Robot body autonomous generation method and system based on artificial intelligence

Abstract

The invention provides a robot body autonomous generation method and system based on artificial intelligence, and the method comprises the steps: carrying out the text coding of the text description of a robot structure, and generating text embedding; according to text embedding, a component generator is used for predicting and generating all components including component types and various attributes through multiple iterations; using a connection relation predictor to predict the connection probability between the components, and screening high-probability connections to form a component connection relation; according to the generated component types and connection relations of the components, father-child connection relations among the components are obtained, the components are connected to father components of the components through joints of the components, various attributes are assigned, a urdf model of the whole robot is formed, and the components in the urdf model are represented by simple geometries assigned by the shape and size attributes of the components; and replacing each simple geometry through a mesh library, and forming a three-dimensional entity model of the whole robot after all the component models are replaced. The robot body can be automatically generated.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for autonomous generation of robot bodies based on artificial intelligence. Background Technology

[0002] Traditional robot design often relies on engineers' experience and manual operation, requiring a long period of manual modeling, optimization, and adjustment. Summary of the Invention

[0003] To address the technical problems existing in the prior art, the present invention provides a method and system for autonomous generation of robot bodies based on artificial intelligence, the technical solution of which is as follows: On the one hand, an artificial intelligence-based method for autonomous generation of robot ontology is provided, which includes: S1. Encode the text description of the robot structure to generate the text embedding; S2. Based on the generated text embedding, use the component generator to predict and generate all components through multiple iterations, including component types and various attributes; S3. Use the connection predictor to predict the connection probability between components, filter high-probability connections, and form component connection relationships. S4. By generating the component types and connection relationships of the components, obtain the parent-child connection relationships between each component. The component is connected to its parent component through its joints and various attributes are assigned to form the URDF model of the entire robot. In the URDF model, each component is represented by a simple geometric shape assigned by its "shape" and "size" attributes. S5. Replace each simple geometry using the mesh library. After all component models have been replaced, a three-dimensional solid model of the entire robot is formed.

[0004] Optionally, the component generator is a fully connected network. The input of the first iteration is the text embedding and the encoding of existing component information. The initial existing component information is empty. The output is the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component of the first component. The first iteration of components is generated, including component type and various attributes. The probability of adding a new component output from the first iteration is compared with a preset threshold. When the probability of adding a new component output from the first iteration is greater than the preset threshold, the second iteration continues. The second iteration encodes the component generated in the first iteration and embeds it together with the text as network input. The output is the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component for the second component. The component for the second iteration is generated, including the component type and various attributes. The probability of adding a new component from the second round of iterations is compared with a preset threshold. After multiple rounds of iterations, all components are finally generated, including component types and various attributes.

[0005] Optionally, the connection predictor is a graph neural network (GNN), which encodes all input components and their attributes into vectors, and initializes the connection relationships to 2 based on the number of components n. (n-1+n-2+n-3+...+1) dimensional vector; The connection probabilities predicted by the GNN output are a (n-1+n-2+n-3+...+1) dimensional vector. High-probability connections are selected to form component connection relationships.

[0006] Optionally, the training process of the component generator includes: Each set of training data in the training set includes: text description, component information: component type, joint type, axis, position, shape, size, and component connection relationship; Transform each set of training data in the training set, including: Encode the text description into a text embedding; Encode the component information, including: encoding the component type, joint type, axis, position, shape, and size of each currently included component; The output includes the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component. Convert each set of training data in the training set into the above input / output format, with the probability represented by 1 or 0; The loss function is designed into three categories: classification loss for component types using cross-entropy loss, regression loss for component attributes using mean squared error loss, and probability loss for new components using binary cross-entropy loss. The total loss is the sum of the above three. The component generator is trained through multiple iterations of the process of loading data, forward propagation, calculating loss, backpropagation, and optimization.

