World model dynamic generation method, electronic device, and computer program product

Abstract

The application provides a dynamic generation method of a world model, an electronic device and a computer program product. The method is used for multi-dimensional constraint correction of multi-source original data based on a basic model object generated by single original data. The basic model object generated by single original data provides a possibility for dynamic and rapid generation of a basic model object. Through multi-dimensional constraint correction of the basic model object under the action of multi-source original data, a high-precision, scalable and dynamically evolving world model is constructed.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a method for dynamically generating world models, electronic devices, and computer program products. Background Technology

[0002] The concept of "world models" originated in cognitive science and control theory, used to describe how intelligent agents construct internal representations of their external environment in their minds to support prediction, planning, and decision-making. In 2018, Schmidhuber's team first proposed WorldModels and used deep generative models (such as VAEs and RNNs) to build a learning system that can predict the future state of the environment, enabling intelligent agents to complete reinforcement learning training in a "dream." This marked the transition of world models from theoretical abstraction to a trainable neural network framework.

[0003] With the rapid development of artificial intelligence technologies such as Transformer and Diffusion, world models have been further expanded into "spatial models" capable of high complexity, multimodality, and long time series. In the fields of embodied intelligence, autonomous driving, robotics, and general artificial intelligence, as well as in industry applications such as rail transit, water conservancy and water affairs, emergency fire protection, urban governance, smart cities, smart power, and smart factories, world models have become one of the important research directions. World models can be understood, perceived, reasoned about, predicted, planned, and executed, and are not just a mapping of the world.

[0004] For example, US Patent 20250003764A1 discloses the construction of a world model based on autonomous driving scenarios. Specifically, it compares data obtained by autonomous vehicles during operation through sensors (such as light detection and ranging sensors, radar sensors, cameras, or inertial measurement units) with a pre-generated world model based on high-definition map data. By comparing and detecting errors between reality and the model, it generates data for feature semantics and geometric correction of the world model based on these errors. Then, it modifies the original world model using the corrected data and uses the modified world model to navigate the autonomous vehicle.

[0005] For example, Chinese patent CN117993149A discloses a method for constructing a simulated world model based on these configuration parameters, and for arranging spatial layout and furniture within the space formed in this simulated world model. Specifically, it constructs models of walls, doors, etc., based on configuration parameters such as geometric dimensions, and then generates layout parameters such as graphical structures based on the connection methods of various spaces such as the living room, dining room, and bedroom, ultimately realizing the construction of the space.

[0006] Therefore, traditional world model construction typically follows a specific technological path, such as sensor-centric geometric modeling, map-centric structural reconstruction, or rule-based space generation methods centered on configuration parameters. However, regardless of the approach, the core characteristic of traditional world models is their strong coupling dependence on a single technological link or a single type of data. That is, model construction heavily relies on data from a specific source or "data stacking" within a particular framework. For example, stacking semantic labels on a geometric model or dimensional parameters on a rule-based space represents a solution of "simple data overlay."

[0007] Therefore, the more data is piled up, the more redundant and cumbersome the traditional world model construction scheme becomes. Moreover, world models that heavily rely on single-dimensional data will face significant challenges in terms of accuracy and flexibility. Constructing accurate and lightweight world models is a technical challenge in this field, and also a goal and direction that is constantly being pursued and broken through. Only realistic world models can empower the innovative development of the industry. Summary of the Invention

[0008] One objective of this application is to obtain a highly accurate world model, and to provide a method for dynamically generating world models, an electronic device, and a computer program product.

[0009] According to one aspect of the embodiments of this application, a method for dynamically generating a world model is disclosed, the method comprising:

[0010] A basic model object is generated from a type of raw data to create the initial three-dimensional form of the world model it describes. The basic model object is an initial geometric assumption model. The type of raw data is geometric data, laser point cloud data, or visual data.

[0011] Initial parameters are extracted from various types of raw data to obtain geometric structure information and physical information from geometric data, spatial location information and surface information from laser point cloud data, and contour information and semantic information from visual data;

[0012] Using the geometric data, visual data, and laser point cloud data as mutual parameter constraints, geometric parameters and physical function parameters for modifying the basic model object are obtained; wherein the parameter constraint step includes: constraining and correcting the initial geometric parameters of the basic model object by linking the geometric structure information, spatial position information, and contour information; and constraining and correcting the physical function parameters of the basic model object by linking the physical information, surface information, and semantic information.

[0013] The world model is obtained by reconstructing the basic model object of the laser point cloud data based on the geometric parameters and the physical function parameters.

[0014] According to one aspect of the embodiments of this application, an electronic device is disclosed, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method as described in any of the preceding claims.

[0015] According to one aspect of the embodiments of this application, a computer program product is disclosed, including a computer program that, when executed by a processor, implements the steps of the method as described in any of the preceding claims.

[0016] This application embodiment focuses on the basic model object generated from a single source of raw data, and performs multi-dimensional constraint correction on multiple sources of raw data. The generation of basic model objects from a single source of raw data makes it possible to dynamically and quickly generate basic model objects. Furthermore, by performing multi-dimensional constraint correction on the basic model object under the influence of multiple sources of raw data, a high-precision, scalable, and dynamically evolving world model can be constructed.

[0017] Furthermore, by fusing multi-source raw data, namely geometric, laser point cloud, and visual data, we can achieve comprehensive capture of three-dimensional spatial information, geometric features, surface properties, and semantic information, thereby improving the accuracy and completeness of dynamically generated world models. Based on the parameters provided by various types of raw data, we can perform multi-dimensional parameter correction on the basic model objects generated by one type of raw data to obtain corrected geometric parameters and physical function parameters, and then use these to optimize the basic model objects to obtain a reconstructed world model.

[0018] By adjusting the geometric and physical parameters obtained through multi-dimensional parameter linkage constraints, the consistency of the optimized world model in spatial morphology, physical characteristics, and visual semantics can be ensured, reducing the bias caused by single-source original data and improving the reliability of the world model in complex scenarios.

[0019] Since the embodiments of this application generate a basic model object that can implement constraint corrections based on a single original data, it can be applied to different types of original data, making it easy to implement in different scenarios and applications, and has high scalability and versatility.

[0020] In addition, the model object is described by geometric parameters and the physical function of the model object is described by text. The text semantics are expressed in the form of bytes in the underlying code. It can directly call the third-party large language model to realize the perception, understanding and reasoning of the model. The result data of perception, understanding and reasoning can be fed back to the model object expressed by geometric parameters. The model object can become an intelligent agent that plans and executes autonomously, rather than a replica of animation.

[0021] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0022] It should be understood that the above general description and the following detailed description are merely exemplary and do not limit this application. Attached Figure Description

[0023] The above and other objectives, features and advantages of this application will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0024] Figure 1 A flowchart illustrating a method for dynamically generating a world model according to an embodiment of this application is shown.

[0025] Figure 2 A schematic diagram of a world model from the perspective of an exhibition hall space according to an embodiment of this application is shown.

[0026] Figure 3 A schematic diagram of a world model of a small-scale indoor space of a laboratory according to an embodiment of this application is shown.

[0027] Figure 4 A schematic diagram illustrating a joint constraint based on geometric data, laser point cloud data, and visual data according to an embodiment of this application is shown.

[0028] Figure 5 It is based on Figure 1 The flowchart shown in the corresponding embodiment describes the steps of constraining and correcting the initial geometric parameters of the basic model object by linking geometric structure information, spatial position information, and contour information.

[0029] Figure 6 It is based on Figure 1 The corresponding embodiment shows a flowchart describing the steps of constraining the physical functional parameters of a basic model object by linking physical information, surface information, and semantic information.

[0030] Figure 7 It is based on Figure 1 The flowchart described in the corresponding embodiment describes the steps of reconstructing a world model from a basic model object based on geometric parameters and physical function parameters.

[0031] Figure 8 This is a schematic diagram illustrating a basic model object generated based on geometric data, laser point cloud data, or visual data according to an embodiment of this application.

[0032] Figure 9 yes Figure 8A rendering showing the effect of mutual constraints between walls (partially) in a space.

[0033] Figure 10 It is based on Figure 4 The flowchart described in the corresponding embodiment describes the steps of reconstructing a world model from a basic model object based on geometric parameters and physical function parameters.

[0034] Figure 11 This is a schematic diagram illustrating the spatial relationships and subordination relationships of a world model in a certain scenario according to an embodiment of this application.

[0035] Figure 12 It is based on Figure 4 The flowchart described in the corresponding embodiment describes the steps of reconstructing a world model from a basic model object based on geometric parameters and physical function parameters.

[0036] Figure 13 This is a schematic diagram illustrating the long and short memory mechanism of a world model in a certain scenario according to an embodiment of this application.

[0037] Figure 14 It is based on Figure 4 The flowchart described in the corresponding embodiment describes the steps of reconstructing a world model from a basic model object based on geometric parameters and physical function parameters.

[0038] Figure 15 This is a schematic diagram illustrating social relationships under a certain logical relationship in a world model according to an embodiment of this application.

[0039] Figure 16 It is based on Figure 4 The flowchart described in the corresponding embodiment describes the steps of reconstructing a world model from a basic model object based on geometric parameters and physical function parameters. Detailed Implementation

[0040] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided to make the description of this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The drawings are merely illustrative of this application and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.

[0041] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more exemplary embodiments. Numerous specific details are provided in the following description to give a full understanding of exemplary embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced with one or more of the specific details omitted, or other methods, components, steps, etc., can be employed. In other instances, well-known structures, methods, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0042] The construction of a "world model" requires multi-dimensional data fusion and expression. It should integrate current diverse technologies and different data, rather than relying solely on a single technology, fragmented data integration, or stacked data combinations to achieve a "replication" of the model.

[0043] Our understanding of the world should return to the perspective of "humanity." Integrating rational understanding "from the inside out" with emotional understanding "from the surface to the core," and conducting a holistic analysis, understanding, and reasoning about the world and the changes that affect it, forms the foundational model for spatial intelligence.

[0044] Therefore, constructing a world model cannot rely on just one dimension. It should be done from a "technology equality" perspective, integrating data from various dimensions with different technologies to create a world model with a "skeleton, skin, and clothing" structure and / or "attribute relationships, memory relationships, and social relationships." This approach fully leverages the fusion and mutual constraints of multi-dimensional data, rather than simply replicating a world model through rigid data stacking. The resulting world model is not only realistic but also infinitely close to being understood, perceived, reasoned about, predicted, planned, and executed.

[0045] See Figure 1 , Figure 1 A flowchart illustrating a method for dynamically generating a world model according to an embodiment of this application is shown. An embodiment of this application provides a method for dynamically generating a world model, comprising:

[0046] Step S110: Generate a basic model object of the initial three-dimensional form of the world model it describes using a type of raw data. The basic model object is the initial geometric assumption model. The current type of raw data is geometric data, laser point cloud data or visual data.

[0047] Step S120: Initial parameters are extracted from various types of raw data to obtain geometric structure information and physical information from geometric data, spatial location information and surface information from laser point cloud data, and contour information and semantic information from visual data.

[0048] Step S130: Using geometric data, visual data, and laser point cloud data as mutual parameter constraints, obtain geometric parameters and physical function parameters for modifying the basic model object; wherein the parameter constraint step includes: constraining and correcting the initial geometric parameters of the basic model object by linking geometric structure information, spatial position information, and contour information; and constraining and correcting the physical function parameters of the basic model object by linking physical information, surface information, and semantic information.

[0049] Step S140: Reconstruct the world model from the basic model object based on geometric parameters and physical function parameters.

[0050] The above steps will be explained in detail with specific application examples, as follows:

[0051] The basic model object is the starting point for dynamically generating the corresponding world model. It is generated from any type of raw data and serves as the initial carrier for subsequent multi-source information fusion and dynamic optimization. The basic model object is an initial geometric assumption model that abstractly represents the target world. It contains at least several initial geometric parameters and physical functional information. The geometric parameters describe the spatial morphology and structural outline.

[0052] In practical implementation, when the input is geometric data, the geometric structure and physical information are extracted to construct a geometric model with basic structural constraints. When the input is laser point cloud data, an initial morphological model corresponding to the 3D point set is generated based on spatial location and surface information. When the input is visual data, a geometric hypothesis model with preliminary spatial inference capabilities is generated based on contour and semantic information. In summary, a basic model object will be generated from the current type of raw data, namely, any one of the following: the geometric model, the initial morphological model corresponding to the 3D point set, or the geometric hypothesis model.

