Method, device, medium and equipment for generating body-aware scene description file

Abstract

The application discloses a generation method and device of a body intelligent scene description file, a medium and equipment, and the method comprises the following steps: when a target industry body intelligent training scene needs to be constructed, standardization parameters of common elements are called from a basic scene description layer, the basic scene description layer is used for standardization representation of common elements of at least one dimension extracted from a plurality of industry body intelligent scenes; corresponding target industry plug-ins are selected and loaded from a personalized scene element layer according to target industry requirements, the personalized scene element layer comprises industry plug-ins for describing individual requirements of each industry in a plug-in manner; through a semantic mapping relationship between the basic scene description layer and the personalized scene element layer pre-established by a semantic association layer, the standardization parameters of the common elements are automatically bound and constraint propagation is performed on individual requirement information described by the target industry plug-ins, and a complete description file of the target industry scene is generated.

Description

Technical Field

[0001] This application relates to the field of embodied intelligence technology, and more specifically, to a method, apparatus, medium, and device for generating embodied intelligence scene description files. Background Technology

[0002] Embodied intelligence, as an important development direction of artificial intelligence, emphasizes that intelligent agents perceive their environment through physical interaction, make autonomous decisions, and execute complex tasks, with broad application value in manufacturing, logistics, special operations, public services, and other fields. Driven by the intelligent upgrading of industries, training scenarios that can be quickly transferred, efficiently reused, and flexibly expanded have become the core support for training the general capabilities of intelligent agents and reducing the cost of industry implementation.

[0003] Currently, the construction of embodied intelligent training scenarios relies on scenario description models to achieve digital representation of scenarios. Existing models are mostly customized designs for single industries, modeling only the specific scenario features, entity attributes, and task rules of a particular industry, without systematically sorting out and uniformly abstracting the common features of scenarios across multiple industries. The model architecture is deeply bound to industry scenarios, and the description logic is rigid, making it difficult to adapt to the needs of cross-industry scenario construction.

[0004] Therefore, existing technologies have the following prominent problems: First, the scenarios are significantly fragmented, with inconsistent modeling standards and heterogeneous data formats across industries, making it impossible for scenarios to be shared and reused, resulting in redundant construction, waste of resources, and high overall construction costs; Second, the cross-industry migration capability is weak, and the models lack a universal representation mechanism, requiring redesign of modeling logic and development of adaptation modules for new industry scenarios, which leads to long development cycles and high expansion difficulty; Third, the architecture has high coupling and poor maintainability, with different dimensions of elements in the scenario being interdependent and closely related, and adjusting or adding scenario features can easily trigger a chain of adaptation problems, making it difficult to support the rapid construction and iterative optimization of large-scale, cross-industry embodied intelligent training scenarios. Summary of the Invention

[0005] This application provides a method, apparatus, medium, and device for generating embodied intelligence scene description files, which can solve the technical problems of existing embodied intelligence training scenes, such as fragmentation, poor cross-industry adaptability, low reusability, high construction cost, difficulty in expansion and maintenance, and inability to efficiently support cross-industry large-scale training.

[0006] The specific technical solution is as follows: In a first aspect, embodiments of this application provide a method for generating an embodied intelligent scene description file, the method comprising: When it is necessary to construct an embodied intelligence training scenario for a target industry, standardized parameters of common elements are retrieved from the basic scenario description layer. The basic scenario description layer is used to standardize the common elements of at least one dimension extracted from embodied intelligence scenarios of multiple industries. Based on the needs of the target industry, select and load the corresponding target industry plugins from the personalized scene element layer. The personalized scene element layer contains industry plugins in a plug-in manner to describe the personalized needs of each industry. By establishing a semantic mapping relationship between the basic scene description layer and the personalized scene element layer in advance through the semantic association layer, the standardized parameters of the common elements are automatically bound and constrained with the personalized demand information described by the target industry plugin, generating a complete description file of the target industry scene. The constrained propagation refers to applying the constraints defined in the target industry plugin to the standardized parameters of the bound common elements and restricting or adjusting the related parameters in a linked manner.

[0007] In one possible implementation, the common elements include static common elements and dynamic common elements; The static common elements include at least one dimension from the following: agent elements, environment elements, object elements, interactive subject elements, and rule elements; and / or, The dynamic common elements include motion atoms and operational atoms used to describe the behavior of embodied intelligent agents.

[0008] In one possible implementation, the method for obtaining the static common elements includes: Collect corpora of embodied intelligence scenarios from multiple industries; Candidate elements are extracted from embodied intelligent scene corpora of the aforementioned industries using named entity recognition technology; Candidate elements that meet preset screening conditions are selected from the candidate elements. The preset screening conditions include that the cross-industry coverage rate of the candidate elements is greater than or equal to a preset coverage rate threshold, and the weighted word frequency is greater than or equal to a preset word frequency threshold. The selected candidate elements are semantically clustered and merged to converge into static common elements with at least one preset dimension.

[0009] In one possible implementation, the industry plugin includes at least one of an industry-specific constraint description module, a special operation task description module, and a special entity description module.