[0007] Optionally, the training process of the connection predictor includes: Each set of training data in the training set includes: text description, component information: component type, joint type, axis, position, shape, size, and component connection relationship; The training data is converted into the following format: node features, including the type and attributes of each component, with the type of each component represented as a one-hot vector and the attributes represented as numerical vectors; edge features, represented by the adjacency matrix of the graph, which is a two-dimensional matrix representing the connection relationship between components. If node i and node j are connected, then there is a pair (i, j) and (j, i) in edge_index. The loss function is designed as a binary cross-entropy loss, which calculates the difference between the connection probability predicted by the model and the actual connection label. The connection prediction engine is trained through multiple iterations of loading data, forward propagation, loss calculation, backpropagation, and optimization.

[0008] On the other hand, an artificial intelligence-based autonomous robot body generation system is provided, the system comprising: The text encoding module is used to encode the text description of the robot structure and generate the text embedding. The component generation module is used to predict and generate all components, including component types and various attributes, based on the generated text embedding and through multiple iterations using a component generator. The connection prediction module is used to predict the connection probability between components using a connection predictor, filter high-probability connections, and form component connection relationships. The URDF model forming module is used to obtain the parent-child connection relationship between various components by generating the component type and connection relationship of the components. The components are connected to their parent components through their joints and various attributes are assigned to form the URDF model of the entire robot. In the URDF model, each component is represented by a simple geometric shape assigned by its "shape" and "size" attributes. The 3D solid model generation module is used to replace various simple geometries using the mesh library. After all component models are replaced, a 3D solid model of the entire robot is formed.

[0009] Optionally, the component generator is a fully connected network. The input of the first iteration is the text embedding and the encoding of existing component information. The initial existing component information is empty. The output is the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component of the first component. The first iteration of components is generated, including component type and various attributes. The probability of adding a new component output from the first iteration is compared with a preset threshold. When the probability of adding a new component output from the first iteration is greater than the preset threshold, the second iteration continues. The second iteration encodes the component generated in the first iteration and embeds it together with the text as network input. The output is the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component for the second component. The component for the second iteration is generated, including the component type and various attributes. The probability of adding a new component from the second round of iterations is compared with a preset threshold. After multiple rounds of iterations, all components are finally generated, including component types and various attributes.

[0010] Optionally, the connection predictor is a graph neural network (GNN), which encodes all input components and their attributes into vectors, and initializes the connection relationships to 2 based on the number of components n. (n-1+n-2+n-3+...+1) dimensional vector; The connection probabilities predicted by the GNN output are a (n-1+n-2+n-3+...+1) dimensional vector. High-probability connections are selected to form component connection relationships.

[0011] Optionally, the training process of the component generator includes: Each set of training data in the training set includes: text description, component information: component type, joint type, axis, position, shape, size, and component connection relationship; Transform each set of training data in the training set, including: Encode the text description into a text embedding; Encode the component information, including: encoding the component type, joint type, axis, position, shape, and size of each currently included component; The output includes the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component. Convert each set of training data in the training set into the above input / output format, with the probability represented by 1 or 0; The loss function is designed into three categories: classification loss for component types using cross-entropy loss, regression loss for component attributes using mean squared error loss, and probability loss for new components using binary cross-entropy loss. The total loss is the sum of the above three. The component generator is trained through multiple iterations of the process of loading data, forward propagation, calculating loss, backpropagation, and optimization.

[0012] Optionally, the training process of the connection predictor includes: Each set of training data in the training set includes: text description, component information: component type, joint type, axis, position, shape, size, and component connection relationship; The training data is converted into the following format: node features, including the type and attributes of each component, with the type of each component represented as a one-hot vector and the attributes represented as numerical vectors; edge features, represented by the adjacency matrix of the graph, which is a two-dimensional matrix representing the connection relationship between components. If node i and node j are connected, then there is a pair (i, j) and (j, i) in edge_index. The loss function is designed as a binary cross-entropy loss, which calculates the difference between the connection probability predicted by the model and the actual connection label. The connection prediction engine is trained through multiple iterations of loading data, forward propagation, loss calculation, backpropagation, and optimization.

[0013] On the other hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, which is loaded and executed by the processor to realize the above-described method for autonomous generation of robot bodies based on artificial intelligence.

[0014] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to realize the above-described method for autonomous generation of robot bodies based on artificial intelligence.

