Bionic snake robot snake skin manufacturing method, system and equipment

Abstract

The application relates to a bionic snake robot snake skin manufacturing method, system and equipment, and the method comprises the following steps: a novel lightweight snake scale parameter estimation neural network is constructed and trained in advance; a snake robot parameter is input into a multi-branch one-dimensional feature extraction network for feature extraction to obtain first robust features; the first robust features are input into a lightweight soft threshold attention for feature enhancement to obtain second robust features; the second robust features are input into a deep residual multi-layer perception machine for information integration to obtain optimal snake scale manufacturing parameters; and the optimal snake scale manufacturing parameters are used to process and manufacture best simulation snake scales suitable for the snake robot. Compared with the prior art, the application can adaptively generate snake scale manufacturing parameters according to the size of the snake robot to form the effect of the best simulation real snake scale; and the neural network is used to solve the problems of low fitting degree of the snake scale manufacturing parameters calculated manually with the snake robot, control failure, path planning failure and the like.

Description

Technical Field

[0001] This invention relates to the field of neural network and material preparation technology, and in particular to a method, system and equipment for making snake skin for a biomimetic snake robot. Background Technology

[0002] Snake-like robots, as a type of biomimetic robot, draw their locomotion and morphology from snakes. They hold promise for various applications, such as detection, rescue, surveying, and pipeline inspection, and have therefore attracted considerable attention in robotics research and development. However, they also present a series of technological challenges.

[0003] Firstly, there's the locomotion mechanism of snake-like robots. Their unique movement, such as the wriggling and crawling of a snake, largely relies on friction between the robot's surface and its environment. This friction allows the snake-like robot to move flexibly in various terrains and environments, but it also introduces wear and tear. Prolonged continuous movement or frequent use can cause wear between the robot's surface and the surfaces it contacts, affecting its performance and lifespan. Especially in scenarios requiring long-term task execution, this wear can significantly increase the cost of maintaining and replacing parts for snake-like robots.

[0004] Another concern is the design of existing snake-like robots' shells. Current snake-like robots employ a bowl-shaped structure, with the scales directly printed onto the surface using 3D printing technology. While this method is relatively simple, prolonged movement can cause these scales to wear down, affecting the robot's appearance and usability. Furthermore, replacing the shell for modular snake-like robots is typically a cumbersome process, limiting their practicality in certain applications. Additionally, 3D-printed snake scales do not adaptively generate scale parameters based on the pre-designed snake-like robot dimensions, making it difficult to achieve optimal simulation of realistic snake scales tailored to the robot's structure. Moreover, the scale parameters required for 3D printing are often calculated manually, resulting in poor fit with the snake-like robot and potential problems such as control and path planning failures.

[0005] Therefore, how to manufacture a biomimetic snake-like robot skin with good friction performance and easy replacement is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to overcome the defects of the prior art by providing a method, system and equipment for manufacturing the snake skin of a biomimetic snake robot, which can solve the problems of low surface friction and difficulty in replacement due to wear in biomimetic snake robots.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] According to a first aspect of the present invention, the present invention provides a method for manufacturing the snake skin of a biomimetic snake-like robot, comprising the following steps:

[0009] Obtain the parameters of the snake robot, input the snake robot parameters into a pre-built and trained lightweight snake scale parameter estimation neural network to obtain the optimal snake scale manufacturing parameters, and process and manufacture the best simulated snake scales suitable for the snake robot according to the optimal snake scale manufacturing parameters;

[0010] The lightweight snake scale parameter estimation neural network includes a multi-branch one-dimensional feature extraction network, a lightweight soft-thresholding attention mechanism, and a deep residual multilayer perceptron. By inputting snake robot parameters, the lightweight snake scale parameter estimation neural network first extracts a first robust feature using the multi-branch one-dimensional feature extraction network. This first robust feature is then enhanced by the lightweight soft-thresholding attention mechanism to obtain a second robust feature. This second robust feature is further input into the deep residual multilayer perceptron for information integration, ultimately outputting the optimal snake scale production parameters.