[0053] Regardless of the type of raw data used, the generated basic model object serves as a unified model representation framework, carrying subsequent supplementary constraint information from various data sources. Based on this, the geometric and physical functional parameters of the basic model object are linked and corrected through mutual parameter constraints from multiple sources of raw data, enabling the basic model object to gradually evolve from a single-source hypothetical model into a multi-source consistent, parameter-coordinated world model.

[0054] In step S110, a basic model object is generated from a type of raw data to form the initial three-dimensional form of the world model it describes. The basic model object is the starting point of the three-dimensional structure for abstractly representing the target world, used to carry the parameter correction and fusion processing performed in subsequent steps, and thus dynamically obtains the applicable world model as time and environment evolve.

[0055] For the basic model object obtained in step S110, obtain its various types of original data so that it can be used as observation data to extract initial parameters.

[0056] It should be noted that regardless of the type of raw data, it is all scene-level data, while the basic model object is a single-entity object. Therefore, it is necessary to obtain various types of raw data for the basic model object obtained in step S110 based on the mapping of data to single-entity objects.

[0057] Furthermore, the original data to which a single object, i.e. the generated basic model object belongs, is determined is used to obtain various types of original data during the execution of step S120.

[0058] In step S120, various types of raw data are acquired for the generated basic model objects. This is first determined through the mapping of data to individual objects. In an exemplary embodiment, for each generated basic model object, data consistency matching is performed through its own spatial features, semantic features, and structural features to determine the various types of raw data belonging to the basic model object.

[0059] Specifically, the generated base model object possesses spatial, semantic, and structural features, which will serve as the basis for determining the attribution of various types of raw data and for parameter correction. Spatial features describe the position, orientation, and spatial boundary range of the base model object in the 3D coordinate system, including but not limited to center coordinates, bounding box range, voxel distribution, or boundary contour information. These spatial features are matched with spatial location information in the laser point cloud data to determine the spatial association between the laser point cloud data and the base model object. Semantic features describe the category or functional attributes of the base model object, including object type labels, functional category labels, or semantic confidence information. These semantic features are matched with semantic information in the visual data to determine whether the corresponding region in the visual data belongs to the base model object. Structural features describe the geometric topological relationships or morphological features of the base model object, including boundary structures, surface features, connectivity relationships, or morphological patterns. These structural features are matched with geometric structural information in the geometric data to verify structural consistency.

[0060] The generated base model object is spatially matched with laser point cloud data and visual data to determine the laser point cloud data and visual data belonging to the base model object; semantically matched with visual data to determine the visual data belonging to the base model object; and structurally matched with geometric data to determine the geometric data belonging to the base model object.

[0061] In other words, in the specific implementation process, the attribution confidence is calculated based on the degree of spatial overlap, semantic category consistency and structural similarity. When the attribution confidence exceeds the preset threshold, the corresponding original data is determined to belong to the basic model object.

[0062] Therefore, in an exemplary embodiment, the precise mapping of various raw data to individual basic model objects can be achieved through the joint determination of spatial features, semantic features and structural features, providing a reliable data foundation for subsequent linkage correction of geometric parameters and physical functional parameters.

[0063] In addition to joint determination based on features, separate determination of various features can also be performed to determine the attribution of the original data. In an exemplary embodiment, the execution process of spatial consistency matching through spatial features with laser point cloud data and visual data includes: establishing a three-dimensional spatial enclosing region for the basic model object; mapping the three-dimensional spatial enclosing region and each laser point cloud data and visual data in a unified three-dimensional coordinate system; determining whether the data points of the laser point cloud data and visual data fall within the spatial range of the three-dimensional spatial enclosing region; and classifying the data falling within this spatial range as associated data of the basic model object.

[0064] In another exemplary embodiment, semantic features are extracted from the visual data, and initial category labels are assigned to the basic model object. If the visual data and the basic model object are semantically matched, the visual data is used as associated data belonging to the basic model object.

[0065] In another exemplary embodiment, for a structural feature, the mapped structural feature is identified from the geometric data, and its corresponding boundary is determined in the associated laser point cloud data. If the mapped structural feature and boundary match the base model object, then the geometric data and the associated laser point cloud data can be determined to belong to the base model object.

[0066] By matching and mapping various types of raw data to the basic model object, various types of raw data of the basic model object are obtained. Then, initial parameters are extracted from the obtained raw data. At this time, the obtained raw data are used as observation data to correct and optimize the initial geometric assumptions of the basic model object.

[0067] It should be understood that in step S120, the extracted initial parameters are observation parameters. The parameter deviation is calculated by comparing the observation parameters with the initial parameters of the basic model object. Based on the parameter deviation, the geometric parameters and physical function parameters of the basic model object are constrained and corrected. Through parameter updates driven by observation data, the basic model object gradually approaches the real state from the initial assumed state, achieving accurate reconstruction of the corresponding world model.

[0068] Step S120 involves the extraction of various types of raw data, including: extracting geometric structure information and physical information from geometric data; extracting spatial location information and surface information from laser point cloud data; and extracting contour information and semantic information from visual data. Geometric structure information describes the spatial morphology and structural features of an object; its content includes vertex coordinates, boundary contours, surface parameters, topological connections, and dimensional information, which will not be listed here in detail. Physical information describes physical functional characteristics, including mass parameters, stiffness parameters, material properties, force boundary conditions, or functional attribute parameters. Geometric structure information and physical information are obtained through parsing and processing of geometric data, such as structural analysis, attribute field reading, and structural feature calculation.

[0069] For laser point cloud data, preprocessing is performed first, such as noise reduction, filtering, and coordinate standardization. Then, spatial location and surface information are extracted. Specifically, on one hand, the three-dimensional coordinate information of each data point in the laser point cloud data is extracted to determine the mapped spatial distribution range and attitude information, thus constructing the required spatial location information. On the other hand, the laser point cloud data undergoes point cloud normal vector estimation, curvature calculation, and surface fitting to obtain surface information used to describe surface features.

[0070] Visual data is processed and semantically analyzed to extract contour and semantic information, which facilitates geometric reconstruction and provides relevant labels.

[0071] In summary, geometric structure information, physical information, spatial location information, surface information, contour information, and semantic information are integrated to construct a multidimensional initial parameter set, which can be used as observation parameters to provide a data basis for the linkage constraint correction of geometric parameters and physical function parameters in the subsequent step S130.

[0072] In step S130, the basic model object is optimized and corrected by using geometric data, visual data, and laser point cloud data as mutual parameter constraints to obtain the modified geometric parameters and physical function parameters.

[0073] First, it should be noted that geometric parameters refer to the set of parameters used to describe the shape and structure of a basic model object. These parameters determine the morphological and spatial characteristics of the basic model object in 3D space. For example, geometric parameters include size parameters, viewpoint parameters, and rendering parameters. Size parameters describe the spatial scale information of the basic model object, including length, width, height, radius, thickness, volume ratio, or scale factor, determining the object's geometric size and proportional relationships. Size parameters constrain the actual area occupied by the model in 3D space and serve as the basic scale basis for geometric reconstruction. Viewpoint parameters describe the viewing posture or spatial orientation of the basic model object in 3D space, including position coordinates, rotation angle, direction vector, view frustum parameters, or posture matrix, determining the object's spatial orientation and relative relationships within the scene. Viewpoint parameters ensure spatial consistency when mapping from different data sources. Rendering parameters describe the geometric characteristics of the basic model object at the visual presentation level, including surface lighting response parameters, texture mapping parameters, shadow parameters, transparency parameters, or color parameters. Rendering parameters ensure the consistency of the model's appearance under visual data constraints and assist in contour matching and visual consistency verification.

[0074] In the specific implementation of this application, the geometric state expression vector of the basic model object is composed of size parameters, view parameters, and rendering parameters. By constraining and modifying the geometric state expression vector, the consistency optimization of the spatial shape and visual representation of the basic model object is achieved.

[0075] Physical functional parameters describe the physical and functional attributes of the basic model objects, thereby characterizing their physical interaction forms, relationships, and functional roles. For example, physical functional parameters include at least the facade parameters corresponding to physical information, the relationship parameters corresponding to surface information, and the object parameters corresponding to semantic information. The facade parameters, derived from physical information, describe the physical external structural features of the basic model objects, including but not limited to external structural morphology, material type, thickness parameters, load-bearing capacity parameters, stiffness parameters, or mechanical boundary conditions. Facade parameters reflect the structural strength and physical support capacity of the objects in the physical environment and serve as fundamental parameters for physical consistency verification, ensuring the structural rationality and stability of the objects during reconstruction.

[0076] Relationship parameters, derived from surface information, describe the contact, connection, or interaction relationships between the base model object and the environment or other objects. These include, but are not limited to, contact area parameters, surface friction coefficients, connection method parameters, fit parameters, or surface coupling characteristic parameters. Relationship parameters characterize the physical interactions and surface constraints between base model objects, maintaining physical coordination and consistency of action among them during multi-object fusion or dynamic evolution.

[0077] Object parameters, derived from semantic information, describe the functional categories and role attributes of objects in the basic model. These include object category labels, functional role identifiers, behavioral attribute identifiers, interactive attributes, and operable state parameters. Object parameters determine the functional positioning of objects within the overall world model and provide a foundation for subsequent functional constraint corrections and dynamic behavior deductions.

[0078] The correction of geometric and physical function parameters in the basic model object is achieved through the execution of two sub-processes. In the execution of one sub-process, geometric structure information, spatial location information, and contour information are mapped to a unified parameter space for consistency verification and error calculation. The resulting error is then used to control the optimization of dimensions, etc., so that the optimized world model meets the consistency of multiple types of original data in geometric expression. The execution of the other sub-process will match physical information, surface information, and semantic information to obtain updated parameter values, thereby achieving two-dimensional linkage optimization of geometric and physical function parameters, realizing multi-source data collaborative calibration, and avoiding deviations caused by a single data source.

[0079] The geometric parameters and physical function parameters obtained by correcting the basic model object in step S130 are then used to reconstruct the basic model object in step S140 based on the corrected geometric parameters and physical function parameters, so as to obtain a world model that is adapted to the current situation.

[0080] In an exemplary embodiment, the execution process of step S140 includes: applying the modified geometric parameters to the model space structure update, loading the modified physical function parameters to the model attribute layer, and finally performing a consistent verification of each parameter to complete the optimization of the basic model object and output a complete three-dimensional world model.

[0081] Furthermore, the geometric parameters, including their various components, exist in vector form. For example, the constructed vector is a geometric state vector, i.e. Where m takes a value not less than 3, and These are used for size parameters, viewpoint parameters, rendering parameters, etc.

[0082] Correspondingly, for physical function parameters, the various parameters they contain exist in vector form. For example, the constructed vector is a functional state vector, i.e. Where n takes a value not less than 3, and These are used to represent facade parameters, relational parameters, object parameters, etc.

[0083] At this point, the basic model object The complete state can be represented as S Through the obtained corrected geometric parameters Performing spatial structure updates specifically includes: updating the corrected geometric parameters. Based on the corrected size parameters, the geometry and topology of the basic model object are scaled or locally reconstructed, and its 3D vertex coordinates, boundary and surface control point data are updated. Based on the corrected view parameters, the mapped matrix is ​​recalculated to update the spatial positioning relationship, so that the model's posture and orientation in the unified coordinate system remain consistent with the observation data. The corrected rendering parameters are loaded into the graphics rendering engine, and the material properties, texture mapping paths and lighting model parameters are reconfigured to improve the consistency between the model's visual expression and the real scene.

[0084] This update process involves updating the base model object. Reconstruction of the spatial structure layer.

[0085] Correspondingly, in step S140, the process of loading the corrected physical function parameters into the model attribute layer specifically includes: mapping the material type and physical properties in the facade parameters to the model surface elements, and updating their collision parameters, mechanical coefficients and / or environmental interaction coefficients; constructing a spatial relationship graph between objects based on the relationship parameters in the basic model object attribute relationship data, and updating the continuous relationship and / or subordinate hierarchical relationship between models; binding the semantic information object parameters with the identifiers of the basic model objects to establish a semantic index for subsequent retrieval, reasoning and intelligent decision-making.

[0086] This completes the loading of the model attribute layer of the basic model object, which enhances the information and achieves a unified expression of geometric structure and functional attributes.

[0087] Once the structure update and attribute loading of the basic model object are completed, the parameter consistency verification and model optimization output in step S140 can be executed to obtain a world model that is dynamically adapted to the current environment and has high accuracy.