[0015] The beneficial effects of the technical solution provided by this invention include at least the following: This invention can directly transform robot design tasks into computer models, reducing human input and greatly improving design efficiency. By inputting a brief language description, the model can automatically generate a robot body containing various components, connections, and attributes, saving a lot of design time and engineers' energy. Moreover, the end-to-end network of this invention can learn from a large amount of training data and obtain richer design information from the data, thus possessing stronger scalability. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a robot body autonomous generation method based on artificial intelligence provided in an embodiment of the present invention; Figure 2 This is a block diagram of an AI-based autonomous robot body generation system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0019] This invention provides an artificial intelligence-based method for autonomously generating robot bodies. This method can be implemented by an electronic device, which can be a terminal or a server. Figure 1 The diagram shown is a flowchart of the method. The processing flow may include the following steps: S1. Encode the text description of the robot structure to generate the text embedding; In this embodiment of the invention, a pre-trained Transformer model (such as BERT, GPT) is used to process the input text description and generate a text embedding, which is usually a high-dimensional vector (e.g., a 768-dimensional Transformer output).

[0020] Example: Input: Text description "A robot has a main body, two legs, and two joints in each leg". Output: Text embedding (768-dimensional vector) tensor([[ 0.12, -0.08, 0.15, ..., 0.04, 0.22, -0.19]]).

[0021] S2. Based on the generated text embedding, use the component generator to predict and generate all components through multiple iterations, including component types and various attributes; Example: Input: Text embedding (768-dimensional vector) tensor([[ 0.12, -0.08, 0.15, ..., 0.04, 0.22, -0.19]]) Output: [ {"name": "body1", "joint": "free", "axis": (0, 0, 0), "pos": (0, 0,0), "shape": "box", "size": (0.5, 0.3, 0.2)}, ... {"name": "lower_leg2", "joint": "hinge", "axis": (0, 1, 0), "pos": (0, -0.15, -0.3), "shape": "cylinder", "size": (0.1, 0.3)}] Optionally, the component generator is a fully connected network (such as a multilayer perceptron MLP). The input of the first iteration is the text embedding and the encoding of existing component information. The initial existing component information is empty. The output is the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component of the first component. The first iteration of the component is generated, including the component type and various attributes. The probability of adding a new component output from the first iteration is compared with a preset threshold. When the probability of adding a new component output from the first iteration is greater than the preset threshold, the second iteration continues. The second iteration encodes the component generated in the first iteration and embeds it together with the text as network input. The output is the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component for the second component. The component for the second iteration is generated, including the component type and various attributes. The probability of adding a new component from the second round of iterations is compared with a preset threshold. After multiple rounds of iterations, all components are finally generated, including component types and various attributes.

[0022] Example: The first round encodes the text embedding and existing component information as network input (the existing component information is empty), and outputs node_class_pred = tensor([[0.8, 0.15, 0.05]]), joint_pred = tensor([[0.7, 0.15, 0.05, 0.1]]), axis_pred = tensor([[0, 0, 0]]), pos_pred = tensor([[0, 0, 0]]), shape_pred = tensor([[0.6, 0.3, 0.1]]), size_pred = tensor([[0.5, 0.3, 0.2]]), node_exist_pred = tensor([[0.9]]). Here, `node_class_pred` represents the predicted probability of the component type. Assuming there are three component types (body, upper_leg, lower_leg), then [0.8, 0.15, 0.05] indicates the highest probability (0.8) that the current component is body, and the current component is assigned the value "name": "body1", where body1 indicates that there is only one component of this type. Similarly, assuming there are four joint types (free, hinge, slide, fixed), then [0.7, 0.15, 0.05, 0.1] indicates the highest probability (0.7) for free. `axis` is directly assigned the value (0, 0, 0), and the other attributes are similarly assigned. `node_exist_pred` represents the probability of whether to continue generating. If it is less than the preset threshold (0.1), then no further generation is needed. In the current case (0.9 is greater than the preset threshold), a second iteration is required. The components generated in the first iteration are {"name": "body1", "joint": "free", "axis":(0, 0, 0), "pos": (0, The component is encoded (128-dimensional topological encoding) and then used as network input along with text embedding. This will generate the component for the second round of iteration. After multiple rounds of iteration, all components and their attributes can be generated.

