Multidimensional deformation sensing soft robot system, deformation reconstruction method, electronic device and computer-readable storage medium
By using a sensing fiber group symmetrically arranged with alternating left-hand and right-hand spiral fibers, combined with OFDR demodulation and a deep learning model, the problems of multidimensional deformation decoupling and global shape reconstruction error in soft robots are solved, achieving high-precision multidimensional deformation perception and shape reconstruction.
Patent Information
- Application Number
- CN202611418840.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-11
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies struggle to accurately decouple various deformations such as axial stretching, torsion, and bending while maintaining the flexibility and miniaturization of soft robots, and there are cumulative errors in global shape reconstruction.
A sensing fiber optic group consisting of alternating braids of left-handed and right-handed optical fibers arranged in equal and symmetrical quantities, combined with OFDR demodulation and a deep learning model, achieves continuous and stable sensing of multidimensional deformation and acquisition of high-density strain information through temperature compensation and physical model constraints.
It achieves accurate decoupling of multidimensional deformation of soft robots and high-density strain information acquisition, reduces the cumulative error in global shape reconstruction, improves the shape reconstruction accuracy, and is applicable to magnetic resonance environment in complex and confined spaces.
Smart Images

Figure CN123060744A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of soft robot sensing and shape reconstruction technology, and in particular to a multidimensional deformation sensing soft robot system that utilizes symmetrical weaving of left-hand and right-hand optical fibers, distributed strain measurement of optical frequency domain reflection, and fusion of physical models and data-driven models to realize multidimensional deformation detection and global shape reconstruction of soft robots such as axial tension or compression, torsion, and bending; a deformation reconstruction method for multidimensional deformation sensing soft robots; and an electronic device and a computer-readable storage medium having a processor and a computer program and implementing the aforementioned deformation reconstruction method when executing the computer program. Background Technology
[0002] Soft robots, especially flexible continuum robots, are characterized by high compliance, strong environmental adaptability, and the ability to move in narrow or curved channels. They can be used for interventional procedures, endoscopic examinations, and other confined space operations within cavities such as blood vessels and the digestive tract. In these applications, to implement closed-loop control, motion planning, and safety monitoring, it is necessary to continuously acquire the deformation state of the soft robot at various positions along its axis, and determine the spatial shape and end effector position of the soft robot based on the acquired deformation state.
[0003] In existing deformation detection schemes, external visual sensing typically uses cameras mounted on the outside of the soft robot to acquire image information of the robot body, and estimates the robot's posture, shape, and motion state based on changes in contours, marker points, or surface textures in the image. This scheme can provide non-contact deformation observation information under conditions of stable field of view and no occlusion of the robot body.
[0004] However, visual sensing results rely on a direct visual path between the camera and the robot body. When the robot body is partially occluded by surrounding tissues, instruments, cavity walls, or the robot itself, contours, markers, or texture features in the image are easily lost, potentially causing interruptions in feature extraction, matching, and continuous tracking. Especially in environments lacking stable direct visual paths, such as human cavities, external cameras struggle to continuously acquire complete images, leading to discontinuities in the deformation perception link and making it difficult to meet the continuous full-pose perception requirements of soft robots in complex and confined spaces.
[0005] Another sensing approach based on displacement and driving tension typically measures changes in the displacement or tension of the drive cable and combines this with a drive model to infer the bending state and spatial shape of the soft robot. This approach can reuse cable signals from the drive system, thus eliminating the need for separate position sensors on the robot itself.
[0006] However, the displacement and tension of the drive cable primarily reflect the drive input state. The actual deformation of the robot body is also affected by factors such as transmission backlash, cable friction, local slippage, material hysteresis, and external loads. Torsional deformation caused by contact torque, cavity friction, or asymmetric forces is difficult to decouple from the bending response based on the drive-side signal; when the robot experiences axial compression, axial tension, or local deflection caused by non-drive forces, the drive-side signal may not necessarily change accordingly. Therefore, relying solely on the drive cable signal can easily overlook or misjudge the actual deformation, simplifying multidimensional deformation into a single drive response.
[0007] Another electromagnetic sensing approach typically involves embedding electromagnetic elements within the soft robot, measuring the position and orientation of these elements in an external electromagnetic field, and estimating the robot's spatial shape based on data from multiple measurement locations. This approach can provide tracking information for the robot's local position or orientation.
[0008] However, the measurement results of electromagnetic sensors are easily affected by magnetic field distortions caused by surrounding metal instruments, metal support structures, and other magnetic objects, resulting in drift, jumps, or distortions in position or attitude calculations. Electromagnetic components and their connection structures may also be limited by electromagnetic compatibility and usage conditions in a magnetic resonance imaging environment. Furthermore, electromagnetic sensors typically only form measurement points at discrete locations; local bending, stretching, or torsional changes between adjacent measurement points require interpolation estimation, making it difficult to continuously reflect subtle deformations distributed along the robot's length.
[0009] Another fiber Bragg grating sensing scheme typically involves forming Bragg gratings at multiple locations on the fiber and obtaining strain information at corresponding locations based on the changes in the reflection center wavelength of each grating. This information is then used for curvature calculation and shape reconstruction. This scheme features smaller sensing element size, resistance to electromagnetic interference, and ease of embedding into flexible structures.
[0010] However, the setup of Bragg gratings is limited by factors such as available wavelength bandwidth, spectral crosstalk, and grating writing layout, resulting in a discrete distribution of measurement points along the robot's axis. Local sharp bends, compression zones, or concentrated torsion areas located between adjacent measurement points are difficult to measure directly, requiring shape reconstruction through interpolation, which may lead to a loss of local deformation details. Furthermore, some modeling schemes based on fiber Bragg gratings primarily rely on bending strain for inversion, failing to adequately consider the coupling effects of axial tension, axial compression, and torsion on the strain signal, easily mismapping non-bending deformations as curvature changes.
[0011] Furthermore, the existing distributed optical fiber sensing scheme based on optical frequency domain reflection can utilize Rayleigh scattering within the optical fiber to obtain strain information that is continuously or densely distributed along the fiber length, thereby reducing the limitation of spatial resolution on the spacing of discrete grating measurement points.
[0012] However, distributed scattering responses typically involve strain changes caused by bending, axial tension, axial compression, and torsion simultaneously. Existing solutions still largely employ axially arranged or single-helical-winding optical fibers; in single-helical-winding fibers, both axial stretching and torsion can manifest as changes in fiber path length, and coupling between the two is likely to occur. If axial and torsional deformations are not sufficiently decoupled during shape estimation, local strain changes may be misinterpreted as curvature changes, leading to estimation errors in robot length, centerline shape, and end effector position.
[0013] In some distributed fiber optic sensing solutions, the sensing fiber and the soft robot's supporting structure are independent of each other. Relative slippage may occur between the fiber and the robot body, causing the strain measured by the fiber to not accurately represent the actual strain of the robot body. Adding reinforcing components outside the sensing structure would increase radial space occupation, assembly interfaces, and structural complexity, making it difficult to maintain flexibility and miniaturization while simultaneously achieving deformation sensing and body reinforcement.
[0014] Therefore, how to obtain strain information with high axial density distribution while maintaining the flexibility and miniaturization of soft robots, and reliably decouple various deformations such as axial tension or compression, torsion and bending, has become a technical problem that this disclosure urgently needs to solve.
[0015] In addition, how to reduce the accumulated error in global shape reconstruction while preserving the interpretability of the physical model is another technical problem that this disclosure urgently needs to solve. Summary of the Invention
[0016] This disclosure is made to solve the above-mentioned technical problems. Its main purpose is to provide a multidimensional deformation sensing soft robot system and its deformation reconstruction method. By forming an alternating braided sensing fiber group with an equal number of left-hand and right-hand optical fibers arranged symmetrically along the circumference, axial extension, torsion, and bending can form distinguishable common changes, helical differential changes, and circumferential periodic changes in multiple optical fibers, respectively, so as to realize the continuous, stable, and accurate sensing of multidimensional deformation of the soft robot, to obtain strain information with high density distribution along the axial direction, and reliably decouple various deformations such as axial tension or compression, torsion, and bending.