[0088] To further explain, the updated geometric parameters and physical function parameters of the basic model object undergo consistency verification. This process includes consistency verification of dimensional parameters and physical constraints; matching verification of viewpoint parameters and spatial positioning relationships; material consistency verification of rendering parameters and facade parameters; logical consistency verification of surface relationship parameters and geometric topological relationships; and semantic consistency verification of object parameters and object structure types. When all types of parameters meet the preset error thresholds and logical consistency rules, the optimization of the basic model object is completed.

[0089] When the optimized base model object is only a single object, the optimized base model object constitutes a single world model.

[0090] When optimization yields several individual objects, that is, when the optimized objects correspond to several basic model objects of an individual, it is still necessary to establish the objects and the structural relationships between them as well as global constraints to finally obtain the overall world model.

[0091] In an exemplary embodiment, the establishment of structural relationships and global constraints includes, but is not limited to, the execution processes of spatial stitching, topological integration, relationship graph fusion, and semantic network construction. For example, spatial stitching addresses the coordinate system issue for different objects; each basic model object may have its own original, scale, and rotation references, which are transformed to a unified world coordinate system through spatial stitching. Topological integration ensures that, for instance, building walls must connect to roofs, roads must be continuous with the ground, and pipes must be connected to equipment, thus avoiding structural breaks in the constructed "world." Relationship graph fusion resolves the physical or functional dependencies that should exist between objects; for example, doors should connect to walls, lights should be installed on ceilings, and tables should support computers—all achieved through relationship graph fusion. Similarly, this ensures that the output of the world model is not simply a geometric modeling implementation.

[0092] It should be understood that the basic model object used to dynamically generate the world model can have different structural complexities in different application scenarios. In one case, if the basic model object has one and only one single object, and this single object is an indivisible atomic object, such as a table or a simple geometric solid, then the parameter-optimized basic model object itself constitutes the complete world model. In this case, there is no cross-object spatial splicing, topological integration, or relational fusion process; the optimized set of geometric parameters and physical function parameters can be used as the world model output.

[0093] However, in practical applications, even if there is only one monolithic object, it is usually a closed system. For example, a car, a piece of equipment, or a building contains multiple substructures or functional components, and there are structural connections, functional dependencies, and semantic hierarchical relationships between these components. Therefore, the monolithic object can still be structurally decomposed into several substructural units, forming an internal topology and semantic network.

[0094] In such cases, although the basic model objects are individual entities in number, they still need to undergo processes such as spatial structure integration, topological relationship construction, and semantic network organization to ensure that: the geometric connection relationships between substructures are consistent; the physical and functional dependencies between components are reasonable; and the semantic hierarchy is complete and reasonable.

[0095] Therefore, for a single object with an internal structural hierarchy, the basic model object after parameter optimization is not directly output as a world model. Instead, it is necessary to further construct an internal topological structure and semantic network based on its corrected geometric parameters and physical function parameters, so as to form a complete three-dimensional world model expression. The final output complete world model can accurately express spatial structure, computably express physical attributes, and inferably express semantic relationships.

[0096] The world model is a complete three-dimensional structured expression system based on a parameter-optimized basic model object. Under a unified spatial expression framework, it is formed through geometric structure updates, physical functional attribute loading, and structural relationship organization. It can correspond to atomic-level single objects, as well as complex single-object systems or multi-object systems containing internal hierarchical structures.

[0097] Appendix Figure 2 This is a holistic world model from the perspective of an exhibition hall space. Within this overall world model (a comprehensive expression of geometric data, point cloud data, and texture data), there are individual world models that combine movement and stillness, such as doors and windows, walls, robots, and robot dogs. Individual world models of varying sizes and in different scenes are constructed to cater to different types, quantities, and states, forming a... Figure 2 The resulting holistic world model is a novel intelligent space that serves people's understanding, perception, and reasoning, and acts upon the world to support their planning and execution of the future, positively promoting industrial progress, such as innovative applications in scenarios like rail transit, water conservancy, emergency fire fighting, urban governance, smart cities, smart power, and smart factories.

[0098] Based on this, each individual world model in the overall world model starts with a basic model object implemented through different technical paths, and ultimately achieves a description from the "inside out" through the aforementioned parameter optimization process of constraint correction. Specifically, the internal "skeleton" describing the basic model object can be geometric data, such as Building Information Modeling (BIM) or Computer Aided Design (CAD) data, using mathematical methods to describe the geometric structure of the basic model object. The "skin" attached to the surface of the basic model object's "skeleton" can be laser point cloud data, such as LiDAR point cloud data or 3D Gaussian splatting data, using laser point cloud data to describe the geometric surface of the basic model object. Above the basic model object's "skin" is the visual "clothing," which can be visual data, such as image (e.g., png / jpg) or video (e.g., AVI / MP4) data, using visual data to describe the colorful appearance of the model object. Therefore, acquiring geometric data, laser point cloud data, and visual data constitutes the data foundation structure of the model object "from the inside out."

[0099] Geometric data, laser point cloud data, and visual data each encompass a wide variety of data types and categories. Enumerating all data categories would lead to redundancy and a lack of focus in the basic model objects. Conversely, using all data categories to generate basic model objects increases complexity and difficulty, making them unsuitable for rapidly changing scenarios. Therefore, the digital representation of basic model objects must be accurate, lightweight, and easily understood and perceived.

[0100] Therefore, a basic model object is initially generated using only one type of data, and then optimized using various other types of data to achieve a hierarchical description "from the inside out," with each type of data being both interconnected and independent. Furthermore, for the structure of the basic model object, such as dimensions / positions like length, width, height, thickness, and orientation, and shapes / contours like circles, rectangles, and curved surfaces, the structural description of these basic model objects primarily uses mathematical methods to construct geometric parameters for precise quantitative expression. For the functions of the basic model object, such as material, texture, color, object affiliation, and attribute classification, physical functional parameters are primarily expressed through text. Thus, it is evident that, under the influence of geometric and physical functional parameters, various intelligent applications can achieve a "rational understanding" of the basic model object and the final output world model.

[0101] Furthermore, describing the basic model object through geometric parameters involves expressing continuous mathematical quantities through mathematical functions in the underlying code or programming language, which is precise and requires less code. Describing the physical function of the model object through text involves expressing textual semantics in byte form in the underlying code, which can directly call third-party large language models to achieve perception, understanding, and reasoning of the model. The results of perception, understanding, and inference can be fed back to the basic model object expressed through geometric parameters, thereby building an intelligent agent that plans and executes autonomously based on the basic model object, rather than simply replicating animations.

[0102] The following table illustrates the data architecture of the basic model object using a specific application example:

[0103]

[0104] Combined with appendix Figure 3 The world model shown is based on a small indoor space scene in a laboratory. It constructs individual world models, including walls, doors and windows, floors, a robot dog, a robot, and partitions, through the generation and optimization of basic model objects. Further illustrating this, BIM technology is used to acquire the geometric and physical information of the basic model objects. This includes geometric entities (such as points, lines, surfaces, and volumes) and parametric geometry (such as constraint equations), mathematically describing the geometric structure of the "skeleton" of the basic model objects. Additionally, the material of the basic model objects can be physical information such as rebar (e.g., HRB400 / 500) or cement (e.g., 425 / 525). Correspondingly, laser point cloud technology can acquire the spatial location and surface information of the basic model objects, comprehensively describing the surface of the model objects (similar to "skin"). This includes spatial location information such as coordinates (e.g., point coordinates XYZ) and dimensional scales (e.g., distances between points in space); and surface information such as reflection intensity (the strength of light energy reflected from the object's surface back to the sensor) and attribute classification (e.g., high reflectivity indicates metal, low reflectivity indicates plastic, and medium reflectivity indicates walls). Images acquired through a visual camera (such as JPG or PNG format images) can provide contour and semantic information of the basic model object, comprehensively describing the object's appearance (similar to "clothes"). This includes contour information such as the contour points (e.g., convex / concave points projected onto a 2D plane, represented by coordinates) and contour area ratio (region size, which can constrain model volume or surface coverage); as well as the rendered color of the basic model object (e.g., red, yellow, etc.) and the semantic information describing the model object (table, chair, etc.).

[0105] In the application examples described above, a basic model object corresponding to a specific object can be quickly and directly generated based on any one of the three types of raw data: geometric data, laser point cloud data, or visual data. Specifically, geometric data can directly generate a white model of the BIM structure, and the resulting basic model object is a three-dimensional structure composed of lines. At this point, laser point cloud data can be used to construct the surface of the basic model object, that is, to cover the obtained white model with a laser point "surface". Visual data renders different colors on the "surface" of the white model, or directly attaches the obtained image to the "surface". Since geometric data, laser point cloud data, and visual data are collected independently, when a basic model object is initially generated based on one type of data, the other two types of data will be nested on this basis and integrated into one. Because these data are independent of each other, the basic model objects describing various data are also very different and have deviations between them. For example, if the size of the white model of the geometric structure is too large or too small, the laser point cloud data will completely or partially cover the white model, or the rendered area or color will also have errors. It is like the "skeleton" of a "person" does not match the "skin surface" and is wearing an ill-fitting "clothes", which makes the final world model distorted.

[0106] Based on this, we will use an application example model object for illustration, combined with the appendix. Figure 4 This diagram illustrates the joint constraints based on geometric data, laser point cloud data, and visual data. Specifically, it fully utilizes geometric data, laser point cloud data, and visual data to extract mathematical and textual dimensions. The collected data is shared and reused, that is, geometric data, laser point cloud data, and visual data are shared and reused to construct mutually constraining parameters. The "skeleton, skin, and clothing" of the model object are adapted to each other, "tailor-made and well-fitting," thereby outputting a high-precision 3D world model.

[0107] See Figure 5 , Figure 5 It is based on Figure 1 The flowchart shown in the corresponding embodiment describes the steps of constraining and correcting the initial geometric parameters of the basic model object by linking geometric structure information, spatial position information, and contour information.

[0108] The basic model object mapping can describe the initial geometric parameters of the initial three-dimensional form of the world model. These geometric parameters include size parameters, viewpoint parameters, and rendering parameters. In the embodiment of this application, step S130 involves the execution process of constraining and correcting the initial geometric parameters of the basic model object by linking geometric structure information, spatial position information, and contour information.

[0109] Step S131a: Using the spatial coordinates and size scale in the spatial location information, perform consistency verification of the two-dimensional edge and shape dimensions adapted to the contour information. Apply the size error obtained from the verification to optimize the size parameters of the basic model object until the updated size parameters are obtained by convergence under the size error control.

[0110] Step S132a: Extract the actual contour from the contour information of the visual data, including the contour convexity and concave points and the contour area ratio. Under the initial viewpoint parameters corresponding to the current spatial position information, project the corresponding three-dimensional model structure to generate the predicted contour. Compare the actual contour and the predicted contour. In the comparison, combine the vanishing point position of the geometric structure information and the combination of visible surfaces to calculate the viewpoint error. Optimize the viewpoint parameters mapped by the spatial position information based on the viewpoint error.

[0111] Step S133a: Render the basic model object according to the current rendering parameters to generate a predicted outline. Calculate the rendering consistency error by comparing the predicted outline with the actual outline extracted from the visual data. Adjust the rendering parameters according to the rendering consistency error.

[0112] Step S134a: The mutually constrained size parameters, view parameters, and rendering parameters form the constrainable geometric parameters of the basic model object.

[0113] These steps are explained in detail below.

[0114] In this embodiment, the base model object is mapped to obtain initial geometric parameters that describe the initial three-dimensional shape of the world model. These initial geometric parameters include: size parameters, viewpoint parameters, and rendering parameters.

[0115] Step S130 executes multi-source constraints on the initial geometric parameters of the basic model object mapping on the one hand, and multi-source constraints on the physical function parameters of the basic model object mapping on the other hand, thereby ultimately improving the fitting accuracy of the basic model object to the real observation data, avoiding parameter compensation and structural distortion, and obtaining a globally optimal solution that is consistent with geometry and vision under joint constraints.

[0116] In step S131a, consistency constraint correction of dimensional parameters is performed so that the three-dimensional structural scale corresponding to the basic model object can be consistent with the actual observed contour.

[0117] On the one hand, based on the spatial coordinates and size scale in the spatial location information, combined with model parameters, i.e., the current size parameters and the viewpoint parameters, the predicted two-dimensional contour size is obtained through projection calculation; on the other hand, the actual two-dimensional contour shape size is obtained by extracting the contour information from the visual data. The consistency between the predicted two-dimensional contour size and the actual two-dimensional contour shape size is verified, and the size error between the two is calculated. Then, the size parameters of the basic model object are adjusted in reverse according to the obtained size error.

