A method for posture reconstruction of soft manipulator based on LSTM neural network

Abstract

The present invention provides a method for reconstructing the posture of a soft manipulator based on an LSTM neural network, comprising: inputting a fiber curvature vector output by the soft manipulator into the input layer of the LSTM neural network; predicting the posture of the soft manipulator in the output layer of the LSTM neural network to obtain a predicted posture vector of the soft manipulator; describing the predicted posture vector in the output layer of the LSTM neural network by means of a probability distribution; constructing a loss function, correcting the probability distribution describing the predicted posture vector, and outputting the posture of the soft manipulator. The present invention describes the predicted posture vector by means of a probability distribution and corrects the probability distribution describing the predicted posture vector, thereby reducing the influence of system errors on the accuracy of the posture reconstruction of the soft manipulator and improving the accuracy of the posture reconstruction of the soft manipulator.

Description

Technical Field

[0001] The present invention relates to the field of electronic information technology, and in particular to a method for reconstructing the posture of a software operating arm based on an LSTM neural network. Background Art

[0002] Fiber-optic shape sensing is currently a cutting-edge technology in the field of optical sensing. Since the late 20th century, with advances in fiber Bragg grating (FBG) manufacturing, FBG sensing has become a hot topic in sensing research. Fiber Bragg grating (FBG) sensors exhibit extremely high sensitivity to changes in temperature and strain. Based on this principle, they can be fabricated to meet diverse measurement needs, such as temperature, pressure, tension, acceleration, displacement, and curvature. Fiber-optic sensors offer advantages such as small size, light weight, high sensitivity, and fast real-time response, making them a promising alternative to traditional sensors in many precision measurement applications. For sensing tasks in soft robotics used in minimally invasive surgery, fiber-optic sensors are resistant to electromagnetic interference and acidic and alkaline environments. Their technology and materials are also harmless to the human body. Multi-point sensing and monitoring can be achieved using a single optical fiber, making them ideally suited to the positioning and sensing requirements of soft manipulators in minimally invasive surgery.

[0003] Due to the unique material and performance advantages of fiber optic sensors, some researchers have begun to apply fiber optic shape sensing technology to the condition monitoring of flexible mechanisms. After more than ten years of development, many technological achievements have been produced.

[0004] In 2004, Clements' team first published a patent for "intelligent shape-sensing cable" technology using multi-core fiber gratings arranged in flexible cables. This cable reconstructs three-dimensional curves based on the curvature and torsion information at the grating points fed back by the central fiber optic sensor. This not only achieves relative positioning and direction calibration of the cable's two end points, but also reconstructs the real-time shape of the entire cable.

[0005] In 2017, a research team led by Wang embedded FBG fiber optic sensors in a linear fashion along the central axis of an elephant-trunk-shaped silicone soft robot. Based on the assumptions of piecewise constant curvature and constant torsion, they monitored the robot's three-dimensional posture, limiting the error to within 1.5 cm. In this work, the team identified three main sources of posture sensing error in FBG fiber optic sensors in soft robots and proposed corresponding optimization strategies:

[0006] (1) Due to the limitations of the manufacturing process of soft robots, the arrangement of optical fibers in soft robots cannot accurately follow geometric assumptions. For example, in this experimental case, there is an error between the actual trajectory of the optical fiber and the geometric center axis of the soft arm.

[0007] (2) Error in shape reconstruction algorithm based on fiber Bragg grating.

[0008] (3) Since the grating points are discretely distributed on the optical fiber, the bending information between two points cannot be perceived. Although the interpolation method is constantly being optimized, the error caused by the lack of original information cannot be eliminated.

[0009] The team proposed methods such as filtering Gaussian errors and adding temperature compensation mechanisms to address the above error sources.