[0010] In one possible implementation, the method for generating the industry plugin includes: Obtain the element set for each industry, and remove the common elements from the element set to obtain the candidate set of individual elements for each industry; Calculate the term frequency-inverse document frequency (TF-IDF) value of each element in the candidate set of personality elements in the corresponding industry corpus, and use the TF-IDF value as the industry saliency score of the element; The elements in the individual candidate set are screened based on the industry saliency score; The selected elements are encapsulated into standardized plug-in modules according to at least one of the following categories: industry-specific constraints, special operation tasks, and special entities, to form the industry plug-in.

[0011] In one possible implementation, a semantic mapping relationship is pre-established between the basic scene description layer and the personalized scene element layer, including: The semantic similarity between the common elements and the individual needs information described by the industry plugin is calculated, and the elements with similarity exceeding a preset similarity threshold are identified as target alignment items to complete the concept layer alignment. The common elements are matched with the personalized demand information described by the industry plugin using dimensional standardization and boundary constraint matching to complete the attribute layer alignment. Dimensional standardization matching is to uniformly convert the dimensions of each attribute in the personalized demand information into the same standard dimensional unit as the common elements. Boundary constraint matching is to match or constrain the value range or boundary conditions of each attribute in the personalized demand information with the preset boundary of the corresponding attribute in the common elements. A consistency check is performed on the alignment results of the concept layer and the alignment results of the attribute layer, and any existing consistency conflicts are resolved to obtain the semantic mapping relationship.

[0012] In one possible implementation, the complete description file includes a common element encoding segment, a personalized element plug-in tag segment, and a semantic tag segment; wherein the semantic tag segment is used to describe the semantic information of the task.

[0013] Secondly, embodiments of this application provide an apparatus for generating an embodied intelligent scene description file, the apparatus comprising: The retrieval unit is used to retrieve standardized parameters of common elements from the basic scenario description layer when it is necessary to construct an embodied intelligence training scenario for a target industry. The basic scenario description layer is used to standardize and represent common elements of at least one dimension extracted from embodied intelligence scenarios of multiple industries. The selection loading unit is used to select and load the corresponding target industry plugin from the personalized scene element layer according to the target industry needs. The personalized scene element layer contains industry plugins that describe the personalized needs of each industry in a plug-in manner. The file generation unit is used to automatically bind and constrain the standardized parameters of the common elements with the personalized demand information described by the target industry plugin through the semantic mapping relationship between the basic scene description layer and the personalized scene element layer established in advance by the semantic association layer, and generate a complete description file of the target industry scene. The constraint propagation refers to applying the constraint conditions defined in the target industry plugin to the standardized parameters of the bound common elements and linking or adjusting the related parameters.

[0014] In one possible implementation, the common elements include static common elements and dynamic common elements; The static common elements include at least one dimension from the following: agent elements, environment elements, object elements, interactive subject elements, and rule elements; and / or, The dynamic common elements include motion atoms and operational atoms used to describe the behavior of embodied intelligent agents.

[0015] In one possible implementation, the device further includes: An acquisition unit is used to acquire the static common elements; The acquisition unit includes: The data acquisition module is used to collect corpora of embodied intelligent scenarios from multiple industries; The extraction module is used to extract candidate elements from the embodied intelligent scene corpus of the multiple industries using named entity recognition technology. The first filtering module is used to filter out candidate elements that meet preset filtering conditions from the candidate elements. The preset filtering conditions include that the cross-industry coverage rate of the candidate elements is greater than or equal to a preset coverage rate threshold, and the weighted word frequency is greater than or equal to a preset word frequency threshold. The clustering convergence module is used to perform semantic clustering and merging of the selected candidate elements, converging them into static common elements with at least one preset dimension.

[0016] In one possible implementation, the industry plugin includes at least one of an industry-specific constraint description module, a special operation task description module, and a special entity description module.

[0017] In one possible implementation, the device further includes: A plugin generation unit is used to generate the industry plugins; The plugin generation unit includes: The acquisition and elimination module is used to acquire the element set of each industry and eliminate the common elements from the element set to obtain the individual element candidate set of each industry; The calculation module is used to calculate the word frequency-inverse document frequency (TF-IDF) value of each element in the candidate set of personality elements in the corresponding industry corpus, and use the TF-IDF value as the industry saliency score of the element; The second filtering module is used to filter elements in the individual candidate set based on the industry salience score; The encapsulation module is used to encapsulate the selected elements into standardized plug-in modules according to at least one of the following categories: industry-specific constraint class, special operation task class, and special entity class, so as to form the industry plug-in.

[0018] In one possible implementation, the device further includes: A unit is established to pre-establish the semantic mapping relationship between the basic scene description layer and the personalized scene element layer; The establishment unit includes: The first alignment module is used to calculate the semantic similarity between the common elements and the individual needs information described by the industry plugin, and to determine the elements with similarity exceeding a preset similarity threshold as target alignment items in order to complete the concept layer alignment. The second alignment module is used to perform dimensional standardization matching and boundary constraint matching between the common elements and the personalized demand information described by the industry plugin to complete the attribute layer alignment. The dimensional standardization matching is to uniformly convert the dimensions of each attribute in the personalized demand information into the same standard dimensional unit as the common elements. The boundary constraint matching is to match or constrain the value range or boundary conditions of each attribute in the personalized demand information with the preset boundary of the corresponding attribute in the common elements. The verification module is used to perform consistency verification on the alignment results of the concept layer and the alignment results of the attribute layer, and to resolve any consistency conflicts to obtain the semantic mapping relationship.