[0023] Optionally, the training process of the component generator includes: Each set of training data in the training set includes: text description, component information: component type, joint type, axis, position, shape, size, and component connection relationship; The sample data format is as follows: train_data = [{ "description": "A robot has a main torso, two legs, and two joints in each leg." "components": [ {"name": "body1", "joint": "free", "axis": (0, 0, 0), "pos": (0, 0,0), "shape": "box", "size": (0.5, 0.3, 0.2)}, {"name": "upper_leg1", "joint": "hinge", "axis": (0, 1, 0), "pos": (0, 0.15, 0), "shape": "cylinder", "size": (0.1, 0.3)}, {"name": "lower_leg1", "joint": "hinge", "axis": (0, 1, 0), "pos": (0, 0.15, -0.3), "shape": "cylinder", "size": (0.1, 0.3)}, {"name": "upper_leg2", "joint": "hinge", "axis": (0, 1, 0), "pos": (0, -0.15, 0), "shape": "cylinder", "size": (0.1, 0.3)}, {"name": "lower_leg2", "joint": "hinge", "axis": (0, 1, 0), "pos": (0, -0.15, -0.3), "shape": "cylinder", "size": (0.1, 0.3)}], "edges": [("body1", "upper_leg1"), ("body1", "upper_leg2"), ("upper_leg1", "lower_leg1"), ("upper_leg2", "lower_leg2")]] Transform each set of training data in the training set, including: Encode the text description into a text embedding; Encode the component information, including: encoding the component type, joint type, axis, position, shape, and size of each currently included component; The output includes the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component. Convert each set of training data in the training set into the above input / output format, with the probability represented by 1 or 0; The loss function is designed into three categories: classification loss for component types using cross-entropy loss, regression loss for component attributes using mean squared error loss, and probability loss for new components using binary cross-entropy loss. The total loss is the sum of the above three. The component generator is trained through multiple iterations of the process of loading data, forward propagation, calculating loss, backpropagation, and optimization.

[0024] S3. Use the connection predictor to predict the connection probability between components, filter high-probability connections, and form component connection relationships. Example: enter:[ {"name": "body1", "joint": "free", "axis": (0, 0, 0), "pos": (0, 0,0), "shape": "box", "size": (0.5, 0.3, 0.2)}, ... {"name": "lower_leg2", "joint": "hinge", "axis": (0, 1, 0), "pos": (0, -0.15, -0.3), "shape": "cylinder", "size": (0.1, 0.3)}] Output: [("body1", "upper_leg1"), ("body1", "upper_leg2"), ("upper_leg1", "lower_leg1"), ("upper_leg2", "lower_leg2")]] Optionally, the connection predictor is a graph neural network (GNN), which encodes all input components and their attributes into vectors, and initializes the connection relationships to 2 based on the number of components n. (n-1+n-2+n-3+...+1) dimensional vector; The connection probabilities predicted by the GNN output are a (n-1+n-2+n-3+...+1) dimensional vector. High-probability connections are selected to form component connection relationships.

[0025] Example: The current number of input components is 5 (body1, upper_leg1, lower_leg1, upper_leg2, lower_leg2), so the initial connection relationship is 2. A 10-dimensional vector [[0, 0, 0, 0, 1, 1, 1, 2, 2, 3], [1, 2, 3, 4, 2, 3, 4, 3, 4, 4]] The connection probabilities predicted by the GNN output are a 10-dimensional vector tensor ([[0.92], [0.01], [0.88], [0.03], [0.72], [0.02], [0.12], [0.03], [0.09], [0.86]). The connection probability of component 0 (body1) and component 1 (upper_leg1) is 0.92, the connection probability of component 0 (body1) and component 2 (lower_leg1) is 0.01, the connection probability of component 0 (body1) and component 3 (upper_leg2) is 0.88, and so on. The component connection with the higher connection probability has 4 connection methods: [("body1", "upper_leg1"), ("body1", "upper_leg2"), ("upper_leg1", "lower_leg1"), ("upper_leg2", "lower_leg2")).