[0017] Another important objective of this disclosure is to provide a multidimensional deformation sensing soft robot system and its deformation reconstruction method. The system uses an axially sliding temperature sensing component to obtain the temperature drift and implement temperature compensation. Then, it uses the end-effector position prediction result to constrain the initial reconstructed contour output by the physical model. This can reduce the cumulative error in global shape reconstruction and improve the accuracy of global shape reconstruction while preserving the interpretability of the physical model.
[0018] According to a first aspect of this disclosure, a multi-dimensional deformation sensing soft robot system is provided, comprising a soft robot body, an optical fiber sensing unit, an OFDR demodulation unit, and a processing unit. These units sequentially form a processing chain for deformation transmission, optical acquisition, temperature compensation, local deformation parameter calculation, and global shape correction.
[0019] The fiber optic sensing unit comprises a sensing fiber optic assembly and a temperature measurement component. The sensing fiber optic assembly includes multiple right-handed and multiple left-handed fibers arranged symmetrically, with each fiber having at least two strands. The fibers belonging to the same direction of rotation are arranged at equal angular intervals along the circumference of the soft robot body. The right-handed and left-handed fibers are staggered circumferentially and alternately spirally braided and wrapped around the outer wall of the soft robot body. Thus, the two directions of rotation of the fibers can form a balanced constraint effect in both the circumferential and spiral directions, providing symmetrical multi-channel observation for different deformation modes.
[0020] Here, we assume there are n right-handed and n left-handed optical fibers, where n is an integer not less than 2. In one embodiment, the circumferential angle of the k-th right-handed optical fiber on the reference cross-section is θ0 + 2πk / n, and the circumferential angle of the k-th left-handed optical fiber on the same reference cross-section is θ0 + π / n + 2πk / n, where k is from 0 to n-1. Thus, optical fibers with the same direction of rotation are spaced 360° / n apart, and the two sets of optical fibers with different directions of rotation are staggered by 180° / n and alternate circumferentially.
[0021] When n is 2, two right-handed optical fibers are arranged opposite each other circumferentially, and two left-handed optical fibers are also arranged opposite each other circumferentially, with the four fibers alternating circumferentially. When n is 3, three optical fibers of the same direction of rotation are arranged at 120° intervals, and the two groups of optical fibers with different directions of rotation are staggered by 60°. When n is 4, four optical fibers of the same direction of rotation are arranged at 90° intervals, and the two groups of optical fibers with different directions of rotation are staggered by 45°. By ensuring that the number of left-handed and right-handed optical fibers is equal and arranged symmetrically, the directional enhancement and additional torsion caused by unbalanced combinations such as one and three, two and three, or three and four fibers can be avoided.
[0022] In a preferred embodiment, the sensing fiber optic group consists of three right-handed single-core fibers and three left-handed single-core fibers arranged symmetrically. The three right-handed fibers are spaced 120° apart, and the three left-handed fibers are spaced 120° apart. Furthermore, the left-handed fibers are offset by 60° relative to adjacent right-handed fibers, resulting in six sensing fibers that are alternately distributed at 60° intervals along the circumference. This arrangement strikes a balance between the number of fibers, circumferential sampling resolution, and redundancy observation accuracy.
[0023] Furthermore, the sensing fiber optic assembly is fixedly attached to the outer wall of the soft robot body so that the sensing fiber optic assembly deforms together with the soft robot body. The sensing fiber optic assembly can be fixed using, for example, medical adhesive, but this disclosure is not limited to this; it can also be fixed using elastic adhesives, elastic coatings, or combinations thereof. By limiting the relative slippage between the sensing fiber optic assembly and the robot body, the consistency of the transfer of body deformation to fiber strain can be improved. Additionally, the cross-braided sensing fiber optic assembly can also serve as a reinforcing structure, improving the tensile and torsional resistance of the robot body.
[0024] Furthermore, the OFDR demodulation unit is communicatively connected to the fiber optic sensing unit. Based on Rayleigh scattering within each sensing fiber, it acquires strain data distributed along the fiber length at a specified spatial resolution (e.g., 0.64 mm), and also acquires the temperature drift of the temperature sensing component. The OFDR demodulation unit or processing unit performs temperature compensation on the distributed strain data of the sensing fiber group based on the temperature drift. Thus, high-density mechanical strain information distributed along the robot's length can be obtained without being limited by the discrete grating spacing. Note that the 0.64 mm value is merely an example; the OFDR demodulation unit has different spatial resolutions, and 0.64 mm is its minimum resolution. Moreover, the smaller the spatial resolution, the more accurate the acquired strain data.
[0025] Under pure tensile or pure compressive conditions, the strains of multiple right-handed and multiple left-handed optical fibers change in the same direction; under pure torsion conditions, the strains of right-handed and left-handed optical fibers change in opposite directions; under pure bending conditions, the strain of each sensing fiber changes periodically with its circumferential position. The processing unit utilizes these differentiated strain characteristics to extract the axial common component, the left-hand and right-handed differential components, and the circumferential periodic component, respectively, to determine the axial tensile or compressive deformation, torsional deformation, and bending deformation.
[0026] Preferably, the temperature sensing component is axially slidably disposed within the central channel of the soft robot body. The temperature sensing component can be a central temperature-sensing optical fiber, with a gap between the central temperature-sensing optical fiber and the inner wall of the central channel allowing for relative axial sliding, or the central temperature-sensing optical fiber can be disposed within a low-friction sleeve. Because the central temperature-sensing optical fiber is not fixedly attached to the robot body, the transmission of mechanical strain generated by its axial expansion and contraction with the body is suppressed, thus enabling a more independent characterization of ambient temperature changes.
[0027] Furthermore, the processing unit communicates with the OFDR demodulation unit and incorporates a physical model and a deep learning model. The processing unit inputs temperature-compensated strain data into the physical model, solves for the curvature, axial stretching, and torsional angle in two directions along the axial direction of each cross-section based on the cross-sectional geometry of the soft robot body, and obtains the initial reconstructed contour of the soft robot through integration. The physical model ensures that the initial reconstructed contour satisfies continuum kinematics and provides physically meaningful priors for subsequent shape correction.
[0028] The physical model can be integrated using the Bishop algorithm to measure curvature, axial stretching, and torsion angle in two directions. Compared to schemes that reconstruct based solely on bending curvature, simultaneously incorporating axial stretching and torsion angle can reduce profile deviations caused by non-bending deformations being mismapped as curvature changes.
[0029] In a specific example where the sensing fiber optic array comprises six fibers, the processing unit inputs the strain distribution, x-direction curvature distribution, y-direction curvature distribution, axial stretching distribution, and torsional distribution of the six fibers as ten-channel time-series data into a deep learning model. The deep learning model can be a 1D-CNN-LSTM neural network formed by combining a one-dimensional convolutional neural network and a long short-term memory network. The one-dimensional convolutional layer extracts the spatial features along the axis at a single time step, while the long short-term memory network captures the deformation correlation features between consecutive time steps and outputs the predicted three-dimensional coordinates of the soft robot's end effector.
[0030] The processing unit uses the predicted 3D coordinates of the end effector as constraints to iteratively refine the initial reconstructed contour. In one implementation, the soft robot body is discretized into a pseudo-rigid body model containing multiple ball joints. The optimization problem is solved using the L-BFGS algorithm or an L-BFGS-B algorithm suitable for box constraints, and the global shape reconstruction result is obtained through cubic spline interpolation. By implementing corrections within the physical prior neighborhood, material nonlinearity, manufacturing and assembly errors, and piecewise integration cumulative errors can be compensated for while preserving the physical interpretability of the global shape.
