a height of 10 mm," "the through hole needs to be generated in plane B with a diameter of 8 mm," and "the boss has been generated, and the diameter of the through hole needs to be adjusted," while retaining design intentions such as sketches and dimensional parameters.

[0185] This application embodiment can divide an electronic device or main control device into functional modules according to the above method examples. For example, each function can be divided into its own functional modules, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional module. It should be noted that the module division in this embodiment is illustrative and only represents one logical functional division; in actual implementation, there may be other division methods.

[0186] Figure 2 This is a schematic diagram of a three-dimensional computer-aided design model generation device based on a multimodal intelligent agent, provided as an embodiment of this application. Figure 2 As shown, the device includes: a construction module 210, an acquisition module 220, a determination module 230, a modeling module 240, an execution module 250, an analysis module 260, an update module 270, and a completion module 280.

[0187] The construction module 210 is used to construct the main intelligent agent, the modeling planning intelligent agent and multiple modeling skill intelligent agents, and to initialize the structured state data used to record information about the model generation process.

[0188] The acquisition module 220 is used to acquire the current model, target geometric reference, and modeling skill set through the main intelligent agent.

[0189] The determination module 230 is used to determine the difference information based on the current model and the target geometric reference through the main intelligent agent, and write the difference information into the structured state data.

[0190] Modeling module 240 is used to plan an agent through modeling, and dynamically generate or modify a modeling plan containing ordered steps based on the difference information in the structured state data.

[0191] The execution module 250 is used to schedule the corresponding modeling skill agent through the main agent, and execute the parameterized feature generation process according to the modeling plan and the modeling skill set to obtain the current execution result.

[0192] Analysis module 260 is used to analyze the current execution results through the main intelligent agent and obtain the analysis results.

[0193] The update module 270 is used to update the structured state data and correct the subsequent modeling plan according to the analysis results when the analysis results do not meet the preset generation conditions. It then returns the steps of scheduling the corresponding modeling skill agent through the main agent, executing the parameterized feature generation process according to the modeling plan and the modeling skill set, and obtaining the current execution result.

[0194] Module 280 is used to end the generation process when the analysis results meet the preset generation conditions.

[0195] In one possible implementation, the acquisition module 220 is specifically used for:

[0196] Obtain the initial CAD model of the target 3D object to be processed as the current model.

[0197] Obtain multi-view renderings, 2D sketches, or text descriptions of the target 3D object as geometric references.

[0198] Obtain a pre-set parametric modeling instruction library and extract multiple callable atomic modeling operations contained in the parametric modeling instruction library. Map the atomic modeling operations to a set of modeling skills. The set of modeling skills encapsulates the corresponding tool call interface and geometric verification rules.

[0199] In one possible implementation, the modeling module 240 is specifically used for:

[0200] Analyze the discrepancies in structured state data to identify missing geometric features or existing geometric deviations in the current model.

[0201] Based on a pre-defined planning strategy, geometric features or geometric deviations are transformed into a sequence of modeling tasks containing multiple ordered steps.

[0202] Match the corresponding modeling skills to each step in the modeling task sequence, and generate a modeling plan that includes operation type, target parameters, and modeling skill dependencies.

[0203] In one possible implementation, the execution module 250 is specifically used for:

[0204] In accordance with the order of steps in the modeling plan, the corresponding modeling skills in the modeling skill set are called in sequence.

[0205] During the invocation process, the execution parameters required for each modeling skill are calculated based on the real-time geometric state of the current model. The calculation of execution parameters includes: determining the sketch plane and normal vector, identifying the two-dimensional contour region, determining the stretching direction and termination surface, and performing geometric constraint verification.

[0206] Parametric features are generated based on execution parameters to update the topology of the current model, and the updated current model is used as the current execution result.

[0207] In one possible implementation, the analysis module 260 is specifically used for:

[0208] Obtain the geometric feature data corresponding to the current execution result.

[0209] The geometric feature data is compared with the target geometric reference to calculate the geometric overlap or feature missing rate, and the current generation status of the model is determined based on the geometric overlap or feature missing rate.

[0210] When the geometric overlap is below a preset threshold and / or the feature missing rate indicates the presence of missing features, an analysis result containing correction suggestions is generated.

[0211] When the geometric overlap is higher than or equal to a preset threshold and the feature missing rate indicates no feature missing, the difference information in the structured state data is cleared, and an analysis result containing successful confirmation is generated.