[0011] Preferably, the snake robot parameters include the upper and lower side lengths, height, number of segments, length of each segment, and maximum turning radius of the snake robot, and the snake scale parameters include the longitudinal depth of the snake scale root, scale width, lateral distance between scales, and scale height.

[0012] Preferably, the multi-branch one-dimensional feature extraction network comprises three parallel branches: a one-dimensional convolution Conv1d focusing on local information of the input snake robot parameters, a one-dimensional average pooling AvgPool1d focusing on global average information of the input snake robot parameters, and a one-dimensional max pooling MaxPool1d focusing on global salient information of the input snake robot parameters. The outputs of the three branches are then fused through convolutional operations and fully connected layers, and finally, a loss function is applied to obtain the first robust feature. The processing formula for the input snake robot parameters by the multi-branch one-dimensional feature extraction network is as follows:

[0013] F a =Conv1d(P),

[0014] F b =AvgPool1d(P),

[0015] F c =MaxPool1d(P),

[0016] F1=LeakyReLU(FC(Conv1d(Cat(F a ,F b ,F c ))))

[0017] In the formula, P represents the input snake robot parameters, Conv1d(·) is a one-dimensional convolution, and F... a The first branch output features are represented by AvgPool1d(·), which is a one-dimensional average pooling function. F... b For the output features of the second branch, MaxPool1d(·) is a one-dimensional max pooling function, F c The output features of the third branch are: Cat(·) is the connection operation in the channel dimension, FC(·) is the fully connected layer, LeakyReLU(·) is the LeakyReLU loss function, and F1(·) is the first robust feature.

[0018] Preferably, the loss function LeakyReLU can help the parameters of the lightweight snake scale parameter estimation neural network escape local optima and saddle points during training, thereby solving the problem of gradient backpropagation failure in training the lightweight snake scale parameter estimation neural network and improving the network training effect.

[0019] Preferably, the lightweight soft-threshold attention consists of two parts: an attention mechanism and a soft-threshold filtering mechanism. The first part, the attention mechanism, includes the following steps:

[0020] First, the first robust feature of the input is calculated using three linear mapping layers to obtain Query, Key, and Value respectively. Then, the attention weight map is obtained by multiplying Query and Key element by element to reduce the amount of computation, and normalization is performed using the Softmax function. Finally, the attention mechanism calculation result is obtained by tensor multiplication of Value and the normalized attention weight map.

[0021] The second part is soft-thresholding filtering, where a threshold is calculated from the input first robust feature using neural network parameters. Finally, the obtained soft threshold is used to filter the output of the first attention mechanism, and the original feature information is preserved through a residual structure to obtain the second robust feature. The formula for processing the input first robust feature using the lightweight soft-thresholding attention is as follows:

[0022] Q,K,V = Linear(F1),

[0023] M1=Softmax(ElementMul(Q,K))×V,

[0024] T = SoftThreshold(F1),

[0025] F2 = F1 + Filter(T, M1),

[0026] In the formula, F1 is the first robust feature, Linear(·) is the linear layer, Q, K, V are the Query, Key and Value, ElementMul(·) is the element-wise multiplication, Softmax(·) is the normalization operation, M1 is the result of the attention mechanism, SoftThreshold(·) is the soft threshold operation, T is the result of the soft threshold calculation, Filter(·) is the filtering operation, and F2 is the second robust feature.

[0027] Preferably, the soft threshold neural network part includes a global max pooling operation (GlobalMaxPool), a batch regularization operation (BatchNorm), and a sigmoid activation function. The input first robust feature is processed by the soft threshold neural network to obtain a threshold, thereby filtering out interference information contained in the first robust feature. The processing formula of the soft threshold neural network part for the input first robust feature is as follows:

[0028] T = Sigmoid(BN(GMP(F1))),

[0029] In the formula, F1 is the first robust feature, GMP(·) is the global max pooling operation, BN(·) is the batch regularization operation, Sigmoid(·) is the Sigmoid loss function, and T is the soft threshold calculation result.