[0031] Both the fiber optic sensing unit and the soft robot body can be made of non-metallic materials to improve the system's adaptability to magnetic resonance imaging environments. The helix angle of the sensing fiber can be determined based on the axial length and cross-sectional diameter of the soft robot body; furthermore, it can be affected by the number of braid turns or the braid pitch.
[0032] According to a second aspect of this disclosure, a deformation reconstruction method for a multidimensional deformation-sensing soft robot is provided, comprising: a step of acquiring distributed strain data of a sensing fiber optic group and temperature drift of a temperature-sensing component; a step of temperature compensation based on the temperature drift to compensate the distributed strain data; a step of initial reconstruction using a physical model to solve for the deformation parameters of each cross section and integrating to obtain an initial reconstructed contour; a step of end prediction using the strain data of the sensing fiber optic group and the deformation parameters output by the physical model as multi-channel time-series data to input into a deep learning model to obtain an end-effector three-dimensional coordinate prediction result; and a step of global reconstruction using the end-effector three-dimensional coordinate prediction result as a constraint to iteratively correct the initial reconstructed contour to obtain a global shape reconstruction result.
[0033] According to a third aspect of this disclosure, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the various steps of the above-described deformation reconstruction method.
[0034] Similarly, according to a fourth aspect of this disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, performs the various steps of the above-described deformation reconstruction method.
[0035] [Technical Effects]
[0036] When helical optical fibers used for sensing are fixed to the outer wall of a soft robot body and simultaneously reinforce the robot body, the fiber arrangement alters the robot's equivalent stiffness. If the number of left-handed and right-handed optical fibers differs, or if multiple fibers in the same helix direction are not arranged symmetrically along the circumference, the constraint effect on the robot body at different circumferential positions and in different helix directions becomes unbalanced. This imbalance can easily induce undesirable additional torsion when the robot bends, causing the actual deformation to deviate from the expected bending state and further increasing the difficulty of decoupling bending, torsion, and axial extension / retraction signals.
[0037] In this disclosure, compared to single-helical-winding optical fibers or optical fibers with unequal numbers and different helical directions, a common mode and a differential mode are formed by using an equal number of left-hand and right-hand helical optical fibers arranged symmetrically. Axial tension or compression mainly forms common strain in the two sets of optical fibers, torsion mainly forms opposite strain in the two sets of optical fibers, and bending forms periodic strain related to the circumferential position. Therefore, axial tension, torsion, and bending can be simultaneously identified and decoupled in the same optical fiber sensing structure.
[0038] Furthermore, in this disclosure, optical fibers with the same direction of rotation are arranged at equal angular intervals along the circumference, while optical fibers with two different directions of rotation are staggered and alternately arranged, so that the constraint effects on the soft robot in the circumferential direction and the two directions of rotation tend to be balanced. This can reduce the additional torsion caused by unequal number or asymmetrical position of sensing optical fibers during bending, and reduce the structural coupling between bending and torsion signals.
[0039] Furthermore, in this disclosure, the sensing fiber optic assembly is fixedly attached to the outer wall of the soft robot body. It can deform together with the body to obtain distributed strain, and can also serve as a braided reinforcement structure to improve the tensile and torsional resistance of the soft robot body. There is no need to set up a separate reinforcement structure that occupies a large space or a bulky sensor.
[0040] In existing technologies, after obtaining local deformation parameters, a single physical modeling scheme typically calculates local curvature, axial stretching, and torsion angle based on the geometric relationships and material parameters of the soft robot's cross-section, and then integrates segment by segment along the robot's length to obtain the robot's centerline profile or end effector position. This scheme enables the reconstruction results to satisfy continuum kinematics relationships and has relatively clear physical meaning.
[0041] However, physical models are typically based on ideal material properties, ideal cross-sectional geometry, and ideal sensor arrangement. In actual soft robots, variations in material parameters, local structural differences, adhesive layer thickness variations, changes in the bonding state of sensing fibers, and material losses may occur during manufacturing, assembly, and use, causing the mapping relationship from strain to curvature, extension, or torsion to deviate from the ideal model. Since shape reconstruction requires segment-by-segment integration along the length direction, measurement or model errors from preceding segments will propagate to subsequent segments and gradually accumulate, ultimately resulting in errors in the reconstruction of the global contour and end-effector position. Furthermore, while data-driven models can fit the mapping relationship between sensor signals and shape using calibration data, their output may lack constraints from continuum kinematics and typically require a large number of well-covered training samples. When the training data does not cover specific material states, load states, or complex composite deformations, the generalization ability and output stability of the data-driven model may still be affected.
[0042] In this disclosure, the central temperature-sensing fiber can slide axially within the hollow central channel, reducing the mechanical strain transmission of the fiber to the body due to axial expansion and contraction, making it more suitable for obtaining temperature drift. Accordingly, temperature compensation is applied to the peripheral sensing fiber, which can improve the accuracy of multidimensional deformation parameter calculation.
[0043] Furthermore, the physical model provides a priori contours that satisfy the kinematic relationships of a continuum for the global shape, while the data-driven model provides compensation information for manufacturing and assembly errors, material nonlinearities, and integral accumulation errors. Simultaneously, constraining the initial reconstructed contour with the predicted 3D coordinates of the end face can reduce global contour and end face position errors while preserving physical interpretability.
[0044] In addition, in this disclosure, the system body and fiber optic sensing structure can be made entirely of non-metallic materials, and the fiber optic cable has the ability to resist electromagnetic interference, thus making it suitable for medical intervention, magnetic resonance compatible and other confined space scenarios. Attached Figure Description
[0045] Figure 1 This is a schematic diagram illustrating the overall configuration of a multidimensional deformation sensing soft robot system according to one embodiment of the present disclosure.
[0046] Figure 2 (a) and (b) are schematic diagrams illustrating the braided optical fiber arrangement and AA cross-section of the soft robot body in a multidimensional deformation sensing soft robot system according to an embodiment of the present disclosure.
[0047] Figure 3 Figures (a) to (c) illustrate the sensing principle of a multidimensional deformation sensing soft robot system according to an embodiment of the present disclosure, wherein (a) shows the strain variation under tension or compression, (b) shows the strain variation under torsion, and (c) shows the strain variation under bending.
[0048] Figure 4 This is a schematic diagram of a global shape reconstruction process that integrates a physical model and a data-driven model according to one embodiment of this disclosure.
[0049] Figure 5 This is a flowchart of a deformation reconstruction method for a multidimensional deformation sensing soft robot according to one embodiment of the present disclosure.
[0050] (Symbol Explanation)
[0051] 1. Robotic arm; 2. Drive unit; 3. Soft robots; 3a body outer wall; 4 OFDR demodulators; 30 fiber optic sensing modules; 31. Central passageway; 32-channel central temperature measurement fiber optic cable; 33. Right-handed optical fiber; 34. Left-handed optical fiber; 41. Flexible continuum robot; 42. Input features; 43 convolutional layers; 44 Pooling layer; 45. Flattened layer; 46 LSTM layers; 47. Fully connected layer; 48. End position output; 49. Physical Model; 491 Initial reconstructed outline; 492. Optimize shape; 50. Constraints / Feedback. Detailed Implementation
[0052] The embodiments of this disclosure will be further described below with reference to the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure. Technical features in the various embodiments can be combined with each other without contradiction.