[0010] Although extensive research has been conducted on the application of FBG fiber optic sensors in soft structures, and researchers have proposed corresponding deployment schemes and reconstruction algorithms based on the characteristics and application requirements of various soft structures, FBG fiber optic sensors need to fit tightly with the measured part to obtain accurate deformation data. However, the application scenarios of soft manipulators mostly face unknown interactions with the external environment. Fiber Bragg gratings exposed to the outside are very fragile, and the use of casing encapsulation increases the complexity of the soft arm structure, affecting sensing accuracy. This contradiction in the deployment scheme limits the combination of FBG fiber optic sensors and soft manipulators to laboratory environments, hindering their application in soft manipulators. Therefore, it is necessary to conduct in-depth research on the deployment scheme of FBG fiber optic sensors in soft manipulators to meet the requirements of FBG fiber optic sensors for measuring the soft arm's posture without increasing the complexity of the soft arm structure. Secondly, the high redundancy of the minimally invasive surgical soft arm and its unknown interaction with the external environment have brought difficulties to the establishment of the fiber Bragg grating reconstruction algorithm. Since the grating sensing points of the FBG fiber sensor are discretely distributed, the principle of its shape reconstruction is based on the reconstruction of spatial curves and surfaces based on the curvature and direction of discrete points. One of the drawbacks of this method is that it is insensitive to the position perception between the sensing points. In addition, the irregular deformation and fiber torsion caused by the unknown external environment interaction will cause the posture sensing algorithm based on curvature reconstruction to introduce systematic errors in the raw data stage, affecting the accuracy of the soft arm posture sensing. Summary of the Invention

[0011] In order to solve the technical problem of low accuracy of posture reconstruction of a soft manipulator arm caused by system errors in the prior art, an object of the present invention is to provide a method for posture reconstruction of a soft manipulator arm based on an LSTM neural network, the method comprising:

[0012] In the input layer of the LSTM neural network, the fiber curvature vector output by the soft manipulator is input;

[0013] In the output layer of the LSTM neural network, the posture of the soft manipulator is predicted to obtain the predicted posture vector of the soft manipulator;

[0014] In the output layer of the LSTM neural network, the predicted posture vector is described by probability distribution;

[0015] A loss function is constructed to modify the probability distribution describing the predicted posture vector and output the posture of the soft manipulator.

[0016] In a preferred embodiment, in the input layer of the LSTM neural network, the optical fiber curvature vector output by the soft manipulator is input, and the bending curvature and deflection angle of the soft manipulator collected by the optical fiber are constructed.

[0017] In a preferred embodiment, the predicted posture vector of the soft manipulator is described as follows:

[0018] y′=f ω (s,y),

[0019] Where y′ is the predicted posture vector of the soft manipulator; f ω is the mapping function from the fiber curvature vector output by the soft manipulator to the spatial position of the sensor points on the periphery of the soft manipulator; s is the fiber curvature vector output by the soft manipulator; and y is the air pressure input by the soft manipulator.

[0020] In a preferred embodiment, the predicted posture vector is described by the following probability distribution:

[0021] P(y′|x′)=N(μ(x′),σ(x′)),

[0022] Where y′ is the predicted posture vector of the soft manipulator; x′ is the splicing vector of the fiber curvature vector s output by the soft manipulator and the air pressure y input by the soft manipulator; μ(x′) is the mean of the splicing vector x′; σ(x′) is the variance of the splicing vector x′.

[0023] In a preferred embodiment, the constructed loss function is:

[0024]

[0025] Wherein, T is the time period for inputting air pressure to the soft manipulator 1; x′ is the spliced vector of the fiber curvature vector s output by the soft manipulator and the air pressure y input by the soft manipulator; t and m are variables in the summation formula;

[0026] y′ is the predicted pose vector of the soft manipulator; d is the dimension of the predicted pose vector y′ of the soft manipulator; μ(x′) is the mean of the concatenated vector x′; e is a natural constant;

[0027] Where g = logσ(x′), σ(x′) is the variance of the concatenated vector x′.

[0028] In a preferred embodiment, the soft operating arm has a columnar structure, and three air cavities are formed inside the soft operating arm by evenly dividing the outer cavity wall and the inner cavity wall;

[0029] A wavy convex structure is formed on the outside of the external cavity wall; an optical fiber groove penetrating the soft operating arm is formed in the center of the soft operating arm; and a three-core optical fiber is arranged in the optical fiber groove.

[0030] The present invention provides a method for reconstructing the posture of a soft manipulator based on an LSTM neural network. An FBG fiber sensor is implanted in the central axis of the soft manipulator to reconstruct the peripheral posture of the soft manipulator. The method collects the fiber curvature vector output by the FBG fiber sensor and inputs it into the LSTM neural network to obtain a predicted posture vector for the soft manipulator. The predicted posture vector is described using a probability distribution, and the probability distribution describing the predicted posture vector is corrected. This method reduces the impact of system errors on the accuracy of the soft manipulator's posture reconstruction, thereby improving the accuracy of the soft manipulator's posture reconstruction.