[0030] Preferably, the deep residual multilayer perceptron includes a layer regularization layer (LayerNorm), a fully connected layer, a Tanh activation function, a channel gating unit, a residual structure, and a multilayer perceptron. The second robust feature is input and processed through the above structures to integrate feature information. Finally, the multilayer perceptron obtains the optimal snake scale manufacturing parameters based on the snake robot's parameter features. The deep residual multilayer perceptron processes the second robust feature input using the following formula:

[0031] M2 = Tanh(FC(LayerNorm(F2))),

[0032] S = MLP(F2 + FC(CGU(M2))),

[0033] In the formula, F2 is the second robust feature, LayerNorm(·) is the layer regularization, FC(·) is the fully connected layer, Tanh(·) is the Tanh loss function, M2 is the output of the intermediate network layer, CGU(·) is the channel dimension gated unit, MLP(·) is the multilayer perceptron, and S is the prediction result of the snake scale making parameters.

[0034] Preferably, the Channel Gating Unit includes feature channel slicing, layer regularization, channel mapping, and linear computation. By inputting the output of the intermediate network layer, the relevant parameters of the snake robot are mapped to snake scale manufacturing parameters. The Channel Gating Unit processes the input of the intermediate network layer output using the following formula:

[0035] M3,M4 = Split(M2)

[0036] M5=M4·Projection(LayerNorm(M3)),

[0037] In the formula, M2 is the output of the intermediate network layer, Split(·) is the feature channel dimension slicing operation, M3 and M4 are the outputs of the feature channel dimension slicing operation, LayerNorm(·) is the layer regularization operation, Projection(·) is the channel dimension mapping, and M5 is the output of the channel dimension gate unit.

[0038] According to a second aspect of the present invention, the present invention provides a human-computer interaction system specifically for snake scale production using the biomimetic snake-like robot snake skin production method described above, comprising:

[0039] Host equipment: Used to store and manage data, and to transmit instructions to signal processing equipment to network and communication equipment;

[0040] Signal processing equipment: Receives data from network and communication equipment, executes instructions, and collects data to determine the most suitable design solution;

[0041] Network and communication equipment: used to receive instruction data from host equipment, transmit instructions to signal processing equipment, and receive design schemes output by signal processing equipment.

[0042] Preferably, the signal processing device includes one or more processors; a memory; and one or more programs stored in the memory, the one or more programs including instructions for executing the biomimetic snake-like robot snakeskin production method as described above.

[0043] According to a second aspect of the present invention, a computer-readable storage medium is provided, comprising one or more programs executable by one or more processors of an electronic device, the one or more programs comprising instructions for performing the biomimetic snake-like robot snakeskin fabrication method as described above.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) This invention determines snake scale manufacturing parameters through a novel lightweight snake scale parameter estimation neural network, and manufactures matching snake scales for different snake robots, thereby achieving a robust snake scale manufacturing effect. Compared with traditional manual manufacturing, this invention can adaptively generate snake scale manufacturing parameters based on the size of the snake robot to process and manufacture the best simulated snake scales; it solves the problems of poor fit between manually calculated snake scale manufacturing parameters and snake robot, which leads to failure of snake robot control and path planning.

[0046] (2) The novel lightweight snake scale parameter estimation neural network proposed in this invention reduces the number of model parameters and the complexity of the model by making the network lightweight, thereby improving the operating efficiency, reducing the deployment cost, and speeding up the transmission speed, providing a dual guarantee of output and quality for the actual snake scale and snake skin production process. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a method for manufacturing the snake skin of a biomimetic snake-like robot, as provided by the present invention.

[0048] Figure 2 for Figure 1 The schematic diagram of the lightweight snake scale parameter estimation neural network in the embodiment shown is shown.

[0049] Figure 3 for Figure 2 A schematic diagram of the structure of the deep residual multilayer perceptron in the embodiment shown.