[0053] In this specification, "left-handed" and "right-handed" refer to the direction of view relative to the axial direction of the soft robot body and the predetermined helical direction. The names can be interchanged when the viewing direction changes, but the opposite helical chirality of the two sets of optical fibers remains unchanged. "Soft robot" includes continuous robots formed from elastomers, flexible polymers, or flexible composite materials, as well as multi-segment continuous devices formed by multiple flexible segments connected in series. "Fixed bonding" means that the relative slippage between the sensing fiber optic group and the robot body is limited to a level that allows reliable transmission of deformation of the body to the sensing fiber optic group; it does not require the same continuous bonding method in all positions. "Communication connection" includes wired, fiber optic, and wireless connections capable of transmitting optical signals or data.
[0054] First, combined Figures 1 to 3 The multidimensional deformation sensing soft robot system disclosed herein is described.
[0055] like Figure 1 As shown, the multidimensional deformation sensing soft robot system includes a soft robot 3, an optical fiber sensing unit (sensing optical fiber group 30 and central temperature measuring optical fiber 32), an OFDR demodulation unit 4, and a processing unit. Figure 1In the multidimensional deformation sensing soft robot system shown, the robotic arm 1 is used to support or move the drive unit 2. The drive unit 2 is connected to the soft robot 3 and is used to drive the soft robot 3 to produce deformations such as bending, axial movement, or torsion. The OFDR (Optical Frequency Domain Reflective) demodulation unit 4 is optically connected to the fiber optic sensing unit set on the soft robot 3, such as an OFDR demodulator. The processing unit can be set in the OFDR demodulation unit 4 (OFDR demodulator) or implemented by a host computer, workstation, or embedded controller that is communicatively connected to the OFDR demodulation unit 4 (OFDR demodulator).
[0056] In one specific embodiment, the soft robot 3 includes a soft robot body made of flexible material. The soft robot body is made of flexible resin using an SLA photopolymerization process, with an outer diameter of approximately 6 mm and an axial length of approximately 77.5 mm. Three nickel-titanium drive cables are arranged circumferentially at 120° intervals within the soft robot body, and the soft robot 3 can be bent by changing the displacement or tension of each drive cable. The above dimensions, materials, and drive methods are merely examples; the soft robot 3 can also be formed using silicone rubber casting, thermoplastic elastomer molding, or braided tube composite molding, and can be driven by pneumatics, hydraulics, or other drive methods capable of causing a predetermined deformation of the flexible body.
[0057] like Figure 2 As shown in (a), a sensing fiber optic group 30 is provided on the outer wall of the soft robot 3. The sensing fiber optic group 30 includes three right-handed helical fibers 33 and three left-handed helical fibers 34 arranged symmetrically. The right-handed fibers 33 and the left-handed fibers 34 intersect each other on the surface of the soft robot 3, thereby forming a mesh-like braided structure. The three right-handed fibers 33 are spaced 120° apart circumferentially on any reference cross-section (e.g., along section AA), and the three left-handed fibers 34 are also spaced 120° apart circumferentially on the same reference cross-section, and the two groups of fibers are staggered by 60°. Therefore, on the reference cross-section, the six sensing fibers are distributed alternately and substantially uniformly along the circumference. This symmetrical arrangement provides balanced strain observation at different circumferential positions on the one hand, and balances the enhanced constraints formed by the two helical directions on the soft robot 3 on the other hand.
[0058] More generally, the sensing fiber optic group 30 contains n right-handed optical fibers 33 and n left-handed optical fibers 34, where n ≥ 2. The n fibers of the same direction of rotation are arranged at equal angles along the circumference, and the fibers of the two directions of rotation are staggered by half an adjacent interval along the circumference. This symmetrical arrangement balances the circumferential and helical constraints imposed by the sensing fiber optic group 30 on the soft robot 3, and makes strain observations at different circumferential positions tend to be uniform. If unequal combinations of three left-handed optical fibers and one right-handed optical fiber, or three left-handed optical fibers and two right-handed optical fibers, or if the two fibers of the same direction of rotation are not symmetrically arranged, the enhancement effect of the sensing fiber optic group 30 will exhibit significant directionality, and the soft robot 3 is prone to undesirable torsion during bending. Therefore, it is preferable to have an equal number of fibers of the two directions of rotation and to arrange the fibers of the same direction of rotation symmetrically along the circumference.
[0059] When the sensing fiber optic group 30 uses two right-handed spiral fibers 33 and two left-handed spiral fibers 34, the two right-handed spiral fibers 33 are located in opposite circumferential positions, and the two left-handed spiral fibers 34 are also located in opposite circumferential positions, with the four fibers alternately arranged at approximately 90° intervals along the circumference. This scheme uses fewer sensing fibers and is suitable for instruments with smaller diameters or limited channel numbers. When using four right-handed spiral fibers 33 and four left-handed spiral fibers 34, the four fibers in the same spiral direction are arranged at approximately 90° intervals, and the two groups of fibers are staggered by approximately 45°, which can increase measurement redundancy. The scheme with three right-handed spiral fibers 33 and three left-handed spiral fibers 34 achieves a good balance between fiber quantity, circumferential resolution, and redundancy accuracy.
[0060] Each sensing fiber in the sensing fiber assembly 30 can be a single-core silica fiber, polymer fiber, or other optical fiber suitable for OFDR measurement. Each right-handed helical fiber 33 and left-handed helical fiber 34 is arranged along the outer wall of the soft robot 3 with the same or substantially the same helix angle. The helix angle can be determined based on the axial length and cross-sectional diameter of the robot body, and can also be affected by the number of braid turns or the braid pitch. In one specific embodiment, the six sensing fibers can use a braid pitch of approximately 20 mm.
[0061] The sensing fiber optic assembly 30 can be continuously bonded to the outer wall 3a of the soft robot 3 using, for example, medical adhesive (or other elastic adhesive), or it can be point-bonded at multiple axial locations and have relative slippage restricted by an external elastic coating layer. Preferably, the right-handed optical fiber 33 and the left-handed optical fiber 34 are made to have essentially no relative slippage with the outer wall 3a of the soft robot 3, so that deformation of the body surface can be reliably transmitted to the corresponding optical fiber. At the same time, the left-handed and right-handed optical fibers (n right-handed optical fibers 33 and n left-handed optical fibers 34) form a cross-reinforcement network on the outer wall 3a of the body, thereby improving the tensile and torsional resistance of the body without the need for additional independent reinforcement components. Therefore, in some embodiments, the sensing fiber optic assembly 30 can simultaneously perform the functions of distributed deformation sensing and body reinforcement.
[0062] The soft robot 3 has a central channel 31 extending axially. The central channel 31 can be an existing cavity within the soft robot body or a specially designed temperature measurement channel. A central temperature-measuring optical fiber 32 loosely fits through the central channel 31 and is not fixed to the inner wall of the central channel 31 with medical adhesive. Figure 2 As indicated by the bidirectional arrows at both ends of the central temperature-sensing fiber 32 in (a), the central temperature-sensing fiber 32 can slide axially relative to the soft robot 3. Therefore, when the soft robot 3 undergoes axial stretching or compression, the central temperature-sensing fiber 32 can reduce the mechanical strain transmitted to it by the axial extension and contraction of the body through relative sliding. In addition to accommodating the central temperature-sensing fiber 32, the central channel 31 can also retain a larger hollow margin, in which case the central channel 31 can be used as an instrument channel, cable channel, or fluid channel as needed.
[0063] The central temperature-sensing optical fiber 32 can be housed within a low-friction sleeve, or sliding friction can be reduced through structures such as a lubrication layer or a guide at the end of the central channel 31. One end of the central temperature-sensing optical fiber 32 can be fixed to the system base while the other end remains free, or both ends can be connected in a manner that allows axial compensation. As long as the central temperature-sensing optical fiber 32 does not experience the same mechanical strain as the peripheral sensing optical fiber due to the axial extension and contraction of the soft robot 3, its measurement results can be used to characterize temperature drift. This reduces the interference of the soft robot 3's stretching, compression, bending, and torsion on the temperature measurement signal.