[0212] In one possible implementation, the set of modeling skills acquired by the acquisition module 220 includes at least one of the following:

[0213] Sketching skills are used to generate two-dimensional outlines on a specified plane.

[0214] Feature manipulation skills are used to generate 3D entities based on 2D contours or to perform Boolean operations.

[0215] Transformation operations are used to perform spatial translation, rotation, or mirror symmetry processing on model features.

[0216] Chamfering and rounding techniques are used to smooth the edges of a model.

[0217] In one possible implementation, the preset generation conditions in the update module 270 and the completion module 280 include:

[0218] The difference information in the structured state data is empty, and the matching degree between the current execution result and the target geometric reference reaches the preset standard; or, the number of iterations of the modeling plan reaches the preset iteration threshold.

[0219] This embodiment provides a three-dimensional computer-aided design model generation device based on multimodal intelligent agents, which can execute a three-dimensional computer-aided design model generation method based on multimodal intelligent agents in the above embodiment. Its implementation principle and technical effect are similar, and will not be described again here.

[0220] In the aforementioned specific implementation of a three-dimensional computer-aided design model generation device based on multimodal intelligent agents, each module can be implemented as a processor. The processor can execute computer execution instructions stored in the memory, thereby enabling the processor to execute the aforementioned three-dimensional computer-aided design model generation method based on multimodal intelligent agents.

[0221] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 3 As shown, the electronic device includes at least one processor 310 and a memory 320. The electronic device also includes a communication component 330. The processor 310, memory 320, and communication component 330 are connected via a bus 340.

[0222] In the specific implementation process, at least one processor 310 executes computer execution instructions stored in memory 320, causing at least one processor 310 to execute a three-dimensional computer-aided design model generation method based on multimodal intelligent agents as executed on the electronic device side.

[0223] The specific implementation process of processor 310 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0224] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0225] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage.

[0226] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0227] The above description of the functions implemented by electronic devices and main control devices has introduced the solutions provided by the embodiments of the present invention. It is understood that, in order to implement the above functions, the electronic device or main control device includes hardware structures and / or software modules corresponding to the execution of each function. By combining the units and algorithm steps of the various examples described in the embodiments of the present invention, the embodiments of the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the technical solutions of the embodiments of the present invention.

[0228] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the above-described method for generating a three-dimensional computer-aided design model based on a multimodal intelligent agent.

[0229] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0230] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in an electronic device or a host device.

[0231] This application also provides a computer program product, which includes a computer program stored in a readable storage medium. At least one processor of an electronic device can read the computer program from the readable storage medium, and the at least one processor executes the computer program to cause the electronic device to perform the solution provided in the above embodiments.

[0232] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disk, or optical disk.

[0233] The technical solutions of this application have been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it is readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. The above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for generating three-dimensional computer-aided design models based on multimodal intelligent agents, characterized in that, include: Construct a main intelligent agent, a modeling and planning intelligent agent, and multiple modeling skill intelligent agents, and initialize structured state data for recording information about the model generation process; The main intelligent agent obtains the current model, target geometric reference, and modeling skill set. The main intelligent agent determines the difference information based on the current model and the target geometric reference, and writes the difference information into the structured state data. The modeling and planning agent analyzes the difference information in the structured state data to identify the missing geometric features or existing geometric deviations in the current model; based on a preset planning strategy, the geometric features or geometric deviations are transformed into a modeling task sequence containing multiple ordered steps. Match the corresponding modeling skill to each step in the modeling task sequence, and generate a modeling plan that includes operation type, target parameters and modeling skill dependencies; The main agent retrieves a modeling skill agent from the modeling skill set that matches the current step to be executed in the modeling plan, and sends an execution request to it. The modeling skill agent then performs a parameterized feature generation process according to the modeling plan and the modeling skill set to obtain the current execution result. Obtain the geometric feature data corresponding to the current execution result; The geometric feature data is compared with the target geometric reference to calculate the geometric overlap or feature missing rate, and the generation status of the current model is determined based on the geometric overlap or feature missing rate. When the geometric overlap is lower than a preset threshold and / or the feature missing rate indicates the presence of missing features, an analysis result containing correction suggestions is generated; when the geometric overlap is higher than or equal to the preset threshold and the feature missing rate indicates no feature missing features, the difference information in the structured state data is set to empty, and an analysis result containing successful confirmation is generated; after the analysis result is generated, an analysis field is written into the structured state data. When the analysis result does not meet the preset generation conditions, the structured state data is updated according to the analysis result and the subsequent modeling plan is corrected. Then, the process of scheduling the corresponding modeling skill agent through the main agent, executing the parameterized feature generation process according to the modeling plan and the modeling skill set, and obtaining the current execution result is returned. The generation process ends when the analysis results meet the preset generation conditions.