[0019] In one possible implementation, the complete description file includes a common element encoding segment, a personalized element plug-in tag segment, and a semantic tag segment; wherein the semantic tag segment is used to describe the semantic information of the task.

[0020] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method as described in any possible implementation of the first aspect.

[0021] Fourthly, embodiments of this application provide an electronic device, which includes: One or more processors; The processor is coupled to a storage device for storing one or more programs; When one or more programs are executed by one or more processors, the electronic device performs the method as described in any possible implementation of the first aspect.

[0022] Fifthly, embodiments of this application provide a computer program product containing instructions that, when executed on a computer or processor, cause the computer or processor to perform the method described in any possible implementation of the first aspect.

[0023] The method, apparatus, medium, and device for generating embodied intelligent scene description files provided in this application embodiment construct a layered scene description architecture. It utilizes a basic scene description layer to achieve standardized representation of common elements across multiple industries, relies on a pluggable personalized scene element layer to flexibly adapt to the differentiated needs of different industries, and leverages a semantic association layer to automatically bind and propagate constraints between general parameters and industry-specific information. This allows for the rapid generation of complete description files for target industry scenes, effectively solving the problems of fragmented scenes, poor cross-industry adaptability, low reusability, high construction costs, and difficult expansion and maintenance in existing technologies. It significantly improves the cross-industry reusability, adaptation efficiency, and maintainability of scenes, greatly reduces the cost of repeated scene construction, and efficiently supports the rapid construction and iterative optimization of cross-industry large-scale embodied intelligent training scenes.

[0024] In addition, the embodiments of this application can achieve at least the following technical effects: 1. This application's embodiments accurately extract cross-industry common elements and form standardized dimensions by collecting multi-industry scenario corpora, extracting candidate elements through named entity recognition, filtering with dual thresholds, and merging semantic clusters. This effectively solves the problems of existing technologies, such as lack of multi-industry common feature system abstraction, strong arbitrariness in element extraction, and insufficient universality. It ensures the representativeness, universality, and standardization of static common elements, provides a reliable foundation for unified modeling of cross-industry scenarios, and further enhances the model's cross-industry adaptability and reuse value.

[0025] 2. This application's embodiments obtain industry-specific candidate elements by eliminating common elements, calculate industry saliency scores based on TF-IDF and accurately screen them, and then encapsulate them into standardized plugins according to constraints, tasks, and entity categories. This achieves structured, modular encapsulation and independent management of industry-differentiated elements, effectively solving the problems of tight coupling between industry-specific and general elements, lack of standardized encapsulation, need for redevelopment for cross-industry adaptation, poor reusability, and high maintenance costs in existing technologies. It enables flexible plug-and-play access to industry-differentiated needs, significantly reduces the difficulty of adapting to new industry scenarios, improves scenario expansion efficiency and module reusability, avoids mutual interference between industry-specific and general elements, and ensures the independence and maintainability of scenario modeling.

[0026] 3. The embodiments of this application achieve precise alignment of the concept layer, standardized units, and unified adaptation of the attribute layer through semantic similarity calculation, and construct a reliable semantic mapping relationship by combining consistency verification and conflict resolution. This effectively solves the problems of semantic ambiguity between general elements and industry-specific elements, error-prone parameter adaptation, mismatched constraints, and complex coupling relationships that lead to low adaptation efficiency, error-proneness, and difficult maintenance in the prior art. It ensures that general parameters and industry-specific information are accurately bound and that the constraint logic is consistent, avoiding parameter conflicts or logical contradictions during cross-industry adaptation. This significantly improves the reliability of semantic mapping, the accuracy of parameter binding, and the stability of scene adaptation, further reducing the cost of cross-industry adaptation and debugging, and enhancing the robustness of the model and the accuracy of scene generation. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0028] Figure 1 A flowchart illustrating a method for generating an embodied intelligent scene description file, provided in an embodiment of this application; Figure 2 This is a block diagram of an apparatus for generating an embodied intelligent scene description file, provided in an embodiment of this application. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0030] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The terms "comprising" and "having," and any variations thereof, in the embodiments and drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0031] This application provides a flowchart illustrating a method for generating an embodied intelligent scene description file. This method can be applied to electronic devices or computer devices, specifically embodied intelligent devices, such as… Figure 1 As shown, the method may include the following steps: S110: When it is necessary to construct an embodied intelligent training scenario for a target industry, retrieve standardized parameters of common elements from the basic scenario description layer.

[0032] The basic scenario description layer is used to standardize and represent common elements of at least one dimension extracted from embodied intelligent scenarios across multiple industries. These common elements include static and dynamic common elements; the static common elements include at least one dimension of agent elements, environment elements, object elements, interaction subject elements, and rule elements; and / or, the dynamic common elements include motion atoms and operation atoms used to describe the behavior of embodied intelligent agents.

[0033] The elements of an intelligent agent include its physical attributes and / or capability boundaries; the physical attributes of an intelligent agent include its size, number of joints, sensing modules, and actuators, etc., and the capability boundaries of an intelligent agent include its limits of motion, operation, and perception.

[0034] Environmental elements include static and / or dynamic environments; static environments include spatial structure, floor material, wall material, furniture layout, equipment layout, etc., while dynamic environments include lighting, temperature and humidity, and obstacle movement, etc.