[0050] Figure 4 This invention provides a structural schematic diagram of a human-computer interaction system for snake scale production using a biomimetic snake-like robot snakeskin production method.

[0051] Figure 5 For application Figure 1 A photograph of a snake scale produced by the method shown in the embodiment.

[0052] Figure 6 For application Figure 1 The isometric drawing produced by the method of the illustrated embodiment.

[0053] Figure 7 For application Figure 1 An exploded view produced by the method of the illustrated embodiment.

[0054] In the image, 1 represents snake scales, and the snake-shaped robot's overall scales are composed of multiple 1s arranged in an alternating pattern. Detailed Implementation

[0055] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0056] refer to Figure 1 As shown, according to a first aspect of the present invention, this embodiment provides a method for manufacturing the snake skin of a biomimetic snake-like robot, comprising the following steps:

[0057] S1: A novel lightweight neural network for estimating snake scale parameters is pre-built and trained.

[0058] Specifically, refer to Figure 2 As shown, the lightweight snake scale parameter estimation neural network includes a multi-branch one-dimensional feature extraction network, a lightweight soft-threshold attention, and a deep residual multilayer perceptron.

[0059] As an optional implementation, the lightweight snake scale parameter estimation neural network is trained using supervised learning. The training dataset consists of large-scale, high-quality, labeled training samples, and data augmentation techniques are used to expand the training dataset, thereby improving the network's robustness and generalization ability. The supervised training process guides the lightweight snake scale parameter estimation neural network to learn the mapping relationship from snake robot parameters to optimal snake scale production parameter labels, ultimately obtaining a high-performance network model after training.

[0060] S2: Input the snake robot parameters into a multi-branch one-dimensional feature extraction network to extract features and obtain the first robust feature.

[0061] Specifically, the parameters of the snake robot include the top and bottom lengths, height, number of segments, length of each segment, and maximum bending rate.

[0062] refer to Figure 2 As shown, the multi-branch one-dimensional feature extraction network comprises three parallel branches: a one-dimensional convolution Conv1d focusing on local information of the input snake robot parameters, a one-dimensional average pooling AvgPool1d focusing on global average information of the input snake robot parameters, and a one-dimensional max pooling MaxPool1d focusing on global salient information of the input snake robot parameters. The outputs of the three branches are then fused through convolutional operations and fully connected layers, and finally, a loss function is applied to obtain the first robust feature. The processing formula for the input snake robot parameters by the multi-branch one-dimensional feature extraction network is as follows:

[0063] F a=Conv1d(P),

[0064] F b =AvgPool1d(P),

[0065] F c =MaxPool1d(P),

[0066] F1=LeakyReLU(FC(Conv1d(Cat(F a ,F b ,F c ))))

[0067] In the formula, P represents the input snake robot parameters, Conv1d(·) is a one-dimensional convolution, and F... a The first branch output features are represented by AvgPool1d(·), which is a one-dimensional average pooling function. F... b For the output features of the second branch, MaxPool1d(·) is a one-dimensional max pooling function, F c The output features of the third branch are: Cat(·) is the connection operation in the channel dimension, FC(·) is the fully connected layer, LeakyReLU(·) is the LeakyReLU loss function, and F1(·) is the first robust feature.

[0068] As an optional implementation, the loss function LeakyReLU can help the parameters of the lightweight snake scale parameter estimation neural network escape local optima and saddle points during training, thereby solving the problem of gradient backpropagation failure in training the lightweight snake scale parameter estimation neural network and improving the network training effect.

[0069] S3: Input the first robust feature into the lightweight soft-threshold attention for feature enhancement to obtain the second robust feature.

[0070] Specifically, refer to Figure 2 As shown, lightweight soft-threshold attention consists of two parts: the attention mechanism and the soft-threshold filtering. The first part, the attention mechanism, includes the following steps:

[0071] First, the first robust feature of the input is calculated using three linear mapping layers to obtain Query, Key, and Value respectively. Then, the attention weight map is obtained by multiplying Query and Key element by element to reduce the amount of computation, and normalization is performed using the Softmax function. Finally, the attention mechanism calculation result is obtained by tensor multiplication of Value and the normalized attention weight map.