[0031] This invention provides a method for reconstructing the posture of a soft manipulator based on an LSTM neural network. By implanting an FBG fiber optic sensor on the central axis of the soft manipulator, the FBG sensor is separated from the measured surface of the soft manipulator, effectively protecting the FBG sensor from damage caused by external collisions. By placing the FBG fiber optic sensor inside the soft manipulator, external posture changes are transmitted to the internal FBG fiber optic sensor through the deformation of the soft manipulator. Because the soft manipulator buffers and transmits deformation between the measured surface and the FBG fiber optic sensor, the loss of deformation information caused by the discrete distribution of sensing points on the measured surface is reduced.

[0032] This invention provides a method for reconstructing the posture of a soft manipulator based on an LSTM neural network. The FBG fiber sensor employed in this method boasts advantages such as lightness, harmlessness, and stable physical and chemical properties. Its grid parameters can be freely customized to meet the sensing needs of various soft structures. The FBG fiber sensor is separated from the measured surface, and the soft manipulator effectively protects the fiber and reduces the problem of blind spots caused by discrete sensing points. The soft manipulator employed in this invention is simple, easy to integrate and disassemble, and exhibits excellent real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1 The structural diagram of the software operating arm is schematically shown.

[0035] Figure 2 Shown Figure 1 Schematic diagram of the AA direction.

[0036] Figure 3 A cross-sectional schematic diagram of the FBG optical fiber sensor with three core optical fibers according to the present invention is shown.

[0037] Figure 4 A schematic diagram showing the bending of the FBG optical fiber sensor of the three-core optical fiber of the present invention is shown.

[0038] Figure 5 A schematic diagram showing the arrangement of sensing points on the measured surface of the software operating arm of the present invention is shown.

[0039] Figure 6 A schematic diagram of the LSTM neural network of the present invention is shown.

[0040] Figure 7 The figure shows the effect of posture reconstruction of the soft manipulator arm under different air pressure inputs in one embodiment of the present invention.

[0041] Figure 8 A schematic diagram showing the X-direction offset of the end position of a soft manipulator arm during bending of the soft manipulator arm for 5 consecutive minutes in one embodiment of the present invention is shown.

[0042] Figure 9 A schematic diagram showing the Y-direction offset of the end position of a soft manipulator arm tracked for 5 consecutive minutes during the bending process of the soft manipulator arm in one embodiment of the present invention.

[0043] Figure 10 A schematic diagram showing the Z-direction offset of the end position of a soft manipulator arm during bending of the soft manipulator arm for 5 consecutive minutes in one embodiment of the present invention is shown. DETAILED DESCRIPTION

[0044] In order to make the above and other features and advantages of the present invention more clear, the present invention is further described below with reference to the accompanying drawings. It should be understood that the specific embodiments given herein are for the purpose of explaining to those skilled in the art and are only exemplary and not restrictive.

[0045] In order to solve the technical problem of low accuracy of soft manipulator arm posture reconstruction caused by system errors in the prior art, according to an embodiment of the present invention, a method for soft manipulator arm posture reconstruction based on LSTM neural network is provided.

[0046] like Figure 1 The structural diagram of the software operating arm is shown in FIG. Figure 2 shown Figure 1 The schematic diagram in the AA direction of the middle shows a soft operating arm provided by the present invention. The soft operating arm 1 has a columnar structure, and three parallel air cavities 3 are evenly divided inside the soft operating arm 1 by an outer cavity wall 2' and an inner cavity wall 2.

[0047] A wavy convex structure is formed on the outer side of the external cavity wall 2', and an optical fiber groove 5 is formed in the center of the soft operating arm 1 to pass through the soft operating arm 1. A three-core optical fiber (FBG optical fiber sensor) 6 is arranged in the optical fiber groove 5. Figure 3 and Figure 4 The structure of the three-core optical fiber (FBG optical fiber sensor) 6 is shown.

[0048] The air cavity 3 of the soft operating arm 1 has an air inlet 4. When gas is introduced into the air cavity 3 of the soft operating arm 1 through the air inlet 4, a wavy convex structure is formed on the outer side of the external cavity wall 2' and expands, causing the soft operating arm 1 to bend and deform.

