CN111459278A - Discrimination method of robot grasping state based on haptic array - Google Patents

Abstract

The present disclosure provides a method for discriminating the grasping state of a robot based on a tactile array, including: step S1: constructing a tactile data set for grasping objects by a robot; step S2: normalizing the data set constructed in step S1; step S3: Constructing a training model based on a multilayer perceptron; step S4: initializing the training model parameters; step S5: performing model training and convergence based on the data set processed in step S2, to obtain the optimal parameters of the model; and step S6: performing an actual grasping operation Complete the discrimination of the grasping state of the robot based on the tactile array. The above method can be applied to the grasping operation of the robot for a variety of objects, according to the grasping force distribution to judge whether the grasping is successful, and adjust the grasping force of the robot in real time to achieve high accuracy The robot grasping operation of high degree, the grasping success rate is over 99%.

Description

Discrimination method of robot grasping state based on haptic array

technical field

The present disclosure relates to the technical fields of machine learning and deep learning, and in particular, to a method for discriminating the grasping state of a robot based on a haptic array.

Background technique

There are two traditional methods for determining the grasping force of a robot: one is to use a preset force threshold method to give the control force that can grasp the object. The setting of the force threshold can be based on experience or based on the robot dynamics to calculate the closed solution of the grasping force; one is to use the sensor to feedback the magnitude of the grasping force to realize the force feedback control of the robot grasping. The use of sensors for grasping force feedback also includes two types, one is indirect grasping force feedback, which is based on a six-dimensional force sensor or joint torque sensor, which directly obtains the six-dimensional force information and joint torque of the robot end effector. information, and then indirectly solve the grasping force of the end effector through robot dynamics. The grasping force data obtained by this method includes the measurement error of the sensor and the error of the dynamic model, which makes the feedback result less accurate and difficult to apply to more accurate grasping operations. One is the direct grasping force feedback. Through the pressure sensor or tactile sensor installed on the robot end effector, the grasping force information of the robot grasping the object can be directly obtained, so as to directly judge whether the grasping is successful according to the feedback force.

The pressure sensor is a single-point judgment, which can only reflect the single-point contact between the robot and the object to be operated. For most of the grasping operations that are in surface contact, it is difficult to perfectly reflect the contact force. Therefore, the single-point pressure sensor can judge grasping The accuracy of the success or not is not high, and it is easy to cause unstable situations such as slipping and rotation. The array tactile sensor can effectively reflect the contact between the robot end effector and the object during grasping, and feed it back to the robot in the form of a data matrix, and judge whether the grasping is successful or not through numerical analysis.

At present, there are grasping discrimination methods, which are divided into two categories: traditional machine learning (support vector machines, etc.) methods and deep convolutional neural network methods. However, most of the data relied on are single-point pressure sensor data, and the data used simultaneously The small size of the set leads to a low discrimination accuracy rate, and the complexity of the deep convolutional neural network leads to a long time for its training and discrimination, and it is difficult to match the control cycle of the robot grasping operation for real-time discrimination. At the same time, since different discrimination methods are based on different pressure sensors or tactile arrays, the scale, measurement method, range and accuracy of each sensor are different, so the discriminant model obtained by this method is not universal. Therefore, a real-time, accurate and efficient grasping discrimination method that can eliminate the physical differences of the haptic array is of great significance for the autonomous grasping of robots.

public content

(1) Technical problems to be solved

Based on the above problems, the present disclosure provides a method for discriminating the grasping state of a robot based on a haptic array, so as to alleviate the low discrimination accuracy in the prior art, and the complexity of the deep convolutional neural network leads to time-consuming training and discrimination. Both are relatively long, it is difficult to coordinate with the control cycle of the robot grasping operation to perform real-time discrimination, and the obtained discriminant model does not have technical problems such as generality.

(2) Technical solutions

The present disclosure provides a method for discriminating the grasping state of a robot based on a haptic array, including:

Step S1: construct a tactile data set for the robot to grasp the object;

Step S2: normalize the data set constructed in step S1;

Step S3: construct a training model based on a multilayer perceptron;

Step S4: Initialize training model parameters;

Step S5: performing model training and convergence based on the data set processed in step S2 to obtain optimal parameters of the model; and

Step S6: Perform an actual grasping operation to complete the discrimination of the grasping state of the robot based on the haptic array.

In the embodiment of the present disclosure, a tactile sensing array is installed on the gripper of the robot arm.