2. The method according to claim 1, characterized in that, The acquisition of the current model, target geometric reference, and modeling skill set includes: Obtain the initial computer-aided design CAD model of the target 3D object to be processed as the current model; Obtain multi-view renderings, 2D sketches, or text descriptions of the target 3D object as geometric references for the target; Obtain a pre-set parametric modeling instruction library, extract multiple callable atomic modeling operations contained in the parametric modeling instruction library, map the atomic modeling operations to the modeling skill set, and the modeling skill set encapsulates the corresponding tool call interface and geometric verification rules.

3. The method according to claim 1, characterized in that, The step of performing a parametric feature generation process based on the modeling plan and the modeling skill set to obtain the current execution result includes: In accordance with the order of steps in the modeling plan, the corresponding modeling skills in the modeling skill set are called sequentially; During the invocation process, the execution parameters required for each modeling skill are calculated based on the real-time geometric state of the current model. The calculation of the execution parameters includes: determining the sketch plane and normal vector, identifying the two-dimensional contour region, determining the stretching direction and termination surface, and performing geometric constraint verification. Based on the execution parameters, parameterized features are generated to update the topology of the current model, and the updated current model is used as the current execution result.

4. The method according to claim 1, characterized in that, The set of modeling skills includes at least one of the following: Sketching skills are used to generate two-dimensional outlines on a specified plane; Feature manipulation skills are used to generate 3D solids based on 2D contours or to perform Boolean operations; Transformation manipulation skills are used to perform spatial translation, rotation, or mirror symmetry processing on model features; Chamfering and rounding techniques are used to smooth the edges of a model.

5. The method according to any one of claims 1 to 4, characterized in that, The preset generation conditions include: The difference information in the structured state data is empty, and the matching degree between the current execution result and the target geometric reference reaches the preset standard; Alternatively, the number of iterations of the modeling plan reaches a preset iteration threshold.

6. A three-dimensional computer-aided design model generation device based on multimodal intelligent agents, characterized in that, include: The building module is used to build the main intelligent agent, the modeling and planning intelligent agent, and multiple modeling skill intelligent agents, and initialize the structured state data used to record information about the model generation process; The acquisition module is used to acquire the current model, target geometric reference, and modeling skill set through the main intelligent agent; The determination module is used to determine the difference information based on the current model and the target geometric reference through the main intelligent agent, and write the difference information into the structured state data; The modeling module is used to analyze the difference information in the structured state data through the modeling planning agent, identify the missing geometric features or existing geometric deviations in the current model, and transform the geometric features or geometric deviations into a modeling task sequence containing multiple ordered steps based on a preset planning strategy. Match the corresponding modeling skill to each step in the modeling task sequence, and generate a modeling plan that includes operation type, target parameters and modeling skill dependencies; The execution module is used to retrieve, through the main agent, a modeling skill agent matching the current step to be executed in the modeling plan from the modeling skill set, and send an execution request to it. The modeling skill agent executes the parameterized feature generation process according to the modeling plan and the modeling skill set to obtain the current execution result. The analysis module is used to obtain the geometric feature data corresponding to the current execution result; The geometric feature data is compared with the target geometric reference to calculate the geometric overlap or feature missing rate, and the generation status of the current model is determined based on the geometric overlap or feature missing rate. When the geometric overlap is lower than a preset threshold and / or the feature missing rate indicates the presence of missing features, the analysis result containing correction suggestions is generated; when the geometric overlap is higher than or equal to the preset threshold and the feature missing rate indicates no feature missing features, the difference information in the structured state data is set to empty, and an analysis result containing successful confirmation is generated; after the analysis result is generated, the analysis field is written into the structured state data. The update module is used to update the structured state data and correct the subsequent modeling plan according to the analysis results when the analysis results do not meet the preset generation conditions, and return to the step of scheduling the corresponding modeling skill agent through the main agent, executing the parameterized feature generation process according to the modeling plan and the modeling skill set to obtain the current execution result. The completion module is used to end the generation process when the analysis results meet the preset generation conditions.

7. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 5.

Images