[0064] The OFDR demodulation unit 4 (OFDR demodulator) inputs swept-frequency light into each sensing fiber in the sensing fiber group 30 and the central temperature-measuring fiber 32, and obtains the apparent strain or temperature characterization quantity distributed along the fiber length direction based on the phase or spectral changes of the Rayleigh scattered light at each position. Each sensing fiber can be connected to different channels of the OFDR demodulation unit 4 (OFDR demodulator) or connected sequentially through an optical switch. The strain data output by the OFDR demodulation unit 4 (OFDR demodulator) has a sequential form arranged along the fiber position. The processing unit converts the fiber coordinates into axial and circumferential coordinates in the unified coordinate system of the soft robot 3 based on the helical path parameters of the right-hand helical fiber 33 and the left-hand helical fiber 34.
[0065] Before use, each right-handed helical fiber 33, left-handed helical fiber 34, and central temperature-sensing fiber 32 in the sensing fiber assembly 30 can be calibrated under known temperature and deformation conditions to obtain the temperature sensitivity coefficient, mechanical strain transfer coefficient, helix angle deviation, and circumferential initial angle deviation. This calibration can correct for sensitivity differences between different fibers, differences in adhesive layer thickness, and deviations in actual placement relative to the ideal position.
[0066] like Figure 3 As shown in (a), under pure axial tension, both right-handed optical fiber 33 and left-handed optical fiber 34 elongate with the soft robot 3, resulting in an overall increase in strain across all channels; under pure axial compression, the overall strain across all channels decreases. The axial strain can be obtained by averaging the strain across each channel circumferentially, or by averaging the strains of right-handed optical fiber 33 and left-handed optical fiber 34 separately and then determining the common component. Therefore, axial expansion and contraction are mainly manifested as a common change in optical fibers with different helical directions.
[0067] like Figure 3 As shown in (b), under pure torsion conditions, one helical path (right-hand or left-hand) is lengthened, while the other helical path (left-hand or right-hand) is shortened. Consequently, the strains of the right-hand fiber 33 and the left-hand fiber 34 change in opposite directions. By subtracting the average strains of the two sets of fibers, the differential components corresponding to the torsion angle per unit length can be obtained. Since the number of right-hand fibers 33 and left-hand fibers 34 is equal and their helical directions are opposite, the common axial components can cancel each other out in the differential calculation, thereby reducing the influence of axial stretching on the torsion solution.
[0068] like Figure 3 As shown in (c), under pure bending conditions, the right-handed helical fiber 33 or the left-handed helical fiber 34 located on the outer side of the bend is under tension, while the corresponding fiber located on the inner side of the bend is under compression. Since the fibers are symmetrically distributed circumferentially, and the helical fibers periodically pass through different circumferential positions with varying axial positions, the strain of the right-handed helical fiber 33 and the left-handed helical fiber 34 exhibits an approximately periodic variation along the fiber position. The amplitude of this periodic variation is related to the magnitude of the curvature, and the phase is related to the bending direction, thus allowing the curvature components in both directions to be determined.
[0069] The equally symmetrically braided structure disclosed herein improves the symmetry of the sensing matrix. For the case of three co-rotating fibers arranged at 120° intervals, the sum of the circumferential sine terms of the same group of fibers is close to zero, thus the common axial component is less likely to be mixed with the bending component; while the right-handed fibers 33 and the left-handed fibers 34 are staggered by 60°, thus forming a more uniform circumferential sampling. Therefore, the structural symmetry not only reduces the additional torsion during bending but also reduces the computational coupling between different deformation parameters.
[0070] The processing unit sets multiple discrete sections along the axis of the soft robot body. At each discrete section, the curvature in the x-direction, curvature in the y-direction, axial elongation, and torsion angle per unit length are obtained based on the strain after temperature compensation. The spacing between the discrete sections can be equal to the OFDR spatial resolution, or it can be determined by window averaging of multiple original sampling points, in order to adjust between spatial resolution and noise suppression.
[0071] In one implementation, the processing unit uses a Bishop frame to integrate the aforementioned deformation parameters. By performing numerical integration along the axial direction starting from the base position and initial posture of the soft robot body, the initial reconstructed profile of the centerline can be obtained. Based on the configuration described above, the physical model can simultaneously incorporate the effects of bending, axial extension, and torsion on the centerline position and cross-sectional posture, avoiding systematic deviations caused by reconstructing solely based on bending curvature.
[0072] The physical model has a clear geometric meaning, but nonlinearities of the soft material, fiber bonding errors, cross-sectional eccentricity, driving hysteresis, and measurement deviations of various cross-sectional deformation parameters accumulate during integration. To reduce accumulated errors, this implementation further utilizes a data-driven model (see...). Figure 4 Predict the end position and use that end position to make a global correction to the initial reconstructed profile.
[0073] like Figure 4 As shown, the flexible continuum robot 41 is used to illustrate the object that generates sensing information in the deformed state of the soft robot 3. For the implementation with three right-handed optical fibers 33 and three left-handed optical fibers 34, the processing unit organizes the strain distribution of the six optical fibers, the curvature distribution in the x-direction, the curvature distribution in the y-direction, the axial stretching distribution, and the torsional distribution as input features 42, forming a total of ten channels of data. The strain of the six optical fibers can be the strain after temperature compensation. As another example, the strain before and after temperature compensation can also be input simultaneously, and the residual temperature effect can be learned by a deep learning model.
[0074] The processing unit can resample each channel in the input feature 42 along the axis of the soft robot 3 to the same number of sampling points, and perform normalization, zero-point correction, low-pass filtering, or outlier removal. Data from multiple consecutive time points constitute a timing window. The length of the timing window can be determined based on the motion speed of the soft robot 3, the sampling frequency of the OFDR demodulation unit 4 (OFDR demodulator), and the desired processing delay.
[0075] Figure 4 The data-driven model shown includes a convolutional layer 43, a pooling layer 44, a flattening layer 45, an LSTM layer 46, and a fully connected layer 47 connected in sequence. The convolutional layer 43 extracts local spatial variation features at the same time step along the axis of the soft robot 3; the pooling layer 44 performs dimensionality reduction and feature compression on the convolutional features; the flattening layer 45 converts multidimensional features into feature vectors suitable for temporal processing; the LSTM layer 46 captures deformation evolution and material hysteresis features based on the hidden states of adjacent time steps; and the fully connected layer 47 generates an end-effector position output 48 based on the temporal features output by the LSTM layer 46. The end-effector position output 48 includes the three-dimensional coordinates P = (P... x P y P z ).
[0076] Training data can be obtained by moving the soft robot body under different bending directions, bending radii, torsion angles, and axial extensions. Additionally, end-effector position labels can be provided using vision acquisition modules, optical positioning systems, magnetically compatible position markers, or high-precision mechanical positioning devices. The loss function for network training can be the mean square error of the end-effector's 3D coordinates, and can be supplemented with velocity continuity, shape smoothness, or physical consistency losses.
[0077] The end position output 48 can reflect material differences, manufacturing and assembly errors, and sensing errors in the actual system, but using only the end coordinates is insufficient to directly determine the complete shape along the axial direction. Therefore, this embodiment uses the initial reconstructed contour 491 obtained from the physical model 49 as a priori, and introduces the end position output 48 into the shape optimization process through constraints / feedback 50, so that the optimization result converges to the actual end position while maintaining the local distribution shape and the kinematic relationship of the continuum.