In the embodiment of the present disclosure, the tactile sensing array includes: 16×10 sensing units, and each sensing unit can sense the pressure distribution exerted on the object when the robot grasps the object.

In the embodiment of the present disclosure, the force sensing range of each sensing unit is 0-5N, and the force resolution is 0.1N.

In the embodiment of the present disclosure, in step S1, the types of grasped objects include: apples, baseballs, water bottles, cups, jars, bottle openers, oranges and tennis balls; the successful grasp of each object is taken as a positive sample, Unsuccessful or weak grasps are used as negative samples; each set of samples contains grayscale images of 16×10 sensing units and the corresponding grasping conditions, which can be expressed as [x ₁ , x ₂ ,...x _i , x ₁₆₀ , y], as the input of the model, where x _i represents the pixel value of the gray image of the i-th sensing unit, y is 1 for successful capture, and 0 for unsuccessful or weak capture.

In the embodiment of the present disclosure, in step S2, in order to improve the accuracy and convergence speed of the model, the data of the model is normalized, as shown in formula (1):

Among them, x _i is taken as the original sample data, x _i ^* is the normalized data, μ is the mean of the original sample data, and σ is the standard deviation of the original sample data.

In the embodiment of the present disclosure, in step S3, the multi-layer perceptron model has a total of 4 layers, including an input layer, 2 hidden layers and an output layer; the number of neurons in each layer can be represented by a vector n=(n ₁ n ₂ n ₃ n ₄ ) ^T to represent, where n ₁ = 160 neurons in the input layer, n ₂ = 100 neurons in the first hidden layer, n 3 = 30 neurons in the second hidden layer, and n ₃ = 30 neurons in the output layer n ₄ =1 neuron.

表示第l-1层的第j个神经元到第l层的第i个神经元的权重；

表示第l层第i个神经元的偏倚；

表示第l层第i个神经元的输出；于是各层间的传递关系可以表示为：In the embodiment of the present disclosure, in step S3, the multi-layer perceptron training model is a multi-layer fully connected neural network, and the hidden layer and the output layer are connected with the neurons of the previous layer through linear relationship weights, biases and activation functions; definition

represents the weight of the jth neuron in the l-1th layer to the ith neuron in the lth layer;

represents the bias of the ith neuron in the lth layer;

Represents the output of the i-th neuron in the l-th layer; then the transfer relationship between the layers can be expressed as:

where σ( ) represents the activation function, and its functional form is σ(x)=max(0, x); the output layer outputs values between 0 and 1, and its functional form is f(x)=(1+e ^{- x} ) ^T , to represent the probability of successful grasping by the robot; the closer the output is to 1, the higher the probability of predicting successful grasping based on the multilayer perceptron model.

为标准差的正态分布，即

设定

的初始值为0。In the embodiment of the present disclosure, in step S4, in order to avoid the problems of gradient disappearance and gradient explosion that may be caused during the model training process, the weights and biases of each layer of the model are initialized, and the weights are randomly generated and subject to 0 as the mean value, with 0 as the mean value.

is a normal distribution with standard deviation, that is

set up

The initial value of is 0.

In the embodiment of the present disclosure, in step S5, the loss function of the model is set as logloss, and its expression is:

Among them, N represents the number of samples, M represents the category of the classification problem, y _{i, j} represents 1 when the ith sample belongs to category j, and 0 when it belongs to other categories; p _{i, j} represents that the ith sample is predicted to be classified as j The probability;

Based on the input parameters and initialized weights and biases, forward propagation calculates the output of each layer, and obtains the predicted value y at the output layer; calculates the loss, and obtains the deviation between the predicted value and the actual value, and then solves the loss function through back propagation For the gradients of weights and biases, update the weights and biases until the loss function converges; during the training process, perform L2 regularization on the model, and the loss function at this time is expressed as:

为第l层权重矩阵。where λ is the L2 regular term parameter, L is the number of model layers,

is the weight matrix of the lth layer.

(3) Beneficial effects

It can be seen from the above technical solutions that the method for discriminating the grasping state of a robot based on the haptic array of the present disclosure has at least one or a part of the following beneficial effects:

(1) It can be applied to robot grasping operations for a variety of objects, judge whether the grasping is successful according to the grasping force distribution, and adjust the grasping force of the robot in real time to achieve high-accuracy robot grasping operation, grasping success rate more than 99%;

(2) The training model is simple, the required data set is small, and the accuracy of the training results is high, and it can effectively avoid the gradient disappearance, gradient explosion and overfitting problems in the model training process;

(3) The physical differences of different tactile array sensors can be effectively eliminated through the grayscale image processing of the collected data, so that the model can be applied to the data sets collected by different tactile sensors, and has certain universality.