[0026] Optionally, the training process of the connection predictor includes: Each set of training data in the training set includes: text description, component information: component type, joint type, axis, position, shape, size, and component connection relationship; The training data is converted into the following format: node features, including the type and attributes of each component, with the type of each component represented as a one-hot vector and the attributes represented as numerical vectors; edge features, represented by the adjacency matrix of the graph, which is a two-dimensional matrix representing the connection relationship between components. If node i and node j are connected, then there is a pair (i, j) and (j, i) in edge_index. The loss function is designed as a binary cross-entropy loss, which calculates the difference between the connection probability predicted by the model and the actual connection label. The connection prediction engine is trained through multiple iterations of loading data, forward propagation, loss calculation, backpropagation, and optimization.

[0027] S4. By generating the component types and connection relationships of the components, obtain the parent-child connection relationships between each component. The component is connected to its parent component through its joints and various attributes are assigned to form the URDF model of the entire robot. In the URDF model, each component is represented by a simple geometric shape assigned by its "shape" and "size" attributes. S5. Replace each simple geometry using the mesh library. After all component models have been replaced, a three-dimensional solid model of the entire robot is formed.

[0028] The Mesh library stores various 3D mesh models of different components and their connection parameters (connection position, joint type, component size, scalability range, etc.) Component attributes are matched against the component model parameters in the mesh library. When a component attribute matches the scaling range of several components, the best-matching component model is scaled up as needed, and the scaled 3D mesh model replaces the current component's simple geometric model. After all component models have been replaced, the entire robot's 3D solid model is obtained.

[0029] like Figure 2 As shown, this embodiment of the invention also provides an artificial intelligence-based autonomous robot body generation system, the system comprising: Text encoding module 210 is used to encode the text description of the robot structure to generate the text embedding; The component generation module 220 is used to predict and generate all components, including component types and various attributes, based on the generated text embedding and through multiple iterations using a component generator. The connection prediction module 230 is used to predict the connection probability between components using a connection predictor, filter high-probability connections, and form component connection relationships. The URDF model forming module 240 is used to obtain the parent-child connection relationship between each component through the component type and connection relationship of the generated components. The component is connected to its parent component through its joints and various attributes are assigned to form the URDF model of the entire robot. In the URDF model, each component is represented by a simple geometric body with the values ​​assigned to its "shape" and "size" attributes. The 3D solid model forming module 250 is used to replace various simple geometries using the mesh library. After all component models are replaced, a 3D solid model of the entire robot is formed.

[0030] Optionally, the component generator is a fully connected network. The input of the first iteration is the text embedding and the encoding of existing component information. The initial existing component information is empty. The output is the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component of the first component. The first iteration of components is generated, including component type and various attributes. The probability of adding a new component output from the first iteration is compared with a preset threshold. When the probability of adding a new component output from the first iteration is greater than the preset threshold, the second iteration continues. The second iteration encodes the component generated in the first iteration and embeds it together with the text as network input. The output is the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component for the second component. The component for the second iteration is generated, including the component type and various attributes. The probability of adding a new component from the second round of iterations is compared with a preset threshold. After multiple rounds of iterations, all components are finally generated, including component types and various attributes.

[0031] Optionally, the connection predictor is a graph neural network (GNN), which encodes all input components and their attributes into vectors, and initializes the connection relationships to 2 based on the number of components n. (n-1+n-2+n-3+...+1) dimensional vector; The connection probabilities predicted by the GNN output are a (n-1+n-2+n-3+...+1) dimensional vector. High-probability connections are selected to form component connection relationships.

[0032] Optionally, the training process of the component generator includes: Each set of training data in the training set includes: text description, component information: component type, joint type, axis, position, shape, size, and component connection relationship; Transform each set of training data in the training set, including: Encode the text description into a text embedding; Encode the component information, including: encoding the component type, joint type, axis, position, shape, and size of each currently included component; The output includes the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component. Convert each set of training data in the training set into the above input / output format, with the probability represented by 1 or 0; The loss function is designed into three categories: classification loss for component types using cross-entropy loss, regression loss for component attributes using mean squared error loss, and probability loss for new components using binary cross-entropy loss. The total loss is the sum of the above three. The component generator is trained through multiple iterations of the process of loading data, forward propagation, calculating loss, backpropagation, and optimization.