[0118] The size parameters of the base model object are iteratively optimized under the control of the error threshold until the size error converges to the prediction error range, and finally the updated size parameters are obtained.

[0119] By executing step S131a, the optimized three-dimensional structure scale can be kept consistent with the actual observation. Through parameter optimization in the dimension of size parameters, the final model can approximate the actual observed state.

[0120] In this embodiment, the optimization of geometric parameters includes not only constraint correction of dimensional parameters, but also optimization of other geometric parameter dimensions such as viewpoint parameters and rendering parameters under the action of steps S132a and S133a. It should be clear that steps S131a, S132a, and S133a in this embodiment implement constraint correction for parameters in different dimensions of the geometric parameters. These steps are not isolated from each other, but rather collaboratively optimize different dimensions of the geometric parameters within a unified error constraint framework.

[0121] In the specific implementation process, steps S131a, S132a and S133a can be executed concurrently, sequentially or selectively according to the execution requirements. For example, the size parameters can be optimized first and then the viewpoint parameters can be corrected; multiple parameter dimensions can be iteratively updated simultaneously under the constraint of a unified error function; or the optimization process of a certain dimension can be stopped when a certain parameter has met the error threshold condition.

[0122] Therefore, this application does not limit the execution order or combination of each step, as long as it can achieve the constraint correction of geometric parameters and improve the consistency between the basic model object and the real observation data.

[0123] In summary, the embodiments of this application achieve coordinated convergence of multi-dimensional parameters under a unified error objective by decomposing geometric parameters into different dimensions and establishing corresponding constraint correction mechanisms, thereby avoiding the error propagation problem caused by single parameter optimization and improving the stability and accuracy of geometric model construction.

[0124] In step S132a, the viewpoint parameters in the geometric parameters are constrained and corrected. In an exemplary embodiment, after the size parameters are updated, the video parameters are further constrained and corrected, specifically including: on the one hand, extracting contour convex and concave points and contour area ratios from the contour information of the visual data to construct the actual contour; on the other hand, under the initial viewpoint parameters corresponding to the current spatial position information, projecting the three-dimensional model structure of the generated basic model object to generate a predicted contour, thereby comparing the predicted contour with the actual contour and calculating the contour overlap and edge offset of the two.

[0125] The overlap and edge offset between the predicted and actual contours reflect the projection deviation between them. The vanishing point position and visible surface combination in the geometric information provide perspective geometric constraints. Therefore, the calculated overlap and edge offset are combined with the vanishing point position and visible surface combination in the geometric information to stably solve the viewpoint error. Finally, the viewpoint parameters mapped by the spatial position information are optimized and updated based on the viewpoint error, so that the predicted contour obtained by projection gradually approaches the actual contour.

[0126] In step S133a, rendering parameter consistency optimization is performed. In an exemplary embodiment, after the size parameters and viewpoint parameters are corrected, the rendering parameters are optimized, specifically including: rendering the basic model object based on the current rendering parameters to generate a predicted contour, comparing the predicted contour with the actual contour extracted from the visual data to calculate the rendering consistency error, and finally adjusting the rendering parameters based on the rendering consistency error.

[0127] For example, the calculated rendering consistency error includes differences in edge sharpness, contour contrast, and occlusion consistency. Adjustments to rendering parameters include corrections to light intensity, material reflectivity, and texture mapping, ensuring that the final rendering result is consistent with visual observation.

[0128] After the above-mentioned size parameters, viewpoint parameters, and rendering parameters are all corrected, the geometric parameters of the basic model object are generated through step S134a to optimize the parameters.

[0129] In step S134a, the geometric parameters for optimizing the basic model object parameters are formed through mutually constrained size parameters, viewpoint parameters, and rendering parameters. Furthermore, since there is a mutual coupling relationship among the size parameters, viewpoint parameters, and rendering parameters—for example, size parameters affect the projected contour, viewpoint parameters affect the contour shape, and rendering parameters affect the contour edge expression—the optimized size parameters, viewpoint parameters, and rendering parameters can ultimately form the constrainable geometric parameters of the basic model object through the implementation of inter-factor constraints among them.

[0130] Specifically, the interconnected constraints among the optimized size parameters, viewpoint parameters, and rendering parameters can be achieved through the current mechanism. That is, any correction to any of the three parameters triggers the error recalculation of the other parameters. Based on the calculated error, the joint optimization adjustment of the three parameters is constrained by constructing an overall error function until the obtained joint error converges to a preset threshold.

[0131] Thus, the final geometric parameters obtained are consistent with each other in terms of size parameters, viewpoint parameters, and rendering parameters, and match in terms of geometric structure information, spatial position information, and visual contour information. These are the constrainable geometric parameters of the basic model object.

[0132] By constraining and correcting the geometric parameters, including size parameters, viewpoint parameters, and rendering parameters, the accumulation of errors caused by independent optimization of a single parameter is avoided. This significantly improves the consistency between the constructed 3D structure and the 2D observation data, providing a precise geometric foundation for subsequent loading of physical function parameters and construction of the world model.

[0133] See Figure 6 , Figure 6 It is based on Figure 1 The corresponding embodiment shows a flowchart describing the steps of constraining the physical functional parameters of a basic model object by linking physical information, surface information, and semantic information.

[0134] Physical functional parameters include facade parameters of physical information, relational parameters of surface information, and object parameters of semantic information;

[0135] The execution process of constraining the physical function parameters of the basic model object by linking physical information, surface information, and semantic information in step S130 of this application embodiment includes:

[0136] Step S131b involves checking geometric consistency and semantic consistency by using the actual normal vector and actual surface curvature in the surface information, and the actual texture and color in the semantic information, respectively, constructing the facade constraint error, and then minimizing the facade constraint error by using the initial value of the facade parameters as a starting point to obtain the updated value of the facade parameters in the physical information.

[0137] Step S132b: The semantic relationship constraint set mapped by the object description and spatial relationship of the semantic information is matched with the current geometric relationship state of the patch topology and size ratio mapping of the physical information and the error is calculated. The updated value of the relationship parameter in the surface information is obtained by iterative optimization of minimizing the obtained relationship constraint error.

[0138] Step S133b: Construct physical feature vectors by classifying the volume and material of physical information, the inverse strength and attribute classification of surface information, and compare the mapped semantic categories with the object parameters in the semantic information by mapping the physical feature vectors to semantic categories and calculate the semantic consistency error. Minimize the semantic consistency error and perform iterative optimization to obtain the updated values ​​of the object parameters in the semantic information.

[0139] Step S134b: The mutually constrained facade parameters, relationship parameters, and object parameters are combined to form constrainable physical function parameters.

[0140] These steps are explained in detail below.

[0141] In this embodiment, physical functional parameters are used to describe the basic model object at both the physical structure and functional expression levels. Physical functional parameters include at least facade parameters in physical information, relational parameters in surface information, and object parameters in semantic information. The facade parameters correspond to the geometric morphology layer of the basic model object, the relational parameters correspond to the spatial relational layer, and the object parameters correspond to the semantic expression layer. These three parameters are interconnected through a constraint mechanism, collectively forming a computable, optimizable, and interpretable physical functional parameter system.

[0142] The facade parameters are used to describe the visible geometric features of the basic model objects and characterize the external geometric structure features in the physical information; the relationship parameters are used to describe the spatial topological relationship and size ratio relationship between different individual units, such as the spatial structure organization in the surface information; the object parameters are used to describe the object type, functional attribute identifier, material semantic category and surface attribute classification in the semantic information, so as to express the functional belonging of the model at the semantic level.

[0143] In step S131b, geometric consistency checks and semantic consistency checks are performed to construct facade constraint errors based on the obtained geometric consistency errors and semantic consistency errors. Starting from the initial values ​​of the facade parameters, the facade constraint errors are minimized iteratively to obtain the updated facade parameters.

[0144] In an exemplary embodiment, the specific execution process for geometric consistency checking includes: obtaining the actual normal vector and actual surface curvature from the surface information; calculating the angle deviation between the actual normal vector and the current model facade normal vector corresponding to the base model object; calculating the distribution difference between the actual surface curvature and the current model curvature; and finally obtaining the geometric consistency error through the obtained deviation using the geometric consistency error function.

[0145] For example, the geometric consistency error function can be in the form of the following formula:

[0146]

[0147] in, It is a geometric consistency error; It is the normal vector of the current model's exterior facade. It is the actual normal vector. It is a weighting factor that can be adjusted as needed. It is the curvature of the current model. It is the actual surface curvature. The index represents the discrete sampling point or patch element of the facade surface, and j represents the index of the discrete surface element or local surface segment used for curvature calculation.

[0148] Under the influence of the geometric consistency error function, the differences in the normal vectors of the discrete sampling units of the facade are squared and accumulated, and the differences in the curvature of the local surface units are weighted and accumulated, thereby constructing the geometric consistency error.

[0149] In an exemplary embodiment, the specific execution process for semantic consistency checking includes: performing a consistency comparison on the semantics mapped to the current model facade based on the actual texture and actual color in the semantic information, such as determining whether the material texture matches, whether the color distribution conforms to the semantic category, and whether the component semantics are consistent with the facade partition, so as to obtain the degree of difference at the semantic level, and then comprehensively obtain the semantic consistency error.

[0150] In step S132b, based on the object description and spatial relationship description in the semantic information, such as inclusion, adjacency, symmetry, and hierarchical structure, a set of semantic relationship constraints is constructed. In the constructed set of semantic relationship constraints, each semantic relationship is mapped to a computable geometric constraint. For example, a window located inside a wall is mapped to a patch inclusion constraint, a symmetrical distribution of columns is mapped to an object axis distribution, and a roof covering the main body is mapped to a coverage projection constraint.

[0151] The current geometric relationship state of the basic model object is obtained to extract the topology and size ratio of the facets from the physical information to construct the current geometric relationship set; the current geometric relationship set is matched item by item with the semantic relationship constraint set, and the relationship constraint error is calculated.

[0152] The optimization solution minimizes the relation constraint error. During the optimization process, topological continuity is maintained to prevent geometric structure breakage and ensure that the size ratio does not produce non-physical distortion. Finally, the relation parameters are updated after the optimization is completed.

[0153] To further explain, the topological structure before and after optimization will be locked by the adjacency constraints in the semantic relation constraint set, and the boundary continuity constraints will lock the boundary of the two faces sharing the edge so as not to split, thereby maintaining topological continuity, so that the adjacency relationship between faces does not break and the boundary sharing relationship remains consistent.

[0154] The specific implementation process to prevent geometric structure breakage includes: optimizing and updating the local transformation matrix of the parameter update mapping, monitoring the determinant sign, and limiting it to keep it positive in order to avoid patch inversion or structural collapse.

[0155] It should be further explained that during the update of the facade parameters and relational parameters, the model vertex coordinates will generate deformation mapping according to the parameter changes. This deformation mapping can be approximated as a linear transformation in the local area, and its corresponding local transformation matrix is ​​used to describe the degree of local geometric deformation.

[0156] In addition, during relation parameter optimization, proportional constraints, scale variation range limitations, and global volume conservation constraints will be used to ensure that the size proportions do not produce non-physical distortions. Specifically, for semantic relations in the semantic relation constraint set, such as the window height being 1 / 3 of the wall height, the proportional preservation term is established as shown in the following formula:

[0157]

[0158] in, It is the actual geometric height of the window, which can be obtained from the facade parameters in the physical information. It represents the actual size of the window object in the vertical direction in the model under the current optimization state. It is the actual geometric height of the wall on which the window is located, which can be obtained through the exterior parameters of the wall object. It represents the overall height of the window in the vertical direction to which it is attached in the current optimized state. It is the ratio of the window height to the wall height as presented by the current geometric structure; It is the target proportion value corresponding to the semantic constraint.

[0159] Therefore, the overall error term can be calculated by the squared deviation between the current model ratio and the semantic target ratio. .

[0160] Maintaining dimensional proportions is not limited to height. By analogy, it can be seen that the proportionality maintenance term controls the proportional constraint between any two related geometric dimensional parameters. That is, the proportionality maintenance term can be generalized as follows:

[0161]

[0162] in, It is the first geometric dimension parameter. It is the second dimension geometric parameter. It is the target value for the proportion.