Description of drawings

FIG. 1 is a schematic flowchart of a method for discriminating a grasping state of a robot based on a haptic array according to an embodiment of the present disclosure.

FIG. 2 is a schematic structural diagram of a method for discriminating a grasping state of a robot based on a haptic array according to an embodiment of the present disclosure.

FIG. 3 is a schematic diagram of a 16×10 haptic array according to an embodiment of the present disclosure.

FIG. 4 is a schematic diagram of a multi-layer perceptron model for grasping determination according to an embodiment of the present disclosure.

Detailed ways

The present disclosure provides a method for judging the grasping state of a robot based on a tactile array. In order to realize the stable grasping operation of the robot, a tactile sensing array is installed on the robot gripper, and the tactile sensing array is used to obtain the grasping state of the robot during the grasping process. The tactile force information is used to establish a data set for different grasping states of multiple types of objects; the robot grasping discrimination problem is regarded as a binary classification problem with tactile information as input, and a training model is constructed based on a multi-layer perceptron; model parameters are obtained through training convergence, In this way, high-precision grasping and discrimination for different objects can be realized. The model has the characteristics of fast training speed and accurate discrimination results, which is superior to other machine learning models and deep learning models.

In order to make the objectives, technical solutions and advantages of the present disclosure clearer, the present disclosure will be further described in detail below with reference to the specific embodiments and the accompanying drawings.

In an embodiment of the present disclosure, a method for discriminating the grasping state of a robot based on a haptic array is provided. With reference to FIGS. 1 to 4 , the method for discriminating the grasping state of a robot based on a haptic array includes:

Step S1: construct a tactile data set for the robot to grasp the object;

Step S2: normalize the data set constructed in step S1;

Step S3: construct a training model based on a multilayer perceptron;

Step S4: Initialize training model parameters;

Step S5: performing model training and convergence based on the data set processed in step S2 to obtain optimal parameters of the model; and

Step S6: Perform an actual grasping operation to complete the discrimination of the grasping state of the robot based on the haptic array.

In step S1, a tactile sensing array is installed on the two-fingered gripper of the robotic arm of the robot, the tactile information of the robot in the grasping process is obtained through the tactile sensing array, and a grasping discrimination algorithm is proposed to judge whether the object is successfully grasped. The robot grasping discrimination problem is regarded as a binary classification problem with the tactile information obtained by the tactile sensor array as the input, and the two categories are grasping success and grasping failure. The definition criterion of successful grasping is that the grasped object does not slip, spin and remain stable within 10s after the grasping is completed.

The tactile sensing array includes: 16×10 sensing units, the array form of which is shown in Figure 3, and each sensing unit can sense the pressure distribution applied to the object when the robot grasps the object. Each unit has a force sensing range of 0-5N and a force resolution of 0.1N. When grabbing an object, the pressure distribution is output as an array in the order from left to right and top to bottom. When collecting data samples, in order to facilitate data unification and subsequent fusion of data sets collected by different arrays, the mapping between pressure perception and array visualization color changes is constructed, and the mapping relationship between force perception range and black-gray-white domain RGB values is established. Eliminating the physical information in the data, the obtained data samples are grayscale images of the array visualization at the time of grabbing.

The types of grabbed objects include: apples, baseballs, water bottles, cups, cans, corkscrews, oranges and tennis balls.

The successful grasping of each object is taken as a positive sample, and the ungraspable and weak grasping situations are taken as negative samples. Each set of samples contains grayscale images of 16 × 10 sensing units and the corresponding grasping conditions, which can be expressed as [x ₁ , x ₂ ,... _xi , x ₁₆₀ , y], as the input to the model. A grasping state takes 20 consecutive frames of data (sampling frequency is 20ms), and each object records positive and negative samples in 15 grasping states. The data of different objects are summarized and deduplicated according to the positive and negative samples, and a total of 1562 samples are obtained. Among them, there are 1085 samples in the training set and 477 samples in the test set.

In step S2, in order to improve the accuracy and convergence speed of the model, the data of the model is normalized. Based on the z-score normalization method, as shown in equation (1), the original data is subtracted from the product of the mean and the standard deviation to obtain the normalization. normalized data.