[0035] Object elements include at least one of the following: physical attributes, functional attributes, and state attributes. Physical attributes include the object's shape, size, material, weight, and stiffness. Functional attributes include operable parts and usage rules. State attributes include attributes such as stationary, moving, intact, and damaged.

[0036] The interactive elements include the interaction elements between the current intelligent agent and the human, and the interaction elements between the current intelligent agent and other intelligent agents; the interaction elements between the current intelligent agent and the human include behavioral habits, interaction intentions, and safe distance, etc., and the interaction elements between the current intelligent agent and other intelligent agents include cooperation rules and communication methods, etc.

[0037] The rule elements include at least one of the following: operation rules, safety rules, and task rules; operation rules include the correct way to operate an object; safety rules include safety constraints on human-machine / machine-object interaction; and task rules include the execution order of multiple tasks.

[0038] Motion atoms are used to describe the displacement of an embodied intelligent agent, including moving straight and avoiding obstacles, while manipulation atoms are used to describe the physical interaction between the end effector and the environment, including grasping and rotating.

[0039] It should be noted that the object elements described in the embodiments of this application specifically refer to non-intelligent passive entities that exist in the scene as objects of intelligent agent operations or manipulations, such as workpieces, medicines, tools, and goods. The distinction between these and intelligent agent elements and interactive subject elements is clear. Intelligent agent elements refer to active operating entities with autonomous perception, decision-making, and execution capabilities; interactive subject elements refer to humans or other collaborating parties with autonomous behavioral capabilities; and object elements refer only to passive objects other than the above two categories that are manipulated by intelligent agents or interactive subjects. In this invention, these three elements are independent and do not overlap according to the role logic of active operator, passive object, and collaborative interaction party.

[0040] In one possible implementation, the method for obtaining static common elements includes: collecting embodied intelligent scene corpora from multiple industries; extracting candidate elements from the embodied intelligent scene corpora from multiple industries using named entity recognition technology; selecting candidate elements from the candidate elements that meet preset selection conditions, the preset selection conditions including that the cross-industry coverage rate of the candidate elements is greater than or equal to a preset coverage rate threshold, and that the weighted word frequency is greater than or equal to a preset word frequency threshold; and performing semantic clustering and merging on the selected candidate elements to converge into static common elements of at least one preset dimension.

[0041] At least one preset dimension includes at least one of the following: intelligent agent element, environmental element, object element, interactive subject element, and rule element. The embodied intelligent scenario corpus for each industry includes scenario description documents, training requirement texts, industry standard documents, etc. The cross-industry coverage rate of candidate elements refers to the proportion of the number of industries in which a candidate element actually appears in all target industries (such as manufacturing, logistics, special operations, and public services) out of the total number of industries. For example, if there are 4 industries and a candidate element appears in 3 industry corpora, its coverage rate is 3 / 4 = 0.75. Weighted word frequency assigns differentiated weights to different industries and types of corpora (weights can be set according to industry representativeness, corpus data scale, and scenario importance; for example, core industries have higher weights, and large-scale authoritative corpora have higher weights). Finally, the weighted frequency of the element in all corpora is calculated to avoid misselection of pseudo-core elements that are high in one industry but low in others. The specific values ​​of the preset coverage rate threshold and preset word frequency threshold can be determined based on practical experience; for example, the preset coverage rate threshold is 0.75, and the preset word frequency threshold is 0.02.

[0042] S120: Select and load the corresponding target industry plugin from the personalized scene element layer according to the needs of the target industry.

[0043] The personalized scenario element layer includes industry plugins in a plug-in format to describe the unique needs of various industries. Each industry plugin includes at least one of the following: an industry-specific constraint description module, a special task description module, and a special entity description module.

[0044] The method for generating industry plugins includes: obtaining an element set for each industry and removing common elements from the element set to obtain a candidate set of unique elements for each industry; calculating the TF-IDF (Term Frequency-Inverse Document Frequency) value of each element in the candidate set of unique elements in the corresponding industry corpus, and using the TF-IDF value as the industry saliency score of the element; filtering the elements in the candidate set of unique elements according to the industry saliency score; and encapsulating the filtered elements into standardized plugin modules according to at least one of the following categories: industry-specific constraint class, special task class, and special entity class, to form industry plugins.

[0045] Specifically, the selected elements are encapsulated into industry-specific constraint description modules according to industry-specific constraint classes, into special task description modules according to special task classes, and into special entity description modules according to special entity classes.

[0046] The method for obtaining the element set for each industry includes: obtaining the embodied intelligent scene corpus for each industry, and then using named entity recognition technology to obtain multiple elements from the embodied intelligent scene corpus of multiple industries to form an element set.

[0047] Methods for screening elements in the individual candidate set based on industry saliency scores include: first, normalizing the TF-IDF values ​​of all elements in the individual candidate set (within the range of 0 to 1) to eliminate numerical bias caused by differences in corpus size and ensure comparable scores; then, retaining elements with normalized TF-IDF values ​​≥ a preset threshold and removing elements with values ​​below the preset threshold (these elements have weak specificity and are mostly marginal or noisy elements); or sorting by normalized TF-IDF in descending order and selecting the top N% (e.g., the top 20%) of high-value elements.