[0078] In one specific implementation, the physical model 49 discretizes the soft robot 3 into ten rigid links and ten ball joints, thus obtaining eleven nodes. The length of the i-th link is L. i The tilt angle of the i-th link relative to the reference axis is α. i The azimuth angle is β i The end position f(α, β, L), determined by the parameters of each link, can be calculated according to the following formula (1): (1) First, the initial reconstructed contour 491 calculated by the physical model 49 is discretized into eleven initial nodes with equal arc lengths. Then, based on the positional relationship between adjacent nodes, the initial tilt angle α_initial_i(t), initial azimuth angle β_initial_i(t), and initial link length L_initial_i(t) corresponding to each ball joint are calculated. The number of discrete joints is not limited to ten. Increasing the number of joints can improve the shape representation ability, but it will increase the number of optimization variables and computation time.
[0079] Let p est (t) represents the coordinates of the end point at time t corresponding to the end position output 48, p(t) is the set of eleven node coordinates calculated by the pseudo-rigid body model at time t, p(t-1) is the set of node coordinates optimized at the previous time, and λ, ω, γ and η are hyperparameters determined by experiments. Then, the optimization problem performed by constraint / feedback 50 can be expressed as Equation (2): (2) The first term of Equation (2) makes the end position of the pseudo-rigid body model close to the neural network prediction result, and the second term restricts the shape abrupt change between adjacent time steps. The range of values for tilt angle, azimuth angle, and link length keeps the optimization result close to the prior shape of the physical model. The hyperparameter λ is used to balance the end position error and time continuity, ω and γ are used to limit the search range of tilt angle and azimuth angle, and η is a length range coefficient less than 1.
[0080] It should be noted that the phrase "using the predicted three-dimensional coordinates of the end as a hard constraint" in the claims means that it can be achieved by directly setting the equality constraint f(α, β, L) = p_est(t) in the optimization problem, or it can be equivalently achieved by setting a sufficiently large weight in the first term of equation (2).
[0081] After optimization, the physical model 49 performs cubic spline interpolation based on the eleven optimization nodes to obtain the optimized shape 492 from the initial reconstructed contour 491, which is the global shape continuous along the three axes of the soft robot. Alternatively, B-splines, Bézier curves, piecewise circular arcs, or other interpolation methods that maintain node continuity can be used. Thus, the local distribution shape provided by the initial reconstructed contour 491 and the global correction information provided by the end-effector position output 48 can work together through constraints / feedback 50 to suppress the cumulative error generated by piecewise integration along the axis without discarding the kinematic priors of the continuum.
[0082] Below, on the combination Figure 5 The deformation reconstruction method of the multidimensional deformation sensing soft robot disclosed herein is described.
[0083] Figure 5 This paper illustrates a specific flowchart of a deformation reconstruction method for a multidimensional deformation-sensing soft robot. The flowchart includes a data acquisition step S101, a temperature compensation step S102, an initial reconstruction step, an end-effector prediction step, and a global reconstruction step. The initial reconstruction step can be further refined into a coordinate mapping step S103, a deformation parameter solution step S104, and a reconstruction step S105. The end-effector prediction step can be further refined into a feature construction step S106 and a prediction step S107. The global reconstruction step can be further refined into an optimization modeling step S108, an iterative correction step S109, and a reconstruction step S110. Furthermore, each step can be performed according to… Figure 5 The execution order is shown. Additionally, provided the data dependencies are satisfied, some data preprocessing or model computations can also be performed in parallel.
[0084] In acquisition step S101, OFDR demodulation unit 4 (OFDR demodulator) inputs swept light into the right-handed spiral fiber 33 and the left-handed spiral fiber 34 in the sensing fiber group 30, acquires Rayleigh scattering signals along the entire length of each sensing fiber and converts them into distributed apparent strain; at the same time, it acquires the temperature characterization of the central temperature-measuring fiber 32. This yields multi-channel distributed strain data from the peripheral sensing fibers and the temperature drift corresponding to changes in ambient temperature.
[0085] In the temperature compensation step S102, temperature compensation is applied to the apparent strain of each peripheral sensing fiber based on the temperature characterization of the central temperature-sensing fiber 32 and the temperature compensation coefficients of each right-handed and left-handed fiber 34. Since the central temperature-sensing fiber 32 can slide axially within the central channel 31, the mechanical strain transmitted by the axial extension and contraction of the soft robot 3 in its signal is relatively small. After subtracting this temperature component from the apparent strain of the peripheral sensing fibers, the temperature-compensated strain, which mainly characterizes the mechanical deformation of the soft robot 3, can be obtained.
[0086] In the coordinate mapping step S103, the processing unit maps the fiber coordinates output by the OFDR demodulation unit 4 (OFDR demodulator) to the unified axial and circumferential coordinates of the soft robot 3 based on the helical path parameters such as the helix angle, braiding pitch, and initial circumferential position of the right-hand helical fiber 33 and the left-hand helical fiber 34. Through this mapping, strain data collected on different helical paths can be mapped to the same axial discrete cross-section of the soft robot 3, providing spatially corresponding data for subsequent multi-channel joint calculation.
[0087] In step S104, the temperature-compensated strains of multiple right-handed and left-handed optical fibers 33 and 34 are input into the sensing matrix of the physical model 49 at discrete cross-sections along each axis to obtain the curvature in the x-direction, curvature in the y-direction, axial strain, and torsion angle per unit length. For the three right-handed optical fibers 33 and the three left-handed optical fibers 34, the common change in the six strain paths is mainly used to characterize axial expansion and contraction, the differential change in the two types of optical fibers is mainly used to characterize torsion, and the periodic change in each channel with circumferential position is used to characterize bending, thus enabling different deformation modes to be solved synchronously on the same cross-section.
[0088] In reconstruction step S105, the position and orientation of the soft robot's three bases are used as boundary conditions. The x-direction curvature, y-direction curvature, axial extension, and torsion angle obtained in step S104 are input into the Bishop frame kinematic model, and numerical integration is performed along the axial direction to obtain... Figure 4 The initial reconstructed profile 491 is shown. Since the axial stretching and torsion angle are integrated synchronously with the curvature in both directions, profile deviations caused by the incorrect attribution of composite deformation to the bending component can be reduced.
[0089] In feature construction step S106, the processing unit organizes the strain distribution, x-direction curvature distribution, y-direction curvature distribution, axial stretching distribution, and torsion distribution of the six optical fibers into... Figure 4 The input features shown are 42, forming a total of ten channels of time-series data. Each channel can undergo uniform sampling, normalization, filtering, or outlier processing before being input into the deep learning model, and the data from multiple consecutive time points are combined into a time-series window of a predetermined length.
[0090] In prediction step S107, the ten-channel temporal data generated in step S106 is input into a 1D-CNN-LSTM neural network consisting of convolutional layer 43, pooling layer 44, flattening layer 45, LSTM layer 46, and fully connected layer 47. Convolutional layer 43 extracts spatial features distributed along the axis of the soft robot 3 from single-time data, LSTM layer 46 extracts deformation evolution features based on data from multiple consecutive time points, and finally, fully connected layer 47 outputs the end-effector position output 48, which is the predicted result of the three-dimensional coordinates of the end of the soft robot 3.
[0091] In the optimization modeling step S108, the initial reconstructed contour 491 obtained in step S105 is discretized into a pseudo-rigid body model containing multiple ball joints. In a specific example, the soft robot 3 is discretized into ten rigid links and ten ball joints, resulting in eleven nodes. The initial values of the parameters of each ball joint and their allowable search range are determined based on the initial reconstructed contour 491. Simultaneously, the end-effector position output 48 obtained in step S107 is introduced into the physical model 49 via constraint / feedback 50, using the predicted three-dimensional coordinates of the end-effector as a global constraint for shape optimization.