Among them, x _i is taken as the original sample data, x _i ^* is the normalized data, μ is the mean of the original sample data, and σ is the standard deviation of the original sample data.

In step S3, the multi-layer perceptron model proposed by the present disclosure has 4 layers in total, including an input layer, a hidden layer (2 layers) and an output layer. The model structure is shown in Figure 4. The number of neurons in each layer can be represented by a vector n=(n ₁ n ₂ n ₃ n ₄ ) ^T , where n ₁ = 160 neurons in the input layer, n ₂ = 100 neurons in the first hidden layer, The second layer has n ₃ =30 neurons, and the output layer has n ₄ =1 neuron.

表示第l-1层的第j个神经元到第l层的第i个神经元的权重。

表示第l层第i个神经元的偏倚。

表示第l层第i个神经元的输出。于是各层间的传递关系可以表示为：The model layers are fully connected, and the hidden layer and output layer are connected to the neurons of the previous layer through linear relationship weights, biases and activation functions. definition

represents the weight of the jth neuron in the l-1th layer to the ith neuron in the lth layer.

represents the bias of the ith neuron in the lth layer.

represents the output of the ith neuron in the lth layer. So the transfer relationship between the layers can be expressed as:

Among them, σ(·) represents the activation function. This model uses the ReLU function as the activation function connecting each layer, and its function form is σ(x)=max(0,x). Since the robot grasping discrimination is treated as a two-class problem, the output layer has only one neuron, so the sigmoid is used as the activation function in the output layer to output values between 0 and 1, and its function form is f(x)=(1 +e ^-x ) ^T , to characterize the probability of successful grasping by the robot. The closer the output is to 1, the higher the probability of predicting successful grasping based on the MLP model.

为标准差的正态分布，即

设定

的初始值为0。In step S4, in order to avoid the gradient disappearance and gradient explosion problems that may be caused during the model training process, the weights and biases of each layer of the model are initialized based on He initialization.

is a normal distribution with standard deviation, that is

set up

The initial value of is 0.

In step S5, the loss function of the model is set as logloss, and its expression is:

Among them, N represents the number of samples, M represents the category of the classification problem, y _{i, j} represents 1 when the ith sample belongs to category j, and 0 when it belongs to other categories; p _{i, j} represents that the ith sample is predicted to be classified as j The probability.

After the model parameters are initialized, based on the input parameters and the initialized weights and biases, the output of each layer is calculated by forward propagation, and the predicted value y is obtained by using the sigmoid function at the output layer. Calculate the loss, and get the deviation between the predicted value and the actual value, and then solve the gradient of the loss function to the weight and bias through backpropagation, and use the Adam algorithm to update the weight and bias until the loss function converges. In the training process, the samples are scrambled in each iteration process, and the L2 regularization term is added to the loss function to perform L2 regularization on the model, so as to optimize the model to make ω as small as possible when converging, thereby avoiding the overfitting problem. .

The loss function at this time is expressed as:

为第l层权重矩阵。设置L2正则项参数λ＝0.0001，优化器输入的样本批次数量200，最大迭代步数500，优化公差1e-4。设置Adam算法中学习率α＝0.001，一阶矩估计的指数衰减率β₁＝0.9，二阶矩估计的指数衰减率β₂＝0.999，数值稳定指标ε＝1e^-8。训练结束得到模型最优参数，同时所对应的最优化模型在测试集上预测的准确率能够达到99.74％。进而即可实现高准确度的机器人抓取操作，抓取成功率超过99％。Where λ is the L2 regular term parameter, L is the number of model layers, where L=4.

is the weight matrix of the lth layer. Set the L2 regular term parameter λ=0.0001, the number of sample batches input by the optimizer is 200, the maximum number of iteration steps is 500, and the optimization tolerance is 1e-4. Set the learning rate α=0.001 in the Adam algorithm, the exponential decay rate β ₁ =0.9 of the first-order moment estimation, the exponential decay rate β ₂ =0.999 of the second-order moment estimation, and the numerical stability index ε=1e ⁻⁸ . At the end of training, the optimal parameters of the model are obtained, and the prediction accuracy of the corresponding optimal model on the test set can reach 99.74%. Then high-accuracy robot grasping operation can be realized, and the grasping success rate exceeds 99%.

So far, the embodiments of the present disclosure have been described in detail with reference to the accompanying drawings. It should be noted that, in the accompanying drawings or the text of the description, the implementations that are not shown or described are in the form known to those of ordinary skill in the technical field, and are not described in detail. In addition, the above definitions of various elements and methods are not limited to various specific structures, shapes or manners mentioned in the embodiments, and those of ordinary skill in the art can simply modify or replace them.