[0048] Industry-specific constraints refer to the unique restrictions, compliance standards, and boundary constraints specific to each industry and distinct from general safety / operation rules. These are differentiated, mandatory rules that must be followed in each industry scenario. For example, industry-specific constraints in the manufacturing industry include explosion-proof rating constraints for welding workstations and temperature and humidity control constraints within ±0.5℃ for precision machining.

[0049] Specialized operational tasks are unique to each industry and are not universally applicable across industries. They represent the core differentiating execution tasks within a specific industry scenario. For example, specialized operational tasks in the logistics industry include fresh food cold chain sorting and packaging, and customs verification for cross-border goods.

[0050] Specialized entities are industry-specific physical devices, tools, carriers, and facilities that are not cross-industry generic objects. They are the core entities that differentiate different industry scenarios. For example, specialized entities in the public service industry include subway security gates, intelligent guide robots, and emergency broadcast terminals.

[0051] S130: By establishing a semantic mapping relationship between the basic scene description layer and the personalized scene element layer in advance through the semantic association layer, the standardized parameters of common elements are automatically bound and constrained to the personalized demand information described by the target industry plugin, generating a complete description file of the target industry scene.

[0052] Among them, constraint propagation refers to applying the constraints defined in the target industry plugin to the standardized parameters of the bound common elements, and imposing linkage restrictions or adjustments on related parameters.

[0053] The specific implementation methods for establishing semantic mapping relationships may include: calculating the semantic similarity between common elements and the personalized demand information described by industry plugins, and identifying elements with similarity exceeding a preset similarity threshold as target alignment items to complete conceptual-level alignment; performing dimensional standardization matching and boundary constraint matching between common elements and the personalized demand information described by industry plugins to complete attribute-level alignment. Dimensional standardization matching involves converting the dimensions of each attribute in the personalized demand information to the same standard dimensional units as the common elements, and boundary constraint matching involves matching or constraining the value range or boundary conditions of each attribute in the personalized demand information with the preset boundaries of the corresponding attributes in the common elements; performing consistency verification on the conceptual-level alignment results and attribute-level alignment results, and resolving any consistency conflicts to obtain the semantic mapping relationship.

[0054] In addition, it is necessary to define the standardized range of parameters for each element, clarify the unified representation format of common elements and the plug-in interface standard for individual elements, and ensure the compatibility and scalability of the model.

[0055] Taking industrial assembly and medical dispensing scenarios as examples, the basic scenario description layer uniformly represents common elements such as the motion parameters of the intelligent agent and the physical properties of the object; the industrial industry plug-in includes modules for anti-static and assembly precision, while the medical industry plug-in includes modules for sterile environment and drug specifications; and a semantic association layer is used to establish a mapping between assembly precision and the operational limits of the intelligent agent, so as to achieve rapid scenario adaptation.

[0056] Taking precision assembly in the manufacturing industry and pharmacy dispensing in the medical industry as examples, the implementation process of the embodiments of this application is explained in detail: (1) Extraction of common elements in five dimensions: For precision assembly scenarios and pharmacy dispensing scenarios, extract intelligent agent elements (such as robotic arm size and sensing module type), environmental elements (such as workshop / pharmacy space layout), object elements (such as physical properties of workpieces / medicines), interactive subject elements (such as worker / medical staff interaction rules), and rule elements (such as assembly accuracy / dispensing safety rules) as common elements; (2) Construction of basic scene description layer: Based on the above five common elements, a standardized basic description module is built to unify the basic representation format of the scene; (3) Personalized scene element layer design: The manufacturing industry plug-in includes a precision assembly accuracy constraint module and an anti-static environment description module; the medical industry plug-in includes a sterile environment constraint module and a drug specification management module, which are connected to the basic scene description layer through standardized interfaces. (4) Semantic association layer configuration: Establish a semantic mapping between “assembly precision” and “operation precision limit” of the agent, and a semantic association between “sterile requirements” and “cleanliness parameters” of environmental elements to ensure that the scenario description and training requirements are accurately matched.

[0057] In one possible implementation, the complete description file includes a common element encoding segment, a unique element plug-in tag segment, and a semantic tag segment; wherein, the semantic tag segment is used to describe the semantic information of the task.

[0058] Common elements can be encoded using a unified binary code, while unique elements can be extended through plug-in tags, supporting precise encoding and rapid parsing of scene elements. Furthermore, compatible interfaces with simulation engines, data acquisition devices, and training control systems are designed to enable seamless cross-system transfer of scene information.

[0059] The complete description file's three-part encoding structure is built upon four main modules: a core vocabulary module, which defines standardized terminology for static and dynamic common elements; a basic syntax module, which defines the combination rules and constraint expressions for task logic; an encoding format module, which converts the core vocabulary and basic syntax into binary common encoding segments according to bit-segment mapping rules; and a semantic tag module, which encapsulates industry-specific non-standard parameters into personalized plugin tag segments in a plug-in format. Therefore, the common element encoding segments in the three-part structure originate from the core vocabulary and encoding format, the personalized plugin tag segments originate from the semantic tag module, and the semantic tag segments originate from the task semantics description in the basic syntax module.