[0072] The second part is soft-thresholding filtering, where a threshold is calculated from the input first robust feature using neural network parameters. Finally, the obtained soft threshold is used to filter the output of the first attention mechanism, and the original feature information is preserved through a residual structure to obtain the second robust feature. The formula for processing the input first robust feature using the lightweight soft-thresholding attention is as follows:

[0073] Q,K,V = Linear(F1),

[0074] M1=Softmax(ElementMul(Q,K))×V,

[0075] T = SoftThreshold(F1),

[0076] F2 = F1 + Filter(T, M1),

[0077] In the formula, F1 is the first robust feature, Linear(·) is the linear layer, Q, K, V are the Query, Key and Value, ElementMul(·) is the element-wise multiplication, Softmax(·) is the normalization operation, M1 is the result of the attention mechanism, SoftThreshold(·) is the soft threshold operation, T is the result of the soft threshold calculation, Filter(·) is the filtering operation, and F2 is the second robust feature.

[0078] As an optional implementation, the soft threshold neural network part includes a global max pooling operation, a batch regularization operation, and a sigmoid activation function. The input first robust feature is processed by the soft threshold neural network to calculate a threshold, thereby filtering out interference information contained in the first robust feature. The processing formula of the soft threshold neural network part for the input first robust feature is as follows:

[0079] T = Sigmoid(BN(GMP(F1))),

[0080] In the formula, F1 is the first robust feature, GMP(·) is the global max pooling operation, BN(·) is the batch regularization operation, Sigmoid(·) is the Sigmoid loss function, and T is the soft threshold calculation result.

[0081] S4: Input the second robust feature into a deep residual multilayer perceptron for information integration to obtain the optimal snake scale production parameters.

[0082] Specifically, the parameters for making snake scales include the longitudinal depth of the snake scale base, the width of the scale, the lateral distance between the scales, and the height of the scale.

[0083] refer to Figure 3 As shown, the deep residual multilayer perceptron includes a layer regularization layer (LayerNorm), a fully connected layer, a Tanh activation function, a channel gating unit, a residual structure, and a multilayer perceptron. The second robust feature is input and processed through the above structures to integrate feature information. Finally, the multilayer perceptron obtains the optimal snake scale manufacturing parameters based on the snake robot's parameter features. The deep residual multilayer perceptron processes the second robust feature input using the following formula:

[0084] M2 = Tanh(FC(LayerNorm(F2))),

[0085] S = MLP(F2 + FC(CGU(M2))),

[0086] In the formula, F2 is the second robust feature, LayerNorm(·) is the layer regularization, FC(·) is the fully connected layer, Tanh(·) is the Tanh loss function, M2 is the output of the intermediate network layer, CGU(·) is the channel dimension gated unit, MLP(·) is the multilayer perceptron, and S is the prediction result of the snake scale making parameters.

[0087] As an optional implementation, the Channel Gating Unit includes feature channel slicing, layer regularization, channel mapping, and linear computation. By inputting the output of the intermediate network layer, the snake robot-related parameters are mapped to snake scale creation parameters. The Channel Gating Unit processes the input intermediate network layer output using the following formula:

[0088] M3,M4 = Split(M2)

[0089] M5=M4·Projection(LayerNorm(M3)),

[0090] In the formula, M2 is the output of the intermediate network layer, Split(·) is the feature channel dimension slicing operation, M3 and M4 are the outputs of the feature channel dimension slicing operation, LayerNorm(·) is the layer regularization operation, Projection(·) is the channel dimension mapping, and M5 is the output of the channel dimension gate unit.

[0091] S5: Determine a flexible material with good friction properties as the material for making snake scales. Use simulation software to model the snake scales based on the optimal snake scale manufacturing parameters output in S4, and further process and manufacture the best simulated snake scales suitable for snake robots.