Based on the above description, those skilled in the art should have a clear understanding of the structure color imaging structure, testing system and preparation method based on silicon nanopillars of the present disclosure.

To sum up, the present disclosure provides a structure color imaging structure based on silicon nanopillars, a testing system and a preparation method, and a cylindrical periodic array is formed by electron beam lithography and electron beam evaporation deposition method, with high precision; The technology is compatible and easy to integrate; the cylindrical nano-array structure is used to excite the Mie resonance, and according to the change of the refractive index, the change of the structural color caused by the change of the sample can be observed at the same time, which is easy to observe, and is environmentally friendly; when designing the imaging display technology, different arrays Cylindrical structures with different periods or different sizes have different reflectance spectrum characteristics, and researchers can make different cylindrical structures as needed to meet the measurement at different wavelengths.

It should also be noted that the directional terms mentioned in the embodiments, such as "up", "down", "front", "rear", "left", "right", etc., only refer to the directions of the drawings, not used to limit the scope of protection of the present disclosure. Throughout the drawings, the same elements are denoted by the same or similar reference numbers. Conventional structures or constructions will be omitted when it may lead to obscuring the understanding of the present disclosure.

Moreover, the shapes and sizes of the components in the figures do not reflect the actual size and proportion, but merely illustrate the contents of the embodiments of the present disclosure. Furthermore, in the claims, any reference signs placed between parentheses shall not be construed as limiting the claim.

Unless known to the contrary, the numerical parameters set forth in this specification and attached claims are approximations that can vary depending upon the desired properties sought to be obtained from the teachings of the present disclosure. Specifically, all numbers used in the specification and claims to indicate compositional contents, reaction conditions, etc., should be understood as being modified by the word "about" in all cases. In general, the meaning expressed is meant to include a change of ±10% in some embodiments, a change of ±5% in some embodiments, a change of ±1% in some embodiments, and a change of ±1% in some embodiments. Example ±0.5% variation.

Furthermore, the word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements.

The ordinal numbers such as "first", "second", "third", etc. used in the description and the claims are used to modify the corresponding elements, which themselves do not mean that the elements have any ordinal numbers, nor do they Representing the order of a certain element and another element, or the order in the manufacturing method, the use of these ordinal numbers is only used to clearly distinguish an element with a certain name from another element with the same name.

Furthermore, unless the steps are specifically described or must occur sequentially, the order of the above steps is not limited to those listed above, and may be varied or rearranged according to the desired design. And the above embodiments can be mixed and matched with each other or with other embodiments based on the consideration of design and reliability, that is, the technical features in different embodiments can be freely combined to form more embodiments.

Those skilled in the art will understand that the modules in the device in the embodiment can be adaptively changed and arranged in one or more devices different from the embodiment. The modules or units or components in the embodiments may be combined into one module or unit or component, and further they may be divided into multiple sub-modules or sub-units or sub-assemblies. All features disclosed in this specification (including accompanying claims, abstract and drawings) and any method so disclosed may be employed in any combination, unless at least some of such features and/or procedures or elements are mutually exclusive. All processes or units of equipment are combined. Each feature disclosed in this specification (including accompanying claims, abstract and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise. Also, in a unit claim enumerating several means, several of these means can be embodied by one and the same item of hardware.

Similarly, it will be appreciated that in the above description of exemplary embodiments of the disclosure, various features of the disclosure are sometimes grouped together into a single embodiment, figure, or its description. However, this method of disclosure should not be interpreted as reflecting an intention that the claimed disclosure requires more features than are expressly recited in each claim. Rather, as the following claims reflect, disclosed aspects lie in less than all features of a single foregoing disclosed embodiment. Thus, the claims following the Detailed Description are hereby expressly incorporated into this Detailed Description, with each claim standing on its own as a separate embodiment of the present disclosure.

The specific embodiments described above further describe the purpose, technical solutions and beneficial effects of the present disclosure in detail. It should be understood that the above-mentioned specific embodiments are only specific embodiments of the present disclosure, and are not intended to limit the present disclosure. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure should be included within the protection scope of the present disclosure.