[0060] Example of encoding format: [Common encoding: agent (size: 1.2m, number of joints: 16)]-[Personalized tag: medical (sterility level: 100)]-[Semantic tag: operation task (dispensing accuracy: ±0.1g)], which can be automatically parsed by the simulation platform to generate a medical dispensing scenario, and at the same time guide the data acquisition equipment to accurately capture the dispensing operation data.

[0061] The method for generating embodied intelligent scene description files provided in this application constructs a layered scene description architecture. It utilizes a basic scene description layer to standardize the representation of common elements across multiple industries, and a pluggable personalized scene element layer to flexibly adapt to the differentiated needs of different industries. Furthermore, it leverages a semantic association layer to automatically bind and propagate industry-specific parameters and constraints. This method can quickly generate complete description files for target industry scenes, effectively solving the problems of fragmented scenes, poor cross-industry adaptability, low reusability, high construction costs, and difficult expansion and maintenance in existing technologies. It significantly improves the cross-industry reusability, adaptation efficiency, and maintainability of scenes, greatly reduces the cost of repeated scene construction, and efficiently supports the rapid construction and iterative optimization of large-scale embodied intelligent training scenes across industries.

[0062] Furthermore, the embodiments of this application accurately extract cross-industry static general elements and form standardized dimensions by collecting multi-industry scenario corpora, extracting candidate elements through named entity recognition, filtering with dual thresholds, and merging semantic clusters. This effectively solves the problems of existing technologies, such as lack of multi-industry general feature system abstraction, strong arbitrariness in element extraction, and insufficient generality. It ensures the representativeness, universality, and standardization of general elements, provides a reliable foundation for unified modeling of cross-industry scenarios, and further enhances the model's cross-industry adaptability and reuse value.

[0063] This application's embodiments obtain industry-specific candidate elements by eliminating common elements, calculate industry saliency scores based on TF-IDF and accurately screen them, and then encapsulate them into standardized plugins according to constraints, tasks, and entity categories. This achieves structured, modular encapsulation and independent management of industry-differentiated elements, effectively solving the problems of tight coupling between industry-specific and general elements, lack of standardized encapsulation, need for redevelopment for cross-industry adaptation, poor reusability, and high maintenance costs in existing technologies. It enables flexible plug-and-play access to industry-differentiated needs, significantly reduces the difficulty of adapting to new industry scenarios, improves scenario expansion efficiency and module reusability, avoids mutual interference between industry-specific and general elements, and ensures the independence and maintainability of scenario modeling.

[0064] This application's embodiments achieve precise alignment at the concept layer through semantic similarity calculation, unified adaptation at the attribute layer through dimensional standardization and boundary constraint matching, and construct reliable semantic mapping relationships by combining consistency verification and conflict resolution. This effectively solves the problems of semantic ambiguity between general elements and industry-specific elements, error-prone parameter adaptation, mismatched constraints, and complex coupling relationships that lead to low adaptation efficiency, error susceptibility, and difficult maintenance in existing technologies. It ensures accurate binding of general parameters and industry-specific information, and consistency of constraint logic, avoiding parameter conflicts or logical contradictions during cross-industry adaptation. This significantly improves the reliability of semantic mapping, the accuracy of parameter binding, and the stability of scene adaptation, further reducing the cost of cross-industry adaptation debugging and enhancing model robustness and scene generation accuracy.

[0065] In one possible implementation, the scenario library can be built in layers, including a basic atom layer, a combined atom layer, a complex task layer, and a general scenario set layer. The basic atom layer includes two independent types of atoms: motion atoms and operation atoms, defining preconditions and constraints between atoms, such as requiring the "grab" atom to have completed its motion atom execution and the embodied intelligent device to reach the target position. The combined atom layer is used to chain motion atoms and operation atoms according to task logic, explicitly defining the temporal dependencies between the two types of atoms (e.g., "move -> position -> grab -> rotate -> place"). When chaining "move-grab," the pose tolerance and chassis stability at the termination of the motion atom are automatically calculated and used as hard physical trigger thresholds for starting the operation atom, forming a physically feasible sub-process combination. The complex task layer is used to chain multiple motion-operation collaborative combinations to form complete business tasks (e.g., breakfast preparation, precision assembly), with the task description including the alternating switching logic of motion paths and operation actions. The general scenario set layer is used to integrate complex tasks by industry, forming a reusable scenario library (e.g., industrial collaboration scenario set, home service scenario set, etc.). Using "motion-operation coordination" as the core reuse unit, multimodal features and physical triggering conditions are timestamped and encapsulated, thereby efficiently supporting cross-industry scenario migration and end-to-end cross-domain generalization training of embodied large models.

[0066] For example, in an industrial bolt assembly scenario: the basic atomic layer contains "linear movement" and "rotation tightening" atoms; the combined atomic layer forms the sub-process of "positioning the workpiece - gripping the bolt - aligning the bolt hole - tightening and fixing"; the complex task layer is integrated into the "engine bolt assembly" task; and finally, it is incorporated into the "automobile manufacturing scenario set" to support robot cross-vehicle adaptation training.