[0033] Optionally, the training process of the connection predictor includes: Each set of training data in the training set includes: text description, component information: component type, joint type, axis, position, shape, size, and component connection relationship; The training data is converted into the following format: node features, including the type and attributes of each component, with the type of each component represented as a one-hot vector and the attributes represented as numerical vectors; edge features, represented by the adjacency matrix of the graph, which is a two-dimensional matrix representing the connection relationship between components. If node i and node j are connected, then there is a pair (i, j) and (j, i) in edge_index. The loss function is designed as a binary cross-entropy loss, which calculates the difference between the connection probability predicted by the model and the actual connection label. The connection prediction engine is trained through multiple iterations of loading data, forward propagation, loss calculation, backpropagation, and optimization.

[0034] The robot body autonomous generation system based on artificial intelligence provided in this embodiment of the invention has a functional structure that corresponds to the robot body autonomous generation method based on artificial intelligence provided in this embodiment of the invention, and will not be described again here.

[0035] Figure 3 This is a schematic diagram of the structure of an electronic device 300 provided in an embodiment of the present invention. The electronic device 300 may vary considerably due to different configurations or performance. It may include one or more central processing units (CPUs) 301 and one or more memories 302. The memory 302 stores at least one instruction, which is loaded and executed by the processor 301 to implement the steps of the above-mentioned AI-based autonomous robot body generation method.

[0036] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions that can be executed by a processor in a terminal to complete the aforementioned AI-based autonomous robot body generation method. For example, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device.

[0037] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0038] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for autonomously generating robot bodies based on artificial intelligence, characterized in that, The method includes: S1. Encode the text description of the robot structure to generate the text embedding; S2. Based on the generated text embedding, use the component generator to predict and generate all components through multiple iterations, including component types and various attributes; S3. Use the connection predictor to predict the connection probability between components, filter high-probability connections, and form component connection relationships. S4. By generating the component types and connection relationships of the components, obtain the parent-child connection relationships between each component. The component is connected to its parent component through its joints and various attributes are assigned to form the URDF model of the entire robot. In the URDF model, each component is represented by a simple geometric shape assigned by its "shape" and "size" attributes. S5. Replace each simple geometry using the mesh library. After all component models have been replaced, a three-dimensional solid model of the entire robot is formed.

2. The method according to claim 1, characterized in that, The component generator is a fully connected network. The input of the first iteration is the text embedding and the encoding of existing component information. The initial existing component information is empty. The output is the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component for the first component. The first iteration of components is generated, including component type and various attributes. The probability of adding a new component output from the first iteration is compared with a preset threshold. When the probability of adding a new component output from the first iteration is greater than the preset threshold, the second iteration continues. The second iteration encodes the component generated in the first iteration and embeds it together with the text as network input. The output is the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component for the second component. The component for the second iteration is generated, including the component type and various attributes. The probability of adding a new component from the second round of iterations is compared with a preset threshold. After multiple rounds of iterations, all components are finally generated, including component types and various attributes.

3. The method according to claim 1, characterized in that, The connection predictor is a graph neural network (GNN), which encodes all input components and their attributes into vectors, and initializes the connection relationships to 2 based on the number of components n. (n-1+n-2+n-3+...+1) dimensional vector; The connection probabilities predicted by the GNN output are a (n-1+n-2+n-3+...+1) dimensional vector. High-probability connections are selected to form component connection relationships.

4. The method according to claim 1, characterized in that, The training process of the component generator includes: Each set of training data in the training set includes: text description, component information: component type, joint type, axis, position, shape, size, and component connection relationship; Transform each set of training data in the training set, including: Encode the text description into a text embedding; Encode the component information, including: encoding the component type, joint type, axis, position, shape, and size of each currently included component; The output includes the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component. Convert each set of training data in the training set into the above input / output format, with the probability represented by 1 or 0; The loss function is designed into three categories: classification loss for component types using cross-entropy loss, regression loss for component attributes using mean squared error loss, and probability loss for new components using binary cross-entropy loss. The total loss is the sum of the above three. The component generator is trained through multiple iterations of the process of loading data, forward propagation, calculating loss, backpropagation, and optimization.