[0163] In step S133b, it should first be noted that the constructed physical feature vector refers to a feature vector formed by volume parameters, material parameters, reflection intensity, and attribute classification encoding. The physical feature vector is matched to the degree of matching in each semantic category to obtain the corresponding semantic consistency error, which is then used to update the object parameters in the semantic information.

[0164] Using semantic categories as the semantic space, we can perform mapping and matching of physical feature vectors to the semantic space, thereby realizing the semantic consistency error calculation of physical feature vectors to object parameters in semantic information.

[0165] The physical feature vector is input into the semantic space, and the matching value of the physical feature vector in each semantic category is calculated to form a semantic matching vector. Each component in the semantic matching vector represents the degree of matching between the corresponding semantic category and the current physical feature. The semantic category with the highest matching degree is compared with the object parameters in the semantic information to construct the semantic consistency error. The semantic consistency error is minimized and iteratively optimized to output the updated object parameters.

[0166] In step S134b, the updated facade parameters, relational parameters, and object parameters are integrated to obtain physical function parameters for subsequent modeling.

[0167] For the implemented parameter optimization, the obtained geometric and physical function parameters can be directly loaded and updated to obtain the corresponding world model. In addition, during the world model reconstruction, parameters that have deviated under the unified framework will be updated, as well as the geometric structure and derived attribute data will be updated accordingly.

[0168] In summary, the "triangular interlocking" constraint design among parameters means that the mutually constrained facade parameters, relational parameters, and object parameters form constrainable physical functional parameters. These parameters are constrained by geometric boundaries, spatial coupling relationships, and physical properties, and then the three together constrain feasible functions in the physical environment, providing multifaceted support for the recognition and reading of large language models.

[0169] Therefore, by combining these three types of parameters to create constraints, and with the support of a large language model, we can construct derivable, understandable, and interactive model objects. Thus, based on physical function parameters, we can reconstruct the basic model objects to obtain a perceptive world model.

[0170] In other embodiments, the underlying model object is reconstructed by combining geometric parameters and physical function parameters, integrating the internal structure with the external expression, and generating an accurate and perceptive world model.

[0171] Figure 7 It is based on Figure 1 The flowchart described in the corresponding embodiment describes the steps of reconstructing a world model from a basic model object based on geometric parameters and physical function parameters.

[0172] The step S140 of this application embodiment, which reconstructs a world model from a basic model object based on geometric parameters and physical function parameters, includes:

[0173] Step S141a: Align the geometric parameters and physical function parameters to their respective parameter dimensions, compare the parameters corresponding to the basic model object, and calculate the deviation;

[0174] Step S142a: For each parameter dimension, determine whether the obtained deviation value exceeds the preset deviation threshold. If it exceeds the deviation threshold, mark the corresponding parameter dimension in the basic model object as the dimension to be updated.

[0175] Step S143a: For the dimension to be updated, perform weighted update or direct replacement update based on the corresponding deviation value and preset update step size to update the corresponding current parameters of the basic model object;

[0176] Step S144a: After completing the parameter update, perform structural consistency synchronization on the base model object to generate the reconstructed world model. Structural consistency synchronization includes recalculating the geometric structure and derived attribute data of the base model object with the updated parameters.

[0177] These steps are explained in detail below.

[0178] In this exemplary embodiment, step S140 is used to reconstruct the basic model object in a consistent manner based on the optimized geometric parameters and physical function parameters, so as to generate a world model that satisfies multidimensional constraints. The reconstruction process is not a simple parameter replacement, but also includes parameter dimension alignment, deviation determination, selective updating, and structural synchronization, thereby systematically completing the model update process.

[0179] First, in step S141a, the geometric parameters and physical function parameters are aligned according to a preset parameter dimension system. This alignment maps parameters from different sources to the same dimensional space for comparison, thereby determining the parameter dimensions that the base model object currently needs to update, facilitating selective updates to the base model object.

[0180] Align the geometric parameters and physical function parameters obtained by optimization in step S141a with the corresponding parameters of the basic model object, and calculate the deviation between the corresponding parameters.

[0181] For the obtained deviations, a comparison with the corresponding deviation threshold is performed in step S142a. If the deviation value of a parameter dimension exceeds the corresponding deviation threshold, this parameter dimension is marked as a dimension to be updated. By marking, repeated updates to parameters that have already met the accuracy requirements are avoided, thereby improving reconstruction efficiency and reducing unnecessary structural disturbances.

[0182] Subsequently, in step S143a, parameter update operations are performed on the dimensions marked as to be updated. The update methods include weighted updates and direct replacement updates. Weighted updates involve progressively correcting the current parameters based on the magnitude of the deviation and the predicted update step size, causing the parameters to gradually converge towards the target value. Direct replacement updates involve directly replacing the current parameters with the target parameter value when the deviation significantly exceeds a threshold or a structural error exists. During the update process, different weights can be assigned according to the importance of the dimensions to ensure that key structural dimensions converge first, while avoiding over-updating that could cause structural oscillations.

[0183] Finally, in step S144a, after the parameter update is completed, a structural consistency synchronization operation is performed on the base model object to generate the reconstructed world model. Structural consistency synchronization includes: recalculating vertex coordinates and spatial positions based on the updated parameters, refreshing patch boundaries and topological adjacency relationships, recalculating normal vectors and curvature information, and recalculating attribute data derived from the geometric structure. This synchronization mechanism ensures that the updated model maintains consistency between the geometric structure and physical functionality levels, and avoids global structural imbalances caused by local updates.

[0184] In summary, through the coordinated execution of steps S141a to S144a, selective updating and synchronous structural reconstruction under a multi-dimensional parameter alignment mechanism are achieved, thereby generating a world model that meets the requirements of geometric accuracy, physical rationality, and semantic consistency. The basic model object is a model structure generated for a specific single object. Therefore, the world model obtained based on this basic model object through parameter optimization and structural reconstruction belongs to the single-object level, i.e., a single-object world model. A single-object world model only describes the complete expression of the corresponding single object at the levels of geometric structure, physical function, and semantic attributes.

[0185] For scenarios containing multiple objects, it is necessary to combine and collaboratively construct multiple individual world models. By aligning spatial relationships, unifying global coordinates, establishing topological relationships between objects, and synchronizing interaction constraints, multiple individual world models are integrated into a scenario-level overall world model, thereby forming an overall world model with global consistency and structural relevance.

[0186] See Figure 8 , attached Figure 8 The image shows a basic model object generated based on geometric data, laser point cloud data, or visual data. This basic model object is obtained through different technical paths, such as BIM 3D modeling, point cloud model reconstruction, or visual image inference model generation. However, this basic model object only reflects the 3D model form of a single type of original data and generally lacks constraint parameters to describe structural boundaries, object attributes, and spatial relationships, resulting in insufficient accuracy and semantic consistency.

[0187] Therefore, in conjunction with the appendix Figure 9 This is a rendering of the mutual constraints of walls (partially) in space. To improve the accuracy and usability of the resulting world model, geometric parameters and physical function parameters are introduced to constrain the basic model objects, thereby correcting the boundary shape, object attributes and physical behavior of the model, and thus obtaining a more accurate, understandable, perceptible and high-quality model object that can be used for subsequent reasoning calculations and analysis.

[0188] Furthermore, based on the geometric and physical function parameters, each parameter is compared with the corresponding parameter in the basic model object to detect the degree of difference. If the deviation of a parameter exceeds a preset threshold (e.g., the deviation exceeds 10%), the original parameter of the basic model object is corrected, and the corrected parameter is written back. The resulting new model object, i.e., the world model, receives dynamic updates and improved accuracy. If the deviation of a parameter is within the preset threshold range, the original parameter remains unchanged.

[0189] In other embodiments, the data of the basic model object is updated according to different scenarios. Specifically, depending on the ownership of the basic model object, if it is a long-term static model object, such as buildings, roads, bridges, etc., which are static and do not change frequently, the threshold for parameter deviation can be set to a small value (e.g., deviation exceeding 5%) before updating and writing back to the original model object. If it is a short-term dynamic model object, such as pedestrians, vehicles, etc., which change frequently due to high frequency or rapid movement, a dual-range threshold of time and deviation is set. For example, if the deviation value is between 90% and 100% and the time is between 10 and 30 seconds, it is determined that the model object is moving, and the corresponding parameters remain unchanged. It should be understood that the preset parameter deviation threshold and the preset time threshold can be determined by using big data analysis, data analysis statistics, or cluster analysis techniques to find the defined ranges, providing a reference for the selection of the threshold.

[0190] See Figure 10 , Figure 10 It is based on Figure 4 The flowchart described in the corresponding embodiment describes the steps of reconstructing a world model from a basic model object based on geometric parameters and physical function parameters.

[0191] The world model generated by reconstructing the basic model object is a single-entity world model, such as... Figure 7 The world model generated in the corresponding embodiment, and the basic model object provided in this application embodiment, further include step S140 of reconstructing the world model based on geometric parameters and physical function parameters to obtain the world model:

[0192] Step S141b: Obtain attribute relationship data between basic model objects. The attribute relationship data includes spatial relationship parameters and subordinate hierarchy relationship parameters between basic model objects.

[0193] Step S142b: Construct an object spatial relationship graph based on the spatial relationship parameters, and obtain the global pose transformation matrix of each individual world model by solving the global pose in a unified coordinate system based on the spatial relationship parameters. Perform pose transformation of the corresponding individual world model through the global pose transformation matrix.

[0194] Step S143b: Construct a hierarchical structure tree for the subordinate hierarchical relationship parameters. In the hierarchical structure tree, the pose transformation of the child nodes is recursively synchronized to the child nodes according to the hierarchical topology order, with the parent node of the pose transformation as the reference. The pose transformation of the child nodes is superimposed with their own local correction matrix. The local correction matrix is ​​the incremental pose matrix generated during the construction of the spatial relationship graph, the optimization of geometric consistency constraints, or the correction of physical function constraints.

[0195] Step S144b: Based on the spatial relationship diagram and the hierarchical structure tree of the individual model, conflict detection is performed on the individual world model in the spatial, hierarchical and functional dimensions. The detected conflict items are corrected and processed, and the overall world model that satisfies spatial consistency, hierarchical consistency and functional consistency is generated.

[0196] These steps are explained in detail below.

[0197] This exemplary embodiment enables the construction process from a single world model to a global world model, applicable to scenarios containing multiple basic model objects, so that the final generated world model maintains global consistency and synergy in spatial, hierarchical, and functional dimensions.

[0198] In step S141b, spatial relationship parameters describe the global spatial relationships between individual objects (i.e., the generated basic model objects), such as position, orientation, and relative spacing; subordinate hierarchy parameters represent the parent-child hierarchy and dependencies of individual objects in the scene. For example, in a vehicle model, the wheels are child objects of the vehicle body, and the tabletop is the parent object of the table legs. By obtaining the above attribute relationship data, the foundation for the association between individual world models can be established, providing structural support for the construction of the overall world model.

[0199] In step S142b, the obtained spatial relationship parameters are used to construct a spatial relationship graph of the model objects. In the spatial relationship graph, each node corresponds to a single world model, and the edges represent the spatial constraint relationships between objects.

[0200] In addition, a global solution for spatial relationships will be performed under a unified coordinate system, calculating the global pose transformation matrix for each individual world model. The global pose transformation matrix is ​​used to map the individual world model from its local model space to the scene's global space, enabling adjustments to the position and orientation of the individual world model under the unified coordinate system. This step ensures that the spatial layout of individual objects within the scene is reasonable and conforms to the overall spatial constraints of the scene.

[0201] In step S143b, a hierarchical structure tree for the individual model is constructed based on the hierarchical relationship parameters. Within this structure tree, the pose transformations of parent nodes are recursively synchronized to child nodes in hierarchical topological order, ensuring that child nodes move and rotate correctly with the transformations of their parent nodes. The final pose transformation of each child node is the superposition of the parent node's transformation and its own local correction matrix. This local correction matrix is ​​an incremental pose matrix generated during the construction of the spatial relationship graph, optimization of geometric consistency constraints, or correction of physical functional constraints, reflecting the local changes brought about by geometric adjustments or functional constraint corrections within the individual object. Through the recursive synchronization of the parent node's hierarchy and the superposition of local corrections, the consistency and continuity of the hierarchical relationships and object structure in the global space are ensured.