[0067] Based on the above method embodiments, another embodiment of this application provides an apparatus for generating embodied intelligent scene description files, such as... Figure 2 As shown, the device includes: The retrieval unit 210 is used to retrieve standardized parameters of common elements from the basic scenario description layer when it is necessary to construct an embodied intelligence training scenario for a target industry. The basic scenario description layer is used to standardize and represent common elements of at least one dimension extracted from embodied intelligence scenarios of multiple industries. Select loading unit 220, used to select and load corresponding target industry plugins from the personalized scene element layer according to the target industry needs. The personalized scene element layer contains industry plugins in a plug-in manner to describe the personalized needs of each industry. The file generation unit 230 is used to automatically bind and constrain the standardized parameters of the common elements with the personalized demand information described by the target industry plugin through the semantic mapping relationship between the basic scene description layer and the personalized scene element layer pre-established by the semantic association layer, and generate a complete description file of the target industry scene. The constraint propagation refers to applying the constraint conditions defined in the target industry plugin to the standardized parameters of the bound common elements and linking or adjusting the related parameters.

[0068] In one possible implementation, the common elements include static common elements and dynamic common elements; The static common elements include at least one dimension from the following: agent elements, environment elements, object elements, interactive subject elements, and rule elements; and / or, The dynamic common elements include motion atoms and operational atoms used to describe the behavior of embodied intelligent agents.

[0069] In one possible implementation, the device further includes: An acquisition unit is used to acquire the static common elements; The acquisition unit includes: The data acquisition module is used to collect corpora of embodied intelligent scenarios from multiple industries; The extraction module is used to extract candidate elements from the embodied intelligent scene corpus of the multiple industries using named entity recognition technology. The first filtering module is used to filter out candidate elements that meet preset filtering conditions from the candidate elements. The preset filtering conditions include that the cross-industry coverage rate of the candidate elements is greater than or equal to a preset coverage rate threshold, and the weighted word frequency is greater than or equal to a preset word frequency threshold. The clustering convergence module is used to perform semantic clustering and merging of the selected candidate elements, converging them into static common elements with at least one preset dimension.

[0070] In one possible implementation, the common elements include at least one dimension among intelligent agent elements, environmental elements, object elements, interactive subject elements, and rule elements.

[0071] In one possible implementation, the industry plugin includes at least one of an industry-specific constraint description module, a special operation task description module, and a special entity description module.

[0072] In one possible implementation, the device further includes: A plugin generation unit is used to generate the industry plugins; The plugin generation unit includes: The acquisition and elimination module is used to acquire the element set of each industry and eliminate the common elements from the element set to obtain the individual element candidate set of each industry; The calculation module is used to calculate the word frequency-inverse document frequency (TF-IDF) value of each element in the candidate set of personality elements in the corresponding industry corpus, and use the TF-IDF value as the industry saliency score of the element; The second filtering module is used to filter elements in the individual candidate set based on the industry salience score; The encapsulation module is used to encapsulate the selected elements into standardized plug-in modules according to at least one of the following categories: industry-specific constraint class, special operation task class, and special entity class, so as to form the industry plug-in.

[0073] In one possible implementation, the device further includes: A unit is established to pre-establish the semantic mapping relationship between the basic scene description layer and the personalized scene element layer; The establishment unit includes: The first alignment module is used to calculate the semantic similarity between the common elements and the individual needs information described by the industry plugin, and to determine the elements with similarity exceeding a preset similarity threshold as target alignment items in order to complete the concept layer alignment. The second alignment module is used to perform dimensional standardization matching and boundary constraint matching between the common elements and the personalized demand information described by the industry plugin to complete the attribute layer alignment. The dimensional standardization matching is to uniformly convert the dimensions of each attribute in the personalized demand information into the same standard dimensional unit as the common elements. The boundary constraint matching is to match or constrain the value range or boundary conditions of each attribute in the personalized demand information with the preset boundary of the corresponding attribute in the common elements. The verification module is used to perform consistency verification on the alignment results of the concept layer and the alignment results of the attribute layer, and to resolve any consistency conflicts to obtain the semantic mapping relationship.

[0074] In one possible implementation, the complete description file includes a common element encoding segment, a personalized element plug-in tag segment, and a semantic tag segment; wherein the semantic tag segment is used to describe the semantic information of the task.

[0075] The embodied intelligent scene description file generation device provided in this application embodiment constructs a layered scene description architecture. It utilizes a basic scene description layer to achieve standardized representation of common elements across multiple industries, relies on a pluggable personalized scene element layer to flexibly adapt to the differentiated needs of different industries, and uses a semantic association layer to complete the automatic binding and constraint propagation of common parameters and industry-specific information. It can quickly generate complete description files for target industry scenes, effectively solving the problems of fragmented scenes, poor cross-industry adaptability, low reusability, high construction costs, and difficult expansion and maintenance in existing technologies. It significantly improves the cross-industry reusability, adaptation efficiency, and maintainability of scenes, greatly reduces the cost of repeated scene construction, and efficiently supports the rapid construction and iterative optimization of cross-industry large-scale embodied intelligent training scenes.

[0076] Based on the above method embodiments, another embodiment of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the above embodiments.

[0077] Based on the above method embodiments, another embodiment of this application provides an electronic device or computer device, including: One or more processors; The processor is coupled to a storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the electronic device or computer device performs the method as described in any of the above embodiments.

[0078] Based on the above embodiments, another embodiment of this application provides a computer program product, which includes instructions that, when executed on a computer or processor, cause the computer or processor to perform the method described in any of the above embodiments.