Claims (10)

1. A robot grabbing state distinguishing method based on a tactile array comprises the following steps:

step S1: constructing a tactile data set of a robot grabbing object;

step S2: normalizing the data set constructed in the step S1;

step S3: constructing a training model based on a multilayer perceptron;

step S4: initializing training model parameters;

step S5: performing model training and convergence based on the data set processed in the step S2 to obtain optimal parameters of the model; and

step S6: and performing actual grabbing operation to finish judging the grabbing state of the robot based on the tactile array.

2. The method for judging the grabbing state of the robot based on the tactile array according to claim 1, wherein the tactile sensing array is arranged on a mechanical arm paw of the robot.

3. The robot grabbing state discrimination method based on the tactile sensing array as claimed in claim 2, wherein the tactile sensing array comprises 16 × 10 sensing units, and each sensing unit can sense the pressure distribution applied to the object when the robot grabs the object.

4. The method for judging the grabbing state of the robot based on the tactile array as claimed in claim 3, wherein the force sensing range of each sensing unit is 0-5N, and the force resolution is 0.1N.

5. The method for judging the grabbing state of the robot based on the tactile array as claimed in claim 1, wherein in step S1, the grabbed object types include apple, baseball, water bottle, cup, pot, bottle opener, orange and tennis ball, the situation that each object is grabbed successfully is used as a positive sample, the situation that the object is grabbed unsuccessfully or grabbed weakly is used as a negative sample, each group of samples comprises 16 × 10 gray scale images of sensing units and corresponding grabbing situations, and can be represented as [ x ] ×₁，x₂，...x_i，x₁₆₀，y]As input to the model, where x_iAnd the pixel value of the gray image of the ith sensing unit is represented, the successful grabbing y is 1, and the unsuccessful or insecure grabbing y is 0.

6. The method for judging the grasping state of the robot based on the haptic array according to claim 1, wherein in step S2, in order to improve the accuracy and convergence rate of the model, the data of the model is normalized as shown in formula (1):

wherein, x is_iAs sample raw data, x_i ^*For the normalized data, μ is the mean of the original sample data, and σ is the standard deviation of the original sample data.

7. The method for discriminating the grabbing state of a robot based on a tactile array according to claim 1, wherein in step S3, the multi-layer perceptron model has 4 layers including an input layer, 2 hidden layers and an output layer; the number of neurons in each layer may be defined by the vector n ═ n (n)₁n₂n₃n₄)^TIs shown, wherein the layer n is input₁160 neurons, hidden layer one layer n₂100 neurons, hidden layer second layer n₃30 neurons, output layer n₄1 neuron.

8. The method for judging the grabbing state of the robot based on the tactile array according to claim 1, wherein in step S3, the multi-layer perceptron training model is a multi-layer fully-connected neural network, and the hidden layer and the output layer are connected with the neurons in the previous layer through linear relation weights, biases and activation functions; definition of

Representing the weight of the jth neuron of the l-1 layer to the ith neuron of the l layer;

representing the bias of the ith neuron in the l layer;

represents the output of the ith neuron of the l layer; the transfer relationship between the layers can then be expressed as:

whereinσ (·) denotes an activation function, whose functional form is σ (x) ═ max (0, x); the output layer outputs a value between 0 and 1, which is a function of f (x) ═ 1+ e^-x)^TTo represent the probability of successful grabbing of the robot; the closer the output is to 1, the higher the probability of predicting the grabbing success based on the multi-layered perceptron model.

9. The method for judging the grabbing state of a robot based on a haptic array as claimed in claim 1, wherein in step S4, in order to avoid the problem of gradient extinction and gradient explosion that may be caused during the training process of the model, weights and biases are initialized for each layer of the model, and the weights are randomly generated and obeyed to mean 0 to obtain the average value

Is normally distributed with standard deviation, i.e.

Setting up

Is 0.

10. The grasping-state discrimination method of a robot based on a haptic array according to claim 1, wherein in step S5, the loss function of the model is set to loglos, which is expressed as:

where N represents the number of samples, M represents the category of the classification problem, y_i，jThe ith sample is 1 when belonging to the classification j, and belongs to other classifications as 0; p is a radical of_i，jRepresenting the probability that the ith sample is predicted as the j class;

calculating loss and obtaining the deviation between the predicted value and the actual value, further solving the gradient of the loss function to the weight and the bias through back propagation, updating the weight and the bias until the loss function converges, and performing L2 regularization on the model in the training process, wherein the loss function at the moment is expressed as:

where λ is the L2 regularization term parameter, above the number of model layers,

is the l-th layer weight matrix.

Images