[0049] The central axis of the soft manipulator 1 of the present invention has poor axial elongation. Therefore, in some embodiments, it is necessary to ensure that no more than two air cavities 3 are inflated at any one time (three air cavities 3 cannot be inflated simultaneously) to avoid damage to the soft manipulator 1. A three-core optical fiber (FBG fiber sensor) 6 is embedded in the central axis of the soft manipulator 1 to collect curvature information of the soft manipulator 1.

[0050] like Figure 3 The cross-sectional schematic diagram of the FBG optical fiber sensor with three core optical fibers of the present invention is shown in FIG. Figure 4 A schematic diagram of a bending FBG fiber sensor for a three-core fiber according to the present invention is shown. The three-core fiber 6 includes a first core 601, a second core 602, and a third core 603. The first core 601, the second core 602, and the third core 603 surround the central axis of the three-core fiber 6 and are evenly arrayed within the three-core fiber 6. Multiple Bragg gratings (not shown) are arrayed at equal intervals on the first core 601, the second core 602, and the third core 603.

[0051] When the soft manipulator 1 is deformed, the wavelength drift of the Bragg grating of each core in the three-core optical fiber 6 is collected, and the strain of each core in the three-core optical fiber 6 is calculated using the following relationship:

[0052] Δλ B =λ B (1-P e )·ε

[0053] Among them, λ B is the wavelength of the Bragg grating, Δλ B is the wavelength drift of the Bragg grating, P eThe elastic-optic coefficient is related to the inherent characteristics of the Bragg grating, and ε is the strain occurring in each core of the three-core optical fiber 6.

[0054] The bending curvature of each fiber core is calculated by the strain ε generated in each fiber core of the three-core optical fiber 6, specifically by the following method:

[0055]

[0056] Where k is the bending curvature of each fiber core, ε is the strain of each fiber core, and δ is the distance between the center of each fiber core in the curved cross section and the central bending surface of the multi-core fiber when the multi-core fiber is bent.

[0057] Combine Figure 3 and Figure 4 Taking the first core 601 of the three-core optical fiber 6 as an example, when the three-core optical fiber 6 is bent and deformed, the first core 601 is also bent and deformed. When the bent section of the three-core optical fiber 6 is cut, the center of the first core 601 of the three-core optical fiber 6 in the bent section 600 forms a distance δ with the cut surface 605 of the central curved surface 604 of the three-core optical fiber 6.

[0058] The bending strain of the first fiber core 601 satisfies:

[0059]

[0060] Convert the above equation into curvature-strain relationship:

[0061]

[0062] It should be noted that the central curved surface 604 of the triple-core optical fiber 6 is the curved surface where the central axis of the triple-core optical fiber 6 is located when the triple-core optical fiber 6 is bent. When a curved cross section 600 is cut from the curved section, the curved cross section 600 is perpendicular to the tangent plane 605 of the central curved surface 604 of the triple-core optical fiber 6.

[0063] like Figure 3 As shown, in the curved cross section 600 , the distances r between the first core fiber 601 , the second core fiber 602 and the third core fiber 603 and the center O of the three-core optical fiber 6 are each other, and the angles between the first core fiber 601 , the second core fiber 602 and the third core fiber 603 are each 120°.

[0064] The intersection line 605 ′ (neutral axis) of the curved cross section 600 and the tangent plane 605 of the central curved surface 604 of the triple-core optical fiber 6 has deflection angles with the first core fiber 601 , the second core fiber 602 , and the third core fiber 603 , respectively:

[0065]

[0066] Let the first core fiber 601 be a, the second core fiber 602 be b, and the third core fiber 603 be c. Then the strain equations for the three cores at any position can be expressed as:

[0067]

[0068] Among them, E a 、E b 、E c They are the strains of the first core fiber 601, the second core fiber 602 and the third core fiber 603 respectively, and E0 is the strain deviation, which is caused by external force or temperature change.

[0069] Derived from the above formula, the curvature k and deflection angle of the three-core optical fiber 6 are obtained Calculation formula:

[0070]

[0071]

[0072] like Figure 5 The schematic diagram of the arrangement of sensing points on the measured surface of the soft manipulator of the present invention is shown. In the embodiment of the present invention, a plurality of sensing points w are set on the surface of the soft manipulator 1. When the soft manipulator 1 is bent, the curvature k and the deflection angle of the three-core optical fiber 6 are used to measure the sensor points. Characterize the curvature k and deflection angle of the sensing point w on the surface of the soft manipulator 1 In this embodiment, the curvature k and deflection angle of 7 sensing points w are collected as an example.