[0092] To enable the snake-like robot to move with greater power, the scales need to generate significant friction with the ground. Furthermore, it must possess the anisotropic friction characteristic of the scales on a snake's abdomen, ensuring high forward friction and low backward friction to allow for forward movement. The snake scale parameters output from the S4 simulator are recorded and input into simulation software. The design, dimensions, and mounting method of the scales are planned and designed to prepare for the subsequent assembly and integration of the scales with the snake-like robot. Using current advanced engraving and printing technologies, the robot is manufactured according to the results obtained from the simulation software, resulting in... Figure 5 The finished snake scales shown in Figures 6 and 7 are tested to verify whether they meet the design requirements and can provide the friction force needed for the snake robot to crawl.

[0093] According to a second aspect of the invention, reference is made to Figure 4 As shown, this invention provides a human-computer interaction system specifically for snake scale production using the above-described biomimetic snake robot snakeskin production method, comprising:

[0094] Host equipment: Used to store and manage data, and to transmit instructions to signal processing equipment to network and communication equipment;

[0095] Signal processing equipment: Receives data from network and communication equipment, executes instructions, and collects data to determine the most suitable design solution;

[0096] Network and communication equipment: used to receive instruction data from host equipment, transmit instructions to signal processing equipment, and receive design schemes output by signal processing equipment.

[0097] The signal processing device includes one or more processors; a memory; and one or more programs stored in the memory, the one or more programs including instructions for executing the biomimetic snake-like robot snakeskin production method as described above.

[0098] According to a third aspect of the present invention, the present invention provides a computer-readable storage medium comprising one or more programs executable by one or more processors of an electronic device, the one or more programs comprising instructions for performing the biomimetic snake-like robot snake skin manufacturing method as described above.

[0099] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0100] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for manufacturing the snakeskin of a biomimetic snake-shaped robot, characterized in that, The method includes the following steps: obtaining snake robot parameters, inputting the snake robot parameters into a pre-constructed and trained lightweight snake scale parameter estimation neural network to obtain optimal snake scale manufacturing parameters, and processing and manufacturing the best simulated snake scales suitable for the snake robot according to the optimal snake scale manufacturing parameters; The lightweight snake scale parameter estimation neural network includes a multi-branch one-dimensional feature extraction network, a lightweight soft-threshold attention, and a deep residual multilayer perceptron. By inputting snake robot parameters, the lightweight snake scale parameter estimation neural network first extracts a first robust feature using the multi-branch one-dimensional feature extraction network. The first robust feature is then enhanced by the lightweight soft-threshold attention to obtain a second robust feature. The second robust feature is further input into the deep residual multilayer perceptron for information integration, ultimately outputting the optimal snake scale production parameters. The multi-branch one-dimensional feature extraction network comprises three parallel branches: a one-dimensional convolution Conv1d focusing on local information of the input snake robot parameters, a one-dimensional average pooling AvgPool1d focusing on global average information of the input snake robot parameters, and a one-dimensional max pooling MaxPool1d focusing on global salient information of the input snake robot parameters. The outputs of the three branches are then fused through convolutional operations and fully connected layers, and finally, a loss function is applied to obtain the first robust feature. The processing formula for the input snake robot parameters by the multi-branch one-dimensional feature extraction network is as follows: In the formula, To input parameters for the snake robot, For one-dimensional convolution, The first branch outputs features. For one-dimensional average pooling, The output features of the second branch, For one-dimensional max pooling, The output features of the third branch For connection operations at the channel dimension, It is a fully connected layer. The loss function is LeakyReLU. This is the first robust feature; The lightweight soft-threshold attention mechanism comprises two parts: an attention mechanism and a soft-threshold filtering mechanism. The attention mechanism includes the following steps: First, the first robust feature of the input is calculated through three linear mapping layers to obtain Query, Key and Value respectively. Then, the attention weight map is obtained by multiplying Query and Key element by element to reduce the amount of computation. The Softmax function is used for normalization. Finally, the attention mechanism calculation result is obtained by tensor multiplication of Value and the normalized attention weight map. The soft thresholding process includes: calculating a threshold from the input first robust feature using neural network parameters; filtering the output of the first attention mechanism using the obtained soft threshold; and obtaining a second robust feature by preserving the original feature information through a residual structure. The formula for processing the input first robust feature using the lightweight soft thresholding attention is as follows: In the formula, This is the first robust feature. For linear layers, For Query, Key and Value, For element-wise multiplication, For normalization operations, The result is calculated based on the attention mechanism. This is a soft threshold operation. This is the result of soft threshold calculation. For filtering operations, This is the second robust feature; The deep residual multilayer perceptron includes a layer regularization layer (LayerNorm), a fully connected layer, a Tanh activation function, a channel gating unit, a residual structure, and a multilayer perceptron. The second robust feature is input and sequentially processed through the aforementioned structures within the deep residual multilayer perceptron for feature information integration. Finally, the multilayer perceptron obtains the optimal snake scale manufacturing parameters based on the snake robot's parameter features. The deep residual multilayer perceptron processes the input second robust feature using the following formula: In the formula, This is the second robust feature. For layer regularization, It is a fully connected layer. For the Tanh loss function, For the output of the intermediate network layer, For channel-dimensional gating units, It is a multilayer perceptron. Parameter prediction results for snake scales.