[0092] In the iterative correction step S109, the predicted result of the three-dimensional coordinates of the end is used as a constraint, and the discrete nodes optimized at the previous moment can be used as time continuity constraints. The optimization problem is constructed according to Equation (2). By iteratively adjusting the tilt angle, azimuth angle and link length of the ball joint, the end position of the pseudo-rigid body model is made close to the end position output 48, and the optimized nodes are kept near the initial prior given by the physical model 49.
[0093] In reconstruction step S110, cubic spline interpolation is performed on the discrete nodes obtained from optimization in step S109 to obtain... Figure 4 The optimized shape 492 is shown, and the continuous global shape of the soft robot 3 is output. Alternatively, B-splines, Bézier curves, piecewise circular arcs, or other interpolation methods that maintain the continuity of discrete nodes can also be used. The processing unit can also simultaneously output the curvature, axial extension, torsion angle, and abnormal deformation regions exceeding a set threshold at each axial position for use in motion control, safety monitoring, or operation prompts.
[0094] In a set of verification embodiments, the soft robot 3 has an outer diameter of 6 mm and a total length of 77.5 mm. The sensing fiber optic group 30 includes six peripheral single-core optical fibers braided alternately with a braiding pitch of approximately 20 mm. A central temperature-sensing optical fiber 32 that can slide axially is installed in the central channel 31. The spatial resolution of the OFDR demodulation unit 4 (OFDR demodulator) is 0.64 mm. Different bending profiles are formed using circular and elliptical shaping molds, and multiple torsional conditions are superimposed to collect training and test data.
[0095] In this set of tests, the average errors in overall shape reconstruction for circular and elliptical contours were approximately 2.58% and 2.39%, respectively, and the average end-effector positioning errors were approximately 2.85 mm and 2.13 mm, respectively. These results are used to illustrate the feasibility of fusing the physical model 49 with the data-driven model and do not constitute limitations on the size, weaving pitch, network structure, or error range of the soft robot 3.
[0096] According to the method described in this disclosure, high-density distributed strain is first converted into physically meaningful local deformation parameters and an initial reconstructed profile 491 via steps S101 to S105. Then, via steps S106 and S107, the data-driven model generates an end-position output 48. Finally, via steps S108 to S110, global shape optimization is performed in the vicinity of the prior provided by the physical model 49. This maintains the kinematic consistency of the reconstructed result as a continuum and compensates for material property differences, manufacturing and assembly errors, and cumulative errors from piecewise integration using the global constraints of the end-position.
[0097] Other advantages and modifications will readily occur to those skilled in the art. Therefore, more broadly, the invention is not limited to the specific details and representative embodiments shown and described herein. Thus, modifications can be made without departing from the spirit or scope of the overall inventive concept as defined by the appended claims and their equivalents.
[0098] In one embodiment of this disclosure, an electronic device is provided, comprising a memory and a processor. The memory stores a computer program and may also store distributed strain data acquired by the OFDR demodulation unit 4 (OFDR demodulator), the temperature drift corresponding to the central temperature-sensing fiber 32, temperature-compensated strain data, deformation parameters output by the physical model 49, the end position output 48, and the global shape reconstruction result corresponding to the optimized shape 492. The processor is communicatively connected to the memory and performs operations when reading and executing the computer program. Figure 5 The deformation reconstruction method shown.
[0099] The processor can execute sequentially. Figure 5The data acquisition process includes data reception, temperature compensation, fiber optic coordinate mapping, sensor matrix solving, Bishop frame integration, multi-channel feature construction, deep learning model inference, pseudo-rigid body model iterative optimization, and spline interpolation. The processor can be a central processing unit, digital signal processor, graphics processor, embedded processor, or a combination thereof; the memory can include non-volatile memory and volatile memory used for program execution.
[0100] In another embodiment of this disclosure, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, is implemented. Figure 5 The deformation reconstruction method is illustrated. Furthermore, the computer-readable storage medium can be a hard disk, solid-state storage, read-only memory, random access memory, optical disk, or other non-transitory media capable of storing computer programs. Therefore, the deformation reconstruction process can be automatically executed by a host computer, workstation, embedded controller, or other electronic device.
[0101] (Other variations)
[0102] For longer soft robots, the robot body can be divided into multiple flexible segments, with an equal number of left-handed and right-handed optical fibers arranged symmetrically in each segment. The fibers of adjacent flexible segments can be independently connected to the OFDR demodulation unit 4, or they can be spliced or connected to form a continuous optical path. The processing unit reconstructs the shape of each flexible segment and combines them into a unified shape based on the pose continuity conditions of the ends of adjacent flexible segments.
[0103] Furthermore, for puncture needles, catheters, or other miniature instruments, the diameter of the soft robot body can be reduced, and a minimal symmetrical combination of two right-handed and two left-handed optical fibers can be used. For devices requiring higher measurement redundancy, a combination of four right-handed and four left-handed optical fibers, or more, can be used. Regardless of the number of fibers, fibers of the same direction of rotation are preferably arranged at equal angles around the circumference, and fibers of the two directions of rotation are preferably staggered and alternately arranged.
[0104] Deep learning models are not limited to 1D-CNN-LSTM; temporal convolutional networks, gated recurrent units, Transformers, or combinations thereof can also be used. End-effector position prediction can also be extended to the position prediction of one or more intermediate nodes, with multiple prediction nodes serving as shape optimization constraints. Physical models are not limited to the Bishop frame; Cosserat rod models, constant curvature models, piecewise constant curvature models, or finite element reduced-order models can also be used.
[0105] The processing unit can be implemented by a computer, workstation, embedded controller, graphics processing unit module, or cloud computing device. The communication connection between the OFDR demodulation unit 4 and the processing unit can be a wired network, serial bus, fiber optic communication, or wireless communication. Any alternative structure that can achieve distributed strain acquisition, multidimensional deformation parameter calculation, and global shape correction based on the predicted location can be used in this disclosure.
[0106] The above are merely preferred embodiments of this disclosure. Those skilled in the art can make adjustments to the number of optical fibers, helix angle, braiding pitch, soft robot materials, temperature compensation methods, deformation calculation models, deep learning models, and optimization algorithms without departing from the concept of this disclosure, and all such adjustments should fall within the protection scope of this disclosure.
Claims
1. A multidimensional deformation sensing soft robot system, characterized in that, include: The body of the soft robot; The fiber optic sensing unit includes a sensing fiber group and a temperature measuring component. The sensing fiber group includes multiple right-handed and multiple left-handed optical fibers symmetrically arranged and wrapped around the outer wall of the soft robot body in a circumferentially alternating spiral braid. The sensing fiber group is fixedly attached to the outer wall of the soft robot body. The temperature measuring component is axially slidably disposed in the central channel of the soft robot body. The OFDR demodulation unit is communicatively connected to the fiber optic sensing unit. The OFDR demodulation unit is configured to acquire distributed strain data of the sensing fiber group based on Rayleigh scattering, and to acquire the temperature drift of the temperature measuring component, and to perform temperature compensation on the distributed strain data based on the temperature drift. as well as The processing unit is communicatively connected to the OFDR demodulation unit. The processing unit incorporates a physical model and a deep learning model. It is configured to input temperature-compensated strain data into the physical model to solve for the curvature, axial expansion, and torsion angle of each cross-section of the soft robot body. Through integration, it obtains the initial reconstructed contour of the soft robot. The strain data from the sensing fiber optic group, along with the curvature, axial expansion, and torsion angle output by the physical model, are input as multi-channel time-series data into the deep learning model. The model outputs the predicted three-dimensional coordinates of the soft robot's end effector. Using the predicted three-dimensional coordinates of the end effector as constraints, the initial reconstructed contour is iteratively corrected. Finally, through interpolation, a global shape reconstruction result is obtained.