[0073] According to an embodiment of the present invention, a method for reconstructing the posture of a soft manipulator based on an LSTM neural network is provided, comprising:

[0074] In the input layer of the LSTM neural network, the fiber curvature vector output by the soft manipulator 1 is input.

[0075] like Figure 6 The schematic diagram of the LSTM neural network of the present invention is shown. The LSTM neural network includes an input layer, a hidden layer, and an output layer, with dimensions of 256, 128, and 64, respectively.

[0076] In the input layer of the LSTM neural network, the fiber curvature vector output by the soft manipulator 1 is input. The fiber curvature vector s output by the soft manipulator 1 is the curvature k and deflection angle of the soft manipulator 1 collected by the three-core optical fiber 6. Build.

[0077] In the output layer of the LSTM neural network, the posture of the soft manipulator 1 is predicted to obtain the predicted posture vector y′ of the soft manipulator 1.

[0078] Over a period of time, T, air pressure is applied to the soft manipulator 1, causing it to bend. The fiber curvature vector s, collected during this period, and the air pressure applied to the soft manipulator 1 are fed into the input layer of the LSTM neural network. The air pressure y input by the soft manipulator and the predicted posture vector y′ of the soft manipulator 1 form the state sequence dataset D = {s, y, y′}.

[0079] The predicted pose vector of the soft manipulator 1 is described as follows:

[0080] y′=f ω (s,y),

[0081] Where y′ is the predicted posture vector of the soft manipulator; f ω is the mapping function from the optical fiber curvature vector output by the soft manipulator to the spatial position of the peripheral sensing points of the soft manipulator; s is the three-core optical fiber curvature vector output by the soft manipulator; y is the air pressure input by the soft manipulator.

[0082] In the output layer of the LSTM neural network, the predicted posture vector is described by probability distribution. Specifically, the predicted posture vector is described by the following probability distribution:

[0083] P(y′|x′)=N(μ(x′),σ(x′)),

[0084] Where y′ is the predicted posture vector of the soft manipulator; x′ is the splicing vector of the three-core optical fiber curvature vector s output by the soft manipulator and the air pressure y input by the soft manipulator; μ(x′) is the mean of the splicing vector x′; σ(x′) is the variance of the splicing vector x′.

[0085] For LSTM neural network regression tasks that expect to obtain a mapping between input and output, a gap function between the LSTM neural network output value and the true value, such as the root mean square error (RMSE), is set as the loss function, and the weights are updated through backpropagation to reduce the loss function so that the output value approaches the true value.

[0086] However, for the soft manipulator of the FBG fiber optic sensor, the uncertainty of the soft manipulator structure and the cumulative error of the FBG fiber optic sensor cannot be ignored. That is, the input value of the LSTM neural network has systematic errors and uncertain fluctuations, which affects the accuracy and reliability of the output results.

[0087] In order to reduce the systematic errors and uncertain fluctuations in the input values of the LSTM neural network, which affect the accuracy and reliability of the output results, the present invention describes the predicted posture vector in the form of probability distribution instead of directly outputting the posture of the software manipulator.

[0088] According to an embodiment of the present invention, a loss function is constructed to modify the probability distribution describing the predicted posture vector and output the posture of the soft manipulator.

[0089] Specifically, the loss function is constructed as:

[0090]

[0091] Wherein, T is the time period for inputting air pressure to the soft manipulator 1; x′ is the spliced vector of the fiber curvature vector s output by the soft manipulator and the air pressure y input by the soft manipulator; t and m are variables in the summation formula;

[0092] y′ is the predicted pose vector of the soft manipulator; d is the dimension of the predicted pose vector y′ of the soft manipulator; μ(x′) is the mean of the concatenated vector x′; e is a natural constant;

[0093] Where g = logσ(x′), σ(x′) is the variance of the concatenated vector x′.

[0094] By constructing the above loss function, the probability distribution P(y′|x′) = N(μ(x′), σ(x′)) describing the predicted pose vector y′ is modified, and the mean μ(x′) of the spliced vector x′ and the logarithm of the variance of the spliced vector x′ g = logσ(x′) are output. The pose of the soft manipulator 1 is reconstructed from the output mean μ(x′) of the spliced vector x′ and the logarithm of the variance of the spliced vector x′ g = logσ(x′).