2. The method for manufacturing snakeskin for a biomimetic snake-like robot according to claim 1, characterized in that, The snake robot parameters input to the lightweight snake scale parameter estimation neural network include the upper and lower side lengths, height, number of segments, length of each segment, and maximum turning radius of the snake robot. The snake scale manufacturing parameters output by the lightweight snake scale parameter estimation neural network include the longitudinal depth of the snake scale root, scale width, lateral distance between scales, and scale height.

3. The method for manufacturing the snakeskin of a biomimetic snake-like robot according to claim 1, characterized in that, The LeakyReLU loss function is used to help the lightweight snake scale parameter estimation neural network parameters escape local optima and saddle points during training.

4. The method for manufacturing the snakeskin of a biomimetic snake-like robot according to claim 1, characterized in that, The neural network portion of the soft threshold includes a global max pooling operation, a batch regularization operation, and a sigmoid activation function. The input first robust feature is processed by the neural network portion of the soft threshold to calculate a threshold, thereby filtering out interference information contained in the first robust feature. The processing formula for the input first robust feature by the neural network portion of the soft threshold is as follows: In the formula, This is the first robust feature. This is a global max pooling operation. For batch regularization operations, The loss function is Sigmoid. This is the result of the soft threshold calculation.

5. The method for manufacturing the snakeskin of a biomimetic snake-like robot according to claim 1, characterized in that, The Channel Gating Unit includes feature channel slicing, layer regularization, channel mapping, and linear computation. By inputting the intermediate network layer output into the Channel Gating Unit, the snake robot's relevant parameters are mapped to snake scale creation parameters. The Channel Gating Unit processes the input intermediate network layer output using the following formula: In the formula, For the output of the intermediate network layer, This is a slicing operation along the feature channel dimension. The output is the feature channel dimension slicing operation. For layer regularization operations, Channel dimension mapping Output of the channel dimension gating unit.

6. The human-computer interaction system for snake scale production according to claim 1, characterized in that, The aforementioned human-computer interaction system for snake scale production using the biomimetic snake-like robot snakeskin production method as described in any one of claims 1 to 5 includes: Host equipment: Used to store and manage data, and to transmit instructions to signal processing equipment to network and communication equipment; Signal processing equipment: Receives data from network and communication equipment, executes instructions, and collects data to determine the most suitable design solution; Network and communication equipment: used to receive instruction data from host equipment, transmit instructions to signal processing equipment, and receive design schemes output by signal processing equipment.

7. A computer-readable storage medium, characterized in that, It includes one or more programs that are executed by one or more processors of an electronic device, the one or more programs including instructions for performing the biomimetic snake-like robot snake skin manufacturing method as described in any one of claims 1 to 5.

Images