[0202] After completing the global pose calculation and hierarchical synchronization through the aforementioned steps, step S144b performs conflict detection on each individual world model in spatial, hierarchical, and functional dimensions based on the spatial relationship graph and the individual model hierarchical structure tree. This includes detecting positional overlap, topological conflicts, parent-child hierarchical inconsistencies, or functional parameter conflicts. Detected conflicts are corrected, for example, by fine-tuning the pose, adjusting the local topology, or recalculating functional parameters to eliminate contradictions. Subsequently, the individual world models are merged to generate a holistic world model that satisfies spatial consistency, hierarchical consistency, and functional consistency. This step ensures that the multiple individual world models in the entire scene are coordinated and unified in terms of spatial layout, hierarchical structure, and functional performance, thereby achieving accurate construction of the overall world model.

[0203] In summary, this exemplary embodiment rationally combines and synchronously updates individual world models across spatial, hierarchical, and functional dimensions, achieving coordinated unity of multiple individual world models within the overall scene. This not only ensures the local accuracy and global consistency of individual world models but also, through recursive synchronization of hierarchical structures and superposition of local correction matrices, accurately maintains complex parent-child relationships and topological structures within the scene, thereby improving the spatial rationality, hierarchical integrity, and functional reliability of the overall world model. A scene's basic model objects include not only their geometric structure and physical functional state—expressions focusing on the "points" of the basic model objects—but also descriptions of the "spatial relationships" between these objects—expressions of the overall world model of the "surfaces" formed by multiple basic model objects. In other words, besides achieving the "self-shaping" (including shape and function) of model objects through geometric and physical functional parameters, it is also necessary to express the global "surface" world model constructed between basic model objects through spatial relationships.

[0204] Based on this, in order to intuitively understand the overall world model "from the surface to the core", it is necessary to further utilize the spatial relationships of the overall world model for constraints and expression, as shown in steps S141b to S144b above.

[0205] The basic model objects are constrained by geometric and physical parameters, that is, individual "points" of the basic model objects. The objects constrained by spatial relationships are the relationships formed between multiple basic model objects, that is, the "surfaces" formed by multiple basic model objects. In other embodiments, the basic model objects are expressed at different stages at different times through the constraints of geometric and / or physical parameters, without affecting the spatial display relationships of the "surfaces" formed between the basic model objects.

[0206]

[0207] As shown in the table above, the spatial and subordinate relationships between basic model objects are as follows: Spatial relationships describe the relative position, orientation, distance, and topological relationships of basic model objects in three-dimensional space, such as a zebra crossing covering a road, a road connecting to a bridge, a river under a bridge, and a bridge 20 meters above the river surface. This describes the relationship between multiple basic model objects in three-dimensional space. Subordinate relationships describe the hierarchical relationship between basic model objects in logical and functional structures, such as component / whole (e.g., a door belongs to a wall), functional subordination (e.g., a light fixture belongs to the room's scene brightness), etc. Subordinate relationships determine the baseline of a given basic model object and its hierarchical structure in spatial transformation and display.

[0208] Therefore, in conjunction with the appendix Figure 11 (Appendix) Figure 11The diagram illustrates the spatial relationships and hierarchical relationships of a world model in a given scenario. The spatial relationships of basic model objects are analyzed, including location and orientation. Further, hierarchical data is analyzed to form hierarchical relationships, such as a house including floors and exterior walls, and floors including walls, beams, columns, and floor slabs. Spatial relationship constraints for basic model objects are constructed, such as the relative spatial orientation, location, distance, and topology between walls, beams, and columns, which are then converted into mathematical constraints. These mathematical constraints can be expressed in various ways, such as the constraint that basic model object a is adjacent to basic model object b, dist(surfacea, surfaceb)=0; the parallel orientation constraint that basic model object a is parallel to basic model object b, Na·Nb=1; and the distance between basic model object a and basic model object b being 5 meters, ||Pa||. Pb‖=5m, etc. Establish spatial hierarchical constraints for basic model objects by creating coordinate representations of parent-child hierarchical relationships. Based on spatial and hierarchical constraints, constrain the spatial display of basic model objects. Simultaneously, combine the basic model objects and call the rendering engine to generate a consistent and visual world model according to the constraints.

[0209] See Figure 12 , Figure 12 It is based on Figure 4 The flowchart described in the corresponding embodiment describes the steps of reconstructing a world model from a basic model object based on geometric parameters and physical function parameters.

[0210] The step S140 of reconstructing the world model based on geometric parameters and physical function parameters according to the basic model object provided in this application embodiment further includes:

[0211] In step S141c, memory relationship data between basic model objects is obtained. The memory relationship data includes short-term memory and long-term memory. Short-term memory represents the state information of the basic model objects at the most recent moment, and long-term memory represents the stable state information of the basic model objects within the historical time range.

[0212] In step S142c, for each basic model object, a prediction state matrix is ​​generated by weighted fusion of short-term memory and long-term memory. The prediction state matrix includes pose and object parameters.

[0213] In step S143c, based on the predicted state matrix, object-by-object matching of the current pose and parameters is performed on the individual world model corresponding to the overall world model, and the difference matrix of the individual world model relative to the current predicted state is calculated.

[0214] In step S144c, the difference matrix is ​​combined with spatial consistency constraints, hierarchical constraints and functional constraints to solve for the first optimal correction matrix;

[0215] In step S145c, the first optimal correction matrix is ​​applied to the corresponding individual world model to perform a per-individual correction update of the individual world model in the overall world model.

[0216] These steps are explained in detail below.

[0217] To further enhance the dynamic consistency and historical continuity of the overall world model, step S140 also includes dynamic correction and updating of the individual world model based on the memory relationships between the basic model objects.

[0218] In step S141c, the short-term and long-term memories of each basic model object are acquired. Short-term memory represents the object's recent state information, including pose, shape parameters, physical function parameters, and semantic attributes, reflecting the object's latest dynamic performance. Long-term memory represents the object's stable state information over a historical timeframe, reflecting the object's average structure, functional characteristics, and semantic properties in past states. By acquiring memory relationship data, temporal continuity and historical stability can be used to predict and correct the single-world model in subsequent steps.

[0219] Based on the short-term and long-term memories obtained in step S141c, a weighted fusion of the short-term and long-term memories is performed on each basic model object in step S140c to generate a predicted state matrix. The predicted state matrix contains the pose information and object parameters at the current moment, reflecting the most likely object state under the constraints of fusing historical and recent states. The weights of the short-term and long-term memories can be adaptively adjusted according to the dynamic characteristics of the object to balance immediate dynamic response and historical stability.

[0220] In step S143c, the predicted state matrix is ​​matched object-by-object with the corresponding individual world model in the overall world model. Specifically, a difference matrix is ​​calculated by comparing the current pose and object parameters of the individual world model with the corresponding values ​​in the predicted state matrix. The difference matrix represents the deviation of the individual world model from the predicted state matrix in terms of pose, geometric parameters, and physical function parameters, providing a basis for subsequent corrections.

[0221] In step S144c, the difference matrix calculated in step S143c is combined with spatial consistency constraints, hierarchical constraints, and functional constraints to solve for the first optimal correction matrix using an optimization method. The optimal correction matrix is ​​used to minimize the prediction error of the single-world model while maintaining the global spatial layout, hierarchical relationships, and functional constraints, thereby achieving a reasonable correction of the state of the single-world model.

[0222] In step S145c, the first optimal correction matrix obtained in step S144c is applied to the corresponding individual world model to update its pose, geometric parameters, and physical function parameters. By correcting and updating each individual world model in the overall world model one by one, it is ensured that the overall world model can maintain spatial consistency, hierarchical consistency, and functional consistency during dynamic changes, while reflecting the continuous evolution of each object over time.

[0223] Through the above steps, this embodiment of the application implements a dynamic prediction and correction mechanism based on short-term and long-term memory, enabling the overall world model to not only maintain the accuracy of its current state but also possess temporal continuity and historical reference stability. By combining the predicted state matrix with the optimal correction matrix, abrupt changes or non-physical distortions in individual world models can be effectively suppressed, and the consistency of multiple individual objects in the overall scene in terms of spatial, hierarchical, and functional dimensions can be guaranteed, thereby realizing the construction of a dynamically evolving world model.

[0224] Appendix Figure 13 The diagram illustrates the short-term and long-term memory mechanisms of a world model in a given scenario. Based on these mechanisms, the state of the basic model objects (or the world model obtained through prior optimization and reconstruction, such as a world model constrained by geometric parameters and / or physical functional parameters) in space is dynamically managed and updated. Short-term memory records rapid changes in the basic model objects over short periods, such as positional movement (coordinates, trajectory points, etc.), posture adjustments (rotation angle, orientation, etc.), or behavioral interactions. Long-term memory records stable data of the model objects over long periods, such as structural features (size, volume, material, etc.), functional attributes (on / off, fault / normal, trigger, speed, etc.), and topological relationships (contact / separation / containment, etc.). Based on this, the various state data accumulated in short-term and long-term memory are organized into a time series, forming a temporal representation of the model objects suitable for large-scale computation and reasoning. This temporal representation represents the state of space, providing a basis for subsequent dynamic analysis, evolutionary prediction, and spatial reasoning.

[0225] Furthermore, at a specific point in time, i.e., a certain time slice in the time series, there are multiple basic model objects applicable to short-term memory or long-term memory mechanisms. The basic model object applicable to short-term memory describes a spatial state that requires real-time data updates or updates over a predetermined short period. Conversely, for basic model objects applicable to long-term memory, the spatial state they describe is relatively stable, and their data updates are triggered by parameter changes exceeding a threshold over a predetermined longer period. Therefore, the basic model objects under both short-term and long-term memory mechanisms comprehensively consider the characteristics of world model changes. By fusing the immediate state in short-term memory with the stable features in long-term memory, they can accurately describe the current state of the basic model object and make reasonable inferences and corrections when state deviations occur, ensuring the basic model object maintains consistency and accuracy in a dynamic spatial environment. Based on this, differentiated and precise spatial operation and management of "changing" and "static" basic model objects can avoid the data processing pressure caused by high-frequency data updates or the stagnation of world model changes caused by long-term data inactivity, enabling the world model to have a comprehensive multi-state expression.

[0226] In this embodiment, the world model is obtained through constraints of geometric parameters and / or physical functional parameters. This world model can also be understood as the world model at the current point in time, and thus the foundational model object for model changes at the next point in time. Therefore, both the foundational model object and the world model are elements describing the model's spatial state, and are also applicable to other scenarios in this application. In this embodiment, a changing world model is generated by fusing the foundational model object (or model object) with the state of space.

[0227] See Figure 14 , Figure 14 It is based on Figure 4 The flowchart described in the corresponding embodiment describes the steps of reconstructing a world model from a basic model object based on geometric parameters and physical function parameters.

[0228] The step S140 of reconstructing a world model from a basic model object based on geometric parameters and physical function parameters provided in this embodiment of the application further includes:

[0229] Step S141d: Obtain social relationship data between basic model objects. The social relationship data includes relationship frequency and dependency distance. Relationship frequency represents the frequency of interaction and / or collaboration between objects, and dependency distance represents the spatial and / or logical range of interaction or influence between objects.

[0230] Step S142d: Based on the relationship frequency and the dependency distance, construct a social relationship weight matrix for each basic model object, and generate a prediction situation matrix between objects in space based on the social relationship weight matrix. The prediction situation matrix is ​​used to represent the potential relative position and dependency relationship of the individual world model in the overall world model.

[0231] Step S143d: Match the predicted situation matrix with the corresponding individual world models in the overall world model, using parameters as objects, and calculate the positional deviation and relational constraint deviation.

[0232] Step S144d: Combine the deviation matrix with the spatial consistency constraint and the social relationship constraint to obtain the second optimal correction matrix;

[0233] Step S145d: Apply the second optimal correction matrix to the corresponding individual world model. In the overall world model, take the individual world model as the object and adjust the position and parameters of each object to obtain the overall world model that takes social relations into account.

[0234] These steps are explained in detail below.

[0235] In this exemplary embodiment, in order to ensure that the overall world model is consistent not only in terms of spatial structure and functional dimensions, but also in terms of rationality in terms of interaction and collaboration between objects, the execution of step S140 also includes a process of constraining and optimizing the individual world model based on social relationship data.

[0236] Specifically, as shown in step S141d, the social relationship data between basic model objects is first acquired. This social relationship data includes relationship frequency and dependency distance. Relationship frequency represents the degree to which objects interact, collaborate, or coordinate; it can be obtained through historical behavior statistics, rule settings, and / or semantic annotation. Dependency distance represents the spatial and logical scope of interaction or influence between objects, such as physical contact range, perceptual coverage range, and / or functional triggering range. The acquisition of social relationship data through step S141d provides a behavioral-level constraint basis for the potential collaborative structure and spatial layout between individual world models.