[0079] The above-described apparatus embodiments correspond to the method embodiments and have the same technical effects. For detailed descriptions, please refer to the method embodiments. The apparatus embodiments are derived from the method embodiments; detailed descriptions can be found in the method embodiments section, and will not be repeated here. Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application.

[0080] Those skilled in the art will understand that the modules in the apparatus of the embodiments can be distributed in the apparatus of the embodiments as described in the embodiments, or they can be located in one or more devices different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for generating an embodied intelligence scene description file, characterized in that, The method includes: When it is necessary to construct an embodied intelligence training scenario for a target industry, standardized parameters of common elements are retrieved from the basic scenario description layer. The basic scenario description layer is used to standardize the common elements of at least one dimension extracted from embodied intelligence scenarios of multiple industries. Based on the needs of the target industry, select and load the corresponding target industry plugins from the personalized scene element layer. The personalized scene element layer contains industry plugins in a plug-in manner to describe the personalized needs of each industry. By establishing a semantic mapping relationship between the basic scene description layer and the personalized scene element layer in advance through the semantic association layer, the standardized parameters of the common elements are automatically bound and constrained with the personalized demand information described by the target industry plugin, generating a complete description file of the target industry scene. The constrained propagation refers to applying the constraints defined in the target industry plugin to the standardized parameters of the bound common elements and restricting or adjusting the related parameters in a linked manner.

2. The method according to claim 1, characterized in that, The common elements include static common elements and dynamic common elements; The static common elements include at least one dimension from the following: agent elements, environment elements, object elements, interactive subject elements, and rule elements; and / or, The dynamic common elements include motion atoms and operational atoms used to describe the behavior of embodied intelligent agents.

3. The method according to claim 1, characterized in that, The method for obtaining the static common elements includes: Collect embodied intelligence scenario corpora from multiple industries; Candidate elements are extracted from embodied intelligent scene corpora of the aforementioned industries using named entity recognition technology; Candidate elements that meet preset screening conditions are selected from the candidate elements. The preset screening conditions include that the cross-industry coverage rate of the candidate elements is greater than or equal to a preset coverage rate threshold, and the weighted word frequency is greater than or equal to a preset word frequency threshold. The selected candidate elements are semantically clustered and merged to converge into static common elements with at least one preset dimension.

4. The method according to claim 1, characterized in that, The industry plugin includes at least one of the following: an industry-specific constraint description module, a special operation task description module, and a special entity description module.

5. The method according to claim 4, characterized in that, The method for generating the industry plugin includes: Obtain the element set for each industry, and remove the common elements from the element set to obtain the candidate set of individual elements for each industry; Calculate the term frequency-inverse document frequency (TF-IDF) value of each element in the candidate set of personality elements in the corresponding industry corpus, and use the TF-IDF value as the industry saliency score of the element; The elements in the individual candidate set are screened based on the industry saliency score; The selected elements are encapsulated into standardized plug-in modules according to at least one of the following categories: industry-specific constraints, special operation tasks, and special entities, to form the industry plug-in.

6. The method according to claim 1, characterized in that, Establishing a semantic mapping relationship between the basic scene description layer and the personalized scene element layer in advance includes: The semantic similarity between the common elements and the individual needs information described by the industry plugin is calculated, and the elements with similarity exceeding a preset similarity threshold are identified as target alignment items to complete the concept layer alignment. The common elements are matched with the personalized demand information described by the industry plugin using dimensional standardization and boundary constraint matching to complete the attribute layer alignment. Dimensional standardization matching is to uniformly convert the dimensions of each attribute in the personalized demand information into the same standard dimensional unit as the common elements. Boundary constraint matching is to match or constrain the value range or boundary conditions of each attribute in the personalized demand information with the preset boundary of the corresponding attribute in the common elements. A consistency check is performed on the alignment results of the concept layer and the alignment results of the attribute layer, and any existing consistency conflicts are resolved to obtain the semantic mapping relationship.

7. The method according to any one of claims 1-6, characterized in that, The complete description file includes a common element encoding segment, a personalized element plug-in tag segment, and a semantic tag segment; wherein, the semantic tag segment is used to describe the semantic information of the task.

8. A device for generating an embodied intelligent scene description file, characterized in that, The device includes: The retrieval unit is used to retrieve standardized parameters of common elements from the basic scenario description layer when it is necessary to construct an embodied intelligence training scenario for a target industry. The basic scenario description layer is used to standardize and represent common elements of at least one dimension extracted from embodied intelligence scenarios of multiple industries. The selection loading unit is used to select and load the corresponding target industry plugin from the personalized scene element layer according to the target industry needs. The personalized scene element layer contains industry plugins that describe the personalized needs of each industry in a plug-in manner. The file generation unit is used to automatically bind and constrain the standardized parameters of the common elements with the personalized demand information described by the target industry plugin through the semantic mapping relationship between the basic scene description layer and the personalized scene element layer established in advance by the semantic association layer, and generate a complete description file of the target industry scene. The constraint propagation refers to applying the constraint conditions defined in the target industry plugin to the standardized parameters of the bound common elements and linking or adjusting the related parameters.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, The electronic device includes: One or more processors; The processor is coupled to a storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the electronic device performs the method as described in any one of claims 1-7.

Images