[0095] like Figure 7 The figure shows the effect of reconstructing the posture of the soft manipulator under different air pressures inputted by the soft manipulator in one embodiment of the present invention. Figure 7 (a) to (i) show the effect of reconstructing the posture of the soft manipulator under different air pressure inputs. The mean absolute error (MAE) obtained on the LSTM neural network test set is 3.4mm.

[0096] Figure 8 FIG. 1 is a schematic diagram showing the X-direction offset of the end position of a soft manipulator arm during bending of the soft manipulator arm for 5 consecutive minutes according to one embodiment of the present invention. Figure 9 FIG. 1 is a schematic diagram showing the Y-direction offset of the end position of a soft manipulator arm during bending for 5 consecutive minutes in one embodiment of the present invention. Figure 10The figure shows a schematic diagram of tracking the offset of the end position of the soft manipulator arm in the Z direction for 5 consecutive minutes during the bending process of the soft manipulator arm in one embodiment of the present invention. In the embodiment of the present invention, by inflating the soft manipulator arm 1 for 5 consecutive minutes and tracking the offset of the end position of the soft manipulator arm 1 (the end opposite to the air inlet) in the X, Y and Z directions during the bending process of the soft manipulator arm 1, it can be seen that after describing the predicted posture vector in the form of the probability distribution provided by the present invention and making corrections, the offset of the end position of the soft manipulator arm 1 in the X, Y and Z directions is closer to the true value.

[0097] The present invention maps the curvature vector collected by the FBG fiber optic sensor and the posture of the soft manipulator 1 based on the LSTM neural network, describes the predicted posture vector in the form of probability distribution, and constructs a loss function to correct the probability distribution, so as to output the posture of the soft manipulator 1 in the form of the mean and variance of the Gaussian distribution of the information collected by the FBG fiber optic sensor, thereby reducing the influence of the system error on the accuracy of the posture reconstruction of the soft manipulator and improving the accuracy of the posture reconstruction of the soft manipulator.

[0098] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for reconstructing the posture of a soft manipulator based on an LSTM neural network, characterized in that: The method comprises: (1) In the input layer of the LSTM neural network, the fiber curvature vector output by the soft manipulator is input; (2) In the output layer of the LSTM neural network, the posture of the soft manipulator is predicted to obtain the predicted posture vector of the soft manipulator; The predicted posture vector of the soft manipulator is described as follows: ， in, is the predicted pose vector of the soft manipulator; A mapping function from the fiber curvature vector output by the soft manipulator to the spatial position of the peripheral sensing points of the soft manipulator; is the fiber curvature vector output by the soft manipulator; Air pressure input to the soft manipulator (3) In the output layer of the LSTM neural network, the predicted posture vector is described by the following probability distribution: ， in, is the predicted pose vector of the soft manipulator; Fiber curvature vector output for the soft manipulator Air pressure input from the software manipulator The concatenated stitching vector; is the concatenation vector The mean of is the concatenation vector Variance (4) Constructing a loss function to modify the probability distribution describing the predicted posture vector and output the posture of the soft manipulator; The constructed loss function is: ， Wherein, T is the time period for inputting air pressure to the soft manipulator 1; Fiber curvature vector output for the soft manipulator Air pressure input from the software manipulator The concatenated stitching vector; and is the variable of the summation formula; is the predicted pose vector of the soft manipulator; is the predicted pose vector of the soft manipulator Dimensions; is the concatenation vector The mean of is a natural constant; in, , is the concatenation vector The variance of .

2. The method according to claim 1, characterized in that In the input layer of the LSTM neural network, the fiber curvature vector output by the soft manipulator is input, which is constructed by the bending curvature and deflection angle of the soft manipulator collected by the optical fiber.

3. The method according to claim 1, characterized in that The soft operating arm has a columnar structure, and three air cavities are evenly divided inside the soft operating arm by the outer cavity wall and the inner cavity wall; A wavy convex structure is formed on the outside of the external cavity wall; an optical fiber groove penetrating the soft operating arm is formed in the center of the soft operating arm; and a three-core optical fiber is arranged in the optical fiber groove.

Images