[0237] In step S142d, a social relationship weight matrix is ​​constructed for each basic model object based on relationship frequency and dependency distance. The social relationship weight matrix is ​​used to quantify the strength of the association between objects, where relationship frequency serves as a weight enhancement factor and dependency distance serves as a range adjustment factor.

[0238] Based on the social relationship weight matrix, a predictive situation matrix is ​​generated between objects in space under a unified coordinate system. The predictive situation matrix is ​​used to represent the potential relative position, interaction priority, and dependency direction of an individual world model within the overall world model, thereby forming a spatial distribution trend driven by social behavior.

[0239] In step S143d, object-by-object matching and deviation calculation are performed on the individual world models within the overall world model. The predicted situation matrix is ​​matched object-by-object with the individual world models within the overall world model, using the pose parameters, geometric parameters, and functional parameters of the individual world models as the matching objects. The spatial deviation of the current position relative to the predicted situation, and the relational constraint deviation of the current object dependencies relative to the social relation constraints are calculated respectively. The spatial deviation reflects the difference between the actual position and the predicted relative position of the individual world model in the overall scene; the relational constraint deviation reflects the degree of inconsistency between the current interaction structure between objects and the dependency structure represented by the social relation weight matrix.

[0240] In step S144d, the positional deviation and relational constraint deviation obtained in step S143d are used to construct a deviation matrix, which is then combined with spatial consistency constraints and social relational constraints to solve for the second optimal correction matrix using constraint optimization methods. During the solution process, while ensuring that the overall world model does not disrupt existing spatial topological relationships and hierarchical structures, the spatial distribution between objects is made consistent with the dependency structure described by the social relational weight matrix, thereby achieving synergistic optimization of spatial structure and social behavioral structure.

[0241] The obtained second optimal correction matrix is ​​applied to the corresponding individual world model to adjust its pose parameters and related functional parameters. In the overall world model, using the individual world model as the update unit, the position and parameters of each object are corrected to make the spatial layout between objects more consistent with their interaction frequency and dependencies. Through this step, the overall world model considering social relationship constraints is obtained.

[0242] In summary, through the execution of steps S141d to S144d, this embodiment of the application, based on the original geometric structure consistency, hierarchical consistency, and functional consistency, introduces constraints of the social relationship dimension, enabling the overall world model to possess not only spatial rationality but also behavioral logical rationality. By constructing the social relationship weight matrix and the prediction situation matrix, the spatial distribution among objects can better conform to their interaction frequency and dependency characteristics, enhancing the realism and interpretability of the overall world model in collaborative scenarios, interactive scenarios, or complex system simulations, thereby improving the comprehensive expressive power and dynamic adaptability of the world model.

[0243] Combined with appendix Figure 15The diagram illustrates social relationships within a specific logical relation in a world model. In the real world, relationships between people, between people and things, and between things exist within the framework of social attributes. For example, consider a scenario on a weekend morning: someone wakes up, randomly chooses a yellow tracksuit, and cooks oatmeal and eggs for breakfast according to their family's request. In this case, "interaction between people" occurs through the social relationship of "family"; "interaction between people and things" is the person randomly "using" the yellow tracksuit based on their preference; and "interaction between things" is the co-occurrence of the preferred combination of oatmeal and eggs. Therefore, in the real world, there also exist social relationships unaffected by physical logic, but expressed by personal subjective will or emotions such as preferences and interdependence.

[0244] Based on this, to construct a more realistic world model, it is also necessary to further digitally represent the social relationships between model objects, as shown in steps S141d to S144d in the exemplary embodiments of this application.

[0245] Specifically, the social relationship data between basic model objects (or world models, such as the world model obtained by constraining geometric parameters and / or physical function parameters in this embodiment) is extracted to extract data that is easy to express digitally, and the situational expression of the space is gradually enriched.

[0246]

[0247] In one embodiment, referring to the table above, the social relationship data of the basic model objects includes relationship frequency and dependency distance. That is, from the two dimensions of "frequency" and "distance," fuzzy, continuous, and contextualized social relationships are transformed into measurable, computable, and analyzable structured data through large-scale models (such as NLP, LLM, etc.) and vertical model agents. It should be noted that the parameter categories listed in the table above include various parameters such as interaction frequency and physical distance, which are used to express the relationship frequency or dependency distance between basic model objects. These parameters are derived from geometry / laser / vision, and the acquisition methods include BIM, point clouds, or images. The above expression is only one embodiment, and the acquisition of similar parameters through other methods or technical means is within the scope of this application.

[0248] Based on this, the digitization of social relations involves transforming unstructured data and implicit social interaction data from the real world into calculable, reasonable, and storable digital features through technologies such as large models or intelligent agents.

[0249] For example, the mathematical expression of the parameter regarding relation frequency, given any two basic model objects N1 and N2, can be uniformly converted to: A / T, where A is the relation content (such as the number of interactions between models a1, usage of a2, co-occurrence of a3, and mutual exclusion of a4), and T is time (such as t seconds / minute). From this, we can obtain the vector of any basic model objects N1 and N2, such as R(N1, N2) = [at1, at3, at3, at4]. If there are multiple basic model objects, the vector between them is calculated in a similar way, thus yielding specific values.

[0250] Based on the same technical concept, the mathematical expression for dependency examples includes physical distance, expressed as absolute distance in spatial coordinates, i.e., Euclidean distance. Behavioral distance is the time series of the positions of the basic model objects, such as basic model objects M1[5,8,13,52,11] and M2[6,11,24,27,37]. At each node in the time series, the absolute value of the difference between basic model objects M1 and M2 is D = |M1-M2|. Emotional dependency involves assigning weights to emotional parameters, such as emotional dependency E = w1·F + w2·P + w3·S, where F is the interaction frequency, P is the proportion of positive emotions, S is the emotional intensity score (0~1), and w1, w2, and w3 are weights with a sum of 1. Functional dependency: For basic model object M3 and model object M4, if basic model object M3 needs to depend on functions g1, g2 and gn to complete a certain task, and these dependent functions g1, g2 and gn come from basic model object M4, then the functional dependency of basic model object M3 is Y = (g1 + g2 + ... + gn) / n, where the value of g is in the range [0, 1].

[0251] Therefore, based on the spatial situation and basic model objects, a world model with social emotions and warmth is generated.

[0252] See Figure 16 , Figure 16 It is based on Figure 4 The flowchart described in the corresponding embodiment describes the steps of reconstructing a world model from a basic model object based on geometric parameters and physical function parameters.

[0253] The step S140 of reconstructing a world model from a basic model object based on geometric parameters and physical function parameters provided in this embodiment of the application further includes:

[0254] Step S141e: Construct a multidimensional relationship feature matrix from the attribute relationship data, memory relationship data, and social relationship data of the basic model objects;

[0255] Step S142e: Perform model reasoning and logical prediction on the multidimensional relation feature matrix, and form the dynamic inference logic of the basic model object by generating dynamic inference rules and state evolution sequences;

[0256] Step S143e: The pose, geometric parameters and physical function parameters of the basic model object at future time are calculated by dynamic deduction logic. The pose, geometric parameters and physical function parameters of the basic model object at future time constitute the predicted state of the basic model object.

[0257] Step S144e involves matching and constraining the predicted state with the parameters of the current individual world model to generate a dynamic correction matrix, which is then applied to the corresponding individual world model on a per-individual basis to obtain the dynamic evolution of the overall world model in the time dimension.

[0258] These steps are explained in detail below.

[0259] In this exemplary embodiment, to achieve continuous evolution and trend prediction of the overall world model over time, step S140 further includes a dynamic evolution process based on multi-dimensional relationship fusion and logical deduction mechanisms. This process integrates attribute relationship data, memory relationship data, and social relationship data to construct a reasonable structure within a unified relationship space, enabling the prediction and updating of the individual world model towards future states.

[0260] In step S141e, the attribute relationship data, memory relationship data, and social relationship data of the basic model objects are uniformly encoded and fused to construct a multi-dimensional relationship feature matrix. Attribute relationship data reflects the spatial and hierarchical relationships between objects; memory relationship data reflects the historical state and stable trends of objects in the time dimension; and social relationship data reflects the interaction frequency and dependence strength between objects. By performing feature normalization and dimension alignment on the above multi-source data, they are mapped to a unified relationship feature space to form a multi-dimensional relationship feature matrix. This matrix is ​​used to characterize the comprehensive relationship structure of objects at the spatial, temporal, and behavioral levels, providing a foundation for subsequent model inference.

[0261] In step S142e, after obtaining the multidimensional relationship feature matrix, model inference and logical prediction processing are performed on it. Model inference can be based on rule-based inference mechanisms, probabilistic models, or learning models to uncover evolutionary patterns and causal relationships between multidimensional relationships. Dynamic inference rules and state evolution sequences are generated through the inference process. The dynamic inference rules describe the driving conditions and constraints of object state changes, while the state evolution sequences characterize the continuous change trend of the object over time. Based on these rules and sequences, the dynamic inference logic of the basic model object is formed to guide the calculation of future states.

[0262] In step S143e, based on dynamic deduction logic, the pose, geometric parameters, and physical function parameters of the basic model object at future moments are calculated. Pose prediction can be derived based on the evolution trend of spatial relationships and social relationship dependencies; geometric parameter prediction can be estimated based on historical change trends and functional constraints; and physical function parameter prediction can be corrected by combining object attribute stability and environmental interaction patterns. The aforementioned pose, geometric parameters, and physical function parameters at future moments together constitute the predicted state of the basic model object. The predicted state is used to reflect the reasonable evolutionary outcome of the object over time.

[0263] In step S144e, the predicted state is first matched with the parameters of the current individual world model on a per-individual basis, and the difference between the predicted state and the current state is calculated. Based on this difference, and combined with spatial consistency constraints, hierarchical constraints, and functional constraints, constraint optimization is performed to obtain the dynamic correction matrix. The dynamic correction matrix is ​​used to ensure the smooth transition of the individual world model towards the predicted state while maintaining the structural stability of the overall world model, avoiding discontinuous jumps or structural breakdowns.

[0264] The dynamic correction matrix is ​​then applied to each individual world model, updating its pose, geometric parameters, and physical function parameters, thereby realizing the dynamic evolution of the overall world model over time. This individual-by-individual update mechanism ensures that the spatial, hierarchical, and social relationships between objects remain consistent throughout time, forming a continuous, predictable, and logically consistent world model evolution process.

[0265] Through the execution of steps S141e to S144e in this exemplary embodiment, a multi-dimensional relationship fusion and dynamic inference mechanism is introduced on the basis of the original spatial consistency and relationship consistency, enabling the overall world model to possess temporal continuity and trend prediction capabilities. By constructing a multi-dimensional relationship feature matrix and dynamic inference logic, the forward-looking calculation of object states can be achieved, and smooth updates can be realized through dynamic correction matrices, thereby improving the applicability and stability of the world model in dynamic simulation, behavior prediction, and complex scene inference.

[0266] In summary, in the implementation of the exemplary embodiments of this application, the basic model object is reconstructed using geometric and physical functional parameters to generate a single-entity world model, and further constructs an overall world model through relationship fusion and constraint optimization. To achieve unified coordination of the overall world model in spatial, hierarchical, behavioral, and temporal dimensions, a multi-dimensional relationship-driven spatiotemporal evolution framework is constructed.

[0267] Thus, this spatiotemporal evolution framework encompasses spatial consistency constraint mechanisms, hierarchical topological recursive synchronization mechanisms, memory fusion prediction mechanisms, social relationship collaborative optimization mechanisms, and multidimensional relationship deduction and evolution mechanisms. This enables the overall world model to simultaneously satisfy spatial structural rationality, hierarchical structural stability, functional consistency, behavioral logic consistency, and temporal continuity in multi-object collaborative scenarios, thereby achieving highly reliable dynamic construction and continuous evolution of multi-object scenarios.

[0268] It should be made clear that for the implementation of each constraint, that is, the implementation of each of the above constraint mechanisms, this constraint mechanism can be implemented based on the initially generated basic model object and at least one other constraint basis, and the world model reconstructed on the basic model object. It is not limited to the sequential execution of each constraint mechanism to obtain a world model with different capabilities, or even a world model with multiple capabilities.

[0269] To be further clarified, for the initially generated basic model object, the world model can be reconstructed through parameter constraints. In addition, other constraint modifications can be applied to the reconstructed world model of the basic model object or the specific constraint modification, without limitation here.