5. The method according to claim 1, characterized in that, The training process of the connection predictor includes: Each set of training data in the training set includes: text description, component information: component type, joint type, axis, position, shape, size, and component connection relationship; The training data is converted into the following format: node features, including the type and attributes of each component, with the type of each component represented as a one-hot vector and the attributes represented as numerical vectors; edge features, represented by the adjacency matrix of the graph, which is a two-dimensional matrix representing the connection relationship between components. If node i and node j are connected, then there is a pair (i,j) and (j, i) in edge_index. The loss function is designed as a binary cross-entropy loss, which calculates the difference between the connection probability predicted by the model and the actual connection label. The connection prediction engine is trained through multiple iterations of loading data, forward propagation, loss calculation, backpropagation, and optimization.

6. An AI-based autonomous robot body generation system, characterized in that, The system includes: The text encoding module is used to encode the text description of the robot structure and generate the text embedding. The component generation module is used to predict and generate all components, including component types and various attributes, based on the generated text embedding and through multiple iterations using a component generator. The connection prediction module is used to predict the connection probability between components using a connection predictor, filter high-probability connections, and form component connection relationships. The URDF model forming module is used to obtain the parent-child connection relationship between various components by generating the component type and connection relationship of the components. The components are connected to their parent components through their joints and various attributes are assigned to form the URDF model of the entire robot. In the URDF model, each component is represented by a simple geometric shape assigned by its "shape" and "size" attributes. The 3D solid model generation module is used to replace various simple geometries using the mesh library. After all component models are replaced, a 3D solid model of the entire robot is formed.

7. The system according to claim 6, characterized in that, The component generator is a fully connected network. The input of the first iteration is the text embedding and the encoding of existing component information. The initial existing component information is empty. The output is the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component for the first component. The first iteration of components is generated, including component type and various attributes. The probability of adding a new component output from the first iteration is compared with a preset threshold. When the probability of adding a new component output from the first iteration is greater than the preset threshold, the second iteration continues. The second iteration encodes the component generated in the first iteration and embeds it together with the text as network input. The output is the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component for the second component. The component for the second iteration is generated, including the component type and various attributes. The probability of adding a new component from the second round of iterations is compared with a preset threshold. After multiple rounds of iterations, all components are finally generated, including component types and various attributes.

8. The system according to claim 7, characterized in that, The connection predictor is a graph neural network (GNN), which encodes all input components and their attributes into vectors, and initializes the connection relationships to 2 based on the number of components n. (n-1+n-2+n-3+...+1) dimensional vector; The connection probabilities predicted by the GNN output are a (n-1+n-2+n-3+...+1) dimensional vector. High-probability connections are selected to form component connection relationships.

9. The system according to claim 6, characterized in that, The training process of the component generator includes: Each set of training data in the training set includes: text description, component information: component type, joint type, axis, position, shape, size, and component connection relationship; Transform each set of training data in the training set, including: Encode the text description into a text embedding; Encode the component information, including: encoding the component type, joint type, axis, position, shape, and size of each currently included component; The output includes the component type probability, joint type probability, axis direction vector, position vector, shape type probability, size vector, and probability of adding a new component. Convert each set of training data in the training set into the above input / output format, with the probability represented by 1 or 0; The loss function is designed into three categories: classification loss for component types using cross-entropy loss, regression loss for component attributes using mean squared error loss, and probability loss for new components using binary cross-entropy loss. The total loss is the sum of the above three. The component generator is trained through multiple iterations of the process of loading data, forward propagation, calculating loss, backpropagation, and optimization.

10. The system according to claim 6, characterized in that, The training process of the connection predictor includes: Each set of training data in the training set includes: text description, component information: component type, joint type, axis, position, shape, size, and component connection relationship; The training data is converted into the following format: node features, including the type and attributes of each component, with the type of each component represented as a one-hot vector and the attributes represented as numerical vectors; edge features, represented by the adjacency matrix of the graph, which is a two-dimensional matrix representing the connection relationship between components. If node i and node j are connected, then there is a pair (i,j) and (j, i) in edge_index. The loss function is designed as a binary cross-entropy loss, which calculates the difference between the connection probability predicted by the model and the actual connection label. The connection prediction engine is trained through multiple iterations of loading data, forward propagation, loss calculation, backpropagation, and optimization.

Images