[0270] At this point, we will further illustrate the “basic model object” and “world model” in the embodiments of this application with specific examples.

[0271] For example, a basic model object is a structured digital representation of an "independent object" in a real or virtual scene. For instance, in a smart home scenario, a basic model object can be a single entity such as a "table," "chair," "robot," or "light fixture"; in a transportation scenario, a basic model object can be a "vehicle," "pedestrian," or "traffic light"; and in an industrial scenario, a basic model object can be a "robotic arm," "conveyor belt," or "workpiece."

[0272] Each basic model object contains at least two types of parameter information: geometric parameters and physical function parameters. When a certain type of parameter information is reconstructed, a complete digital expression of the object is formed, which constitutes the corresponding single-entity world model. For example, the single-entity world model of a table includes not only its three-dimensional dimensions and position, but also its load-bearing capacity, its contact relationship with the ground, and its functional relationship with the items on the table.

[0273] When multiple basic model objects exist simultaneously in the same scene, such as "a computer on a table with a robot standing next to it," it is necessary to integrate the multiple individual world models according to their spatial location, hierarchical structure, and interaction logic to form a holistic world model that describes the state of the entire scene. The holistic world model can be understood as "a complete digital mapping of the entire scene at a certain moment," which not only describes the state of each object itself, but also the relative positional relationships, dependencies, and collaborative states between objects.

[0274] In summary, the resulting overall world model transforms a complex multi-object scenario into a computable, predictable, optimizable, and collaboratively controllable digital operating system. This effectively avoids the conflicts and inefficiencies caused by separate modeling of each object, dispersed states, and lack of unified relational expression in traditional systems, thereby improving the overall operational reliability and intelligence level in complex scenarios.

[0275] In one embodiment, for model objects, the target is an individual "point," so geometric parameters and physical function parameters are stored in the model object. For various relationships in space, which are relationships between multiple model objects, the target is a "surface," so attribute relationship data, memory relationship data, and social relationship data are stored in the spatial cloud storage of the target world model. In other embodiments, spatial cloud storage is used for multiple regions that can be divided into the target world model. For example, the world model can be divided into multiple small spaces according to the scene, such as residential spaces, roads, bridges, etc. This distributed data design avoids the data from being too concentrated, resulting in massive data volumes. Furthermore, for spatial regions that are not of interest, computing power can be reduced, ineffective spatial deductions can be avoided, and resource utilization can be improved. This data storage and data processing improves the efficiency of world model operation.

[0276] In one exemplary embodiment, this application also provides a computer electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method as described above.

[0277] In one exemplary embodiment, this application also provides a computer program product including a computer program that, when executed by a processor, implements the steps of the method as described above.

[0278] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, terminal device, or network device, etc.) to execute the method according to the embodiments of this application.

[0279] According to one embodiment of this application, a program product for implementing the methods in the above-described method embodiments is also provided. This product may employ a portable compact disc read-only memory (CD-ROM) and include program code, and may run on a terminal device, such as a personal computer. However, the program product of this invention is not limited thereto. In this document, a readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0280] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, mobile terminal, or network device, etc.) to execute the methods according to the embodiments of this application.

[0281] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the appended claims.

Claims

1. A method for dynamically generating a world model, characterized in that, The method includes: A basic model object is generated from a type of raw data to create the initial three-dimensional form of the world model it describes. The basic model object is an initial geometric assumption model. The type of raw data is geometric data, laser point cloud data, or visual data. Initial parameters are extracted from various types of raw data to obtain geometric structure information and physical information from the geometric data, spatial location information and surface information from the laser point cloud data, and contour information and semantic information from the visual data; Using the geometric data, visual data, and laser point cloud data as mutual parameter constraints, geometric parameters and physical function parameters for modifying the basic model object are obtained; wherein the parameter constraint step includes: mapping geometric structure information, spatial location information, and contour information to a unified parameter space for consistency verification and error calculation, and updating the geometric parameters with the obtained error; and matching based on physical information, surface information, and semantic information to obtain the updated physical function parameters. The corrected geometric parameters are applied to update the model space structure, the corrected physical function parameters are loaded into the model attribute layer, and finally, after verifying the consistency of each parameter, a complete 3D world model is generated.

2. The method according to claim 1, characterized in that, The basic model object mapping can describe the initial geometric parameters of the initial three-dimensional form of the world model, including size parameters, view parameters, and rendering parameters; The step of mapping geometric structure information, spatial location information, and contour information to a unified parameter space for consistency verification and error calculation, and updating the geometric parameters with the obtained error, includes: Using the spatial coordinates and size scale in the spatial location information, the consistency of the two-dimensional edge and shape size adapted to the contour information is checked. The size error obtained from the check is used to optimize the size parameters of the basic model object until the updated size parameters are obtained by convergence under the size error control. The actual contour is extracted from the contour information of the visual data by extracting the contour concavity and convexity points and the contour area ratio. Under the initial viewpoint parameters corresponding to the current spatial position information, the corresponding three-dimensional model structure is projected to generate the predicted contour. The actual contour and the predicted contour are compared. In the comparison, the viewpoint error is calculated by combining the vanishing point position of the geometric structure information and the combination of visible surfaces. The viewpoint parameters mapped by the spatial position information are optimized according to the viewpoint error. The basic model object is rendered and a predicted outline is generated based on the current rendering parameters. The rendering consistency error is calculated by comparing the predicted outline with the actual outline extracted from the visual data. The rendering parameters are then adjusted based on the rendering consistency error. The mutually constrained size parameters, viewpoint parameters, and rendering parameters form the constrainable geometric parameters of the base model object.

3. The method according to claim 1, characterized in that, The physical function parameters include the facade parameters of the physical information, the relational parameters of the surface information, and the object parameters of the semantic information; The steps for obtaining updated physical function parameters based on matching physical information, surface information, and semantic information include: By checking the geometric consistency and semantic consistency of the actual normal vector and actual surface curvature in the surface information, and the actual texture and color in the semantic information, respectively, the facade constraint error is constructed. Starting from the initial value of the facade parameters, the facade constraint error is minimized and iteratively optimized to obtain the updated value of the facade parameters in the physical information. The semantic relationship constraint set mapped by the object description and spatial relationship of the semantic information is matched with the current geometric relationship state of the patch topology and size ratio mapping of the physical information and the error is calculated. The updated value of the relationship parameter in the surface information is obtained by iterative optimization of minimizing the obtained relationship constraint error. A physical feature vector is constructed by classifying the volume and material of the physical information, the inverse strength and attribute classification of the surface information, and mapping the physical feature vector to the semantic category. The mapped semantic category is compared with the object parameters in the semantic information and the semantic consistency error is calculated. The updated value of the object parameters in the semantic information is obtained by minimizing the semantic consistency error through iterative optimization. The mutually constrained facade parameters, relational parameters, and object parameters are combined to form constrainable physical functional parameters.

4. The method according to claim 2 or 3, characterized in that, The steps of applying the corrected geometric parameters to update the model space structure, loading the corrected physical function parameters into the model attribute layer, and finally performing consistency verification on each parameter to generate a complete 3D world model include: The geometric parameters and physical function parameters are aligned to their respective parameter dimensions, and the parameters corresponding to the basic model object are compared, and the deviation is calculated. For each parameter dimension, determine whether the obtained deviation value exceeds the preset deviation threshold. If it exceeds the deviation threshold, mark the corresponding parameter dimension in the basic model object as the dimension to be updated. For the dimension to be updated, a weighted update or a direct replacement update is performed based on the corresponding deviation value and a preset update step size to update the corresponding current parameters of the basic model object; After the parameter update is completed, structural consistency synchronization is performed on the base model object to generate a reconstructed world model. The structural consistency synchronization includes recalculating the geometric structure and derived attribute data of the base model object with the updated parameters.

5. The method according to claim 4, characterized in that, The world model generated by reconstructing the basic model object is a single-world model. The steps of applying the corrected geometric parameters to update the model space structure, loading the corrected physical function parameters into the model attribute layer, and finally performing consistency verification on each parameter to generate a complete 3D world model also include: Obtain attribute relationship data between the basic model objects, the attribute relationship data including spatial relationship parameters and subordinate hierarchy relationship parameters between the basic model objects; An object space relationship graph is constructed based on the spatial relationship parameters, and the global pose transformation matrix of each individual world model is obtained by solving the global pose in a unified coordinate system using the spatial relationship parameters. The pose transformation of the corresponding individual world model is then performed using the global pose transformation matrix. A hierarchical structure tree for the subordinate hierarchical relationship parameters is constructed. In the hierarchical structure tree for the individual model, the pose transformation of the child nodes is recursively synchronized to the child nodes according to the hierarchical topology order, with the parent node of the pose transformation as the reference. The pose transformation of the child node is superimposed with its own local correction matrix. The local correction matrix is ​​the incremental pose matrix generated during the construction of the spatial relationship graph, the optimization of geometric consistency constraints, or the correction of physical function constraints. Based on the spatial relationship diagram and the hierarchical structure tree of the individual model, conflict detection is performed on the individual world model in the spatial, hierarchical and functional dimensions. The detected conflict items are corrected and fused to generate an overall world model that satisfies spatial consistency, hierarchical consistency and functional consistency.

6. The method according to claim 5, characterized in that, The steps of applying the corrected geometric parameters to update the model space structure, loading the corrected physical function parameters into the model attribute layer, and finally performing consistency verification on each parameter to generate a complete 3D world model also include: Obtain memory relationship data between the basic model objects. The memory relationship data includes short-term memory and long-term memory. The short-term memory represents the state information of the basic model objects at the most recent moment, and the long-term memory represents the stable state information of the basic model objects within a historical time range. For each basic model object, a prediction state matrix is ​​generated by weighted fusion of the short-term memory and long-term memory, and the prediction state matrix includes pose and object parameters. Based on the predicted state matrix, object-by-object matching of the current pose and parameters is performed on the individual world model corresponding to the overall world model, and the difference matrix of the individual world model relative to the current predicted state is calculated. The difference matrix is ​​combined with spatial consistency constraints, hierarchical constraints, and functional constraints to solve for the first optimal correction matrix; The first optimal correction matrix is ​​applied to the corresponding individual world model to perform a unit-by-unit correction update of the individual world model in the overall world model.

7. The method according to claim 6, characterized in that, The steps of applying the corrected geometric parameters to update the model space structure, loading the corrected physical function parameters into the model attribute layer, and finally performing consistency verification on each parameter to generate a complete 3D world model also include: Obtain social relationship data between the basic model objects. The social relationship data includes relationship frequency and dependency distance. The relationship frequency represents the frequency of interaction and / or collaboration between objects, and the dependency distance represents the spatial and / or logical range of interaction or influence between objects. Based on the relationship frequency and the dependency distance, a social relationship weight matrix is ​​constructed for each basic model object, and a prediction situation matrix between objects in space is generated based on the social relationship weight matrix. The prediction situation matrix is ​​used to represent the potential relative position and dependency relationship of the individual world model in the overall world model. The predicted situation matrix is ​​matched with the corresponding individual world model in the overall world model, and the positional deviation and relational constraint deviation are calculated by matching the parameters one by one. By combining the deviation matrix with spatial consistency constraints and social relationship constraints, the second optimal correction matrix can be obtained. The second optimal correction matrix is ​​applied to the corresponding individual world model. In the overall world model, the position and parameters of each individual world model are adjusted to obtain an overall world model that takes social relations into account.

8. The method according to claim 7, characterized in that, The steps of applying the corrected geometric parameters to update the model space structure, loading the corrected physical function parameters into the model attribute layer, and finally performing consistency verification on each parameter to generate a complete 3D world model also include: A multidimensional relationship feature matrix is ​​constructed from the attribute relationship data, memory relationship data, and social relationship data of the basic model object; The multidimensional relation feature matrix is ​​subjected to model reasoning and logical prediction, and the dynamic inference logic of the basic model object is formed by generating dynamic inference rules and state evolution sequences. The pose, geometric parameters, and physical function parameters of the basic model object at future time are calculated through the dynamic inference logic. The pose, geometric parameters, and physical function parameters of the basic model object at future time constitute the predicted state of the basic model object. The predicted state is matched and constrained with the parameters of the current individual world model to generate a dynamic correction matrix, which is then applied to the corresponding individual world model on a per-individual basis to obtain the dynamic evolution of the overall world model in the time dimension.

9. An electronic device comprising a computer program, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method as described in any one of claims 1 to 8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images