2. The multidimensional deformation sensing soft robot system as described in claim 1, characterized in that, The sensing fiber optic assembly consists of three right-handed and three left-handed fibers arranged symmetrically, alternately spirally woven around the outer wall of the soft robot body at a specified helical angle. The sensing fiber optic assembly is fixedly attached to the outer wall of the soft robot body, so that there is no relative slippage between the sensing fiber optic assembly and the outer wall of the soft robot body.
3. The multidimensional deformation sensing soft robot system as described in claim 2, characterized in that, Under pure tensile or pure compressive conditions, the strain of the three right-handed optical fibers changes synchronously with that of the three left-handed optical fibers. Under pure torsional conditions, the strain of the three right-handed optical fibers changes in opposite directions to that of the three left-handed optical fibers. Under pure bending conditions, the strain of the three right-handed and three left-handed optical fibers exhibits periodic sinusoidal fluctuations. The sensing fiber optic group decouples axial tensile / compression deformation, torsional deformation, and bending deformation based on the differentiated strain signal patterns under the three working conditions, and the sensing fiber optic group also serves as a reinforcing structure covering the outer wall of the soft robot body.
4. The multidimensional deformation sensing soft robot system as described in claim 1, characterized in that, The temperature measuring component is a central temperature measuring optical fiber. The central temperature-sensing optical fiber can be axially slidably disposed within the central channel of the soft robot body. The central temperature-measuring optical fiber is configured to collect the temperature drift of the environment in which the soft robot body is located.
5. The multidimensional deformation sensing soft robot system as described in any one of claims 1 to 4, characterized in that, The OFDR demodulation unit achieves distributed strain measurement at a specified spatial resolution based on Rayleigh scattering.
6. The multidimensional deformation sensing soft robot system as described in any one of claims 1 to 4, characterized in that, When the processing unit inputs the temperature-compensated strain data into the physical model, the physical model solves for the x-direction curvature, y-direction curvature, axial extension and twist angle of each section based on the cross-sectional geometry of the soft robot body. The model then integrates the x-direction curvature, y-direction curvature, axial extension and twist angle using the Bishop algorithm to obtain the initial reconstructed contour.
7. The multidimensional deformation sensing soft robot system as described in claim 6, characterized in that, The deep learning model is a 1D-CNN-LSTM neural network. The multi-channel time-series data consists of ten channels, including the original strain distribution of six optical fibers, as well as the curvature distribution, axial expansion / contraction distribution, and torsional distribution output by the physical model. The processing unit is configured to extract single-time axial spatial features through the one-dimensional convolutional layer of the 1D-CNN-LSTM neural network, and capture deformation temporal correlation features through the two-layer LSTM of the 1D-CNN-LSTM neural network, so as to output the three-dimensional coordinate prediction result of the soft robot end effector.
8. The multidimensional deformation sensing soft robot system as described in claim 7, characterized in that, When the processing unit iteratively corrects the initial reconstructed contour, it discretizes the soft robot body into a pseudo-rigid body model containing multiple ball joints, constructs an optimization objective function with the end-effector 3D coordinate prediction result as a hard constraint, uses the LBGS-F algorithm to iteratively correct the discrete points of the initial reconstructed contour, and obtains the global shape reconstruction result through cubic spline interpolation.
9. The multidimensional deformation sensing soft robot system as described in any one of claims 1 to 4, characterized in that, Both the fiber optic sensing unit and the soft robot body are made of non-metallic materials.
10. The multidimensional deformation sensing soft robot system as described in claim 2 or 3, characterized in that, The helix angle can be determined based on the axial length and cross-sectional diameter of the soft robot body.
11. A deformation reconstruction method for a multidimensional deformation-sensing soft robot, characterized in that, Includes the following steps: The acquisition steps involve acquiring distributed strain data of the sensing fiber group in the fiber optic sensing unit based on Rayleigh scattering using the OFDR demodulation unit, as well as acquiring the temperature drift of the temperature measuring component in the fiber optic sensing unit. The sensing fiber group is wrapped around the outer wall of the soft robot body by symmetrically arranged right-handed and left-handed optical fibers in a circumferentially alternating spiral braid. The temperature compensation step involves performing temperature compensation on the distributed strain data based on the temperature drift. In the initial reconstruction step, the temperature-compensated strain data is input into the built-in physical model through the processing unit to solve for the deformation parameters of each cross section of the soft robot body, and the initial reconstruction contour of the soft robot is obtained by integration. In the end-effector prediction step, the processing unit inputs the strain data of the sensing fiber group and the deformation parameters output by the physical model as multi-channel time-series data into the built-in deep learning model, and outputs the three-dimensional coordinate prediction result of the soft robot end effector. as well as The global reconstruction step involves using the processing unit to iteratively correct the initial reconstructed contour based on the predicted three-dimensional coordinates of the end point, thereby obtaining the global shape reconstruction result.
12. The deformation reconstruction method for a multidimensional deformation-sensing soft robot as described in claim 11, characterized in that, The sensing fiber optic assembly comprises six single-core optical fibers. These six single-core optical fibers are formed by symmetrically arranged three right-handed and three left-handed fibers, which are alternately spirally braided around the outer wall of the soft robot body at a specified helical angle. The fibers are completely fitted to the outer wall of the soft robot body without relative slippage. The distributed strain data is obtained by directly acquiring the Rayleigh scattering signals from the six single-core optical fibers through the OFDR demodulation unit.
13. The deformation reconstruction method for a multidimensional deformation-sensing soft robot as described in claim 12, characterized in that, In the initial reconstruction step, based on the structure of the three right-handed and three left-handed optical fibers arranged alternately in the circumference, the torsional deformation is obtained separately from the distributed strain data by utilizing the inverse strain variation law of the three right-handed and three left-handed optical fibers under pure torsional conditions.
14. The deformation reconstruction method for a multidimensional deformation-sensing soft robot as described in claim 11, characterized in that, In the acquisition step, the OFDR demodulation unit acquires the global discrete strain data of the sensing fiber group at a specified spatial resolution based on Rayleigh scattering.
15. The deformation reconstruction method for a multidimensional deformation-sensing soft robot as described in claim 11 or 14, characterized in that, In the initial reconstruction step, when the physical model solves for the deformation parameters of each section, it simultaneously incorporates the axial expansion and contraction parameters and the torsional angle parameters.
16. The deformation reconstruction method for a multidimensional deformation-sensing soft robot as described in claim 11, characterized in that, The initial reconstruction steps include: The temperature-compensated strain data is input into the physical model; Based on the cross-sectional geometry of the soft robot body, the curvature in the x-direction, curvature in the y-direction, axial extension and twist angle of each cross-section are obtained. The initial reconstructed contour is obtained by combining the Bishop algorithm integral.
17. The deformation reconstruction method for a multidimensional deformation-sensing soft robot as described in claim 11 or 16, characterized in that, The sensing fiber optic group includes six optical fibers. The terminal prediction step includes: The original strain distribution of the six optical fibers and the curvature distribution, axial stretching distribution and torsional distribution output by the physical model, totaling ten channels of time-series data, are input into a 1D-CNN-LSTM neural network; Single-time axial spatial features are extracted using the one-dimensional convolutional layer of the 1D-CNN-LSTM neural network; The 1D-CNN-LSTM neural network captures temporal correlation features of deformation; and Output the predicted three-dimensional coordinates of the soft robot's end effector.
18. The deformation reconstruction method for a multidimensional deformation-sensing soft robot as described in any one of claims 11, 16, and 17, characterized in that, The global reconstruction steps include: The soft robot body is discretized into a pseudo-rigid body model containing multiple ball joints; Using the predicted three-dimensional coordinates of the end point as a hard constraint, an optimization objective function is constructed, and the LBGS-F algorithm is used to iteratively correct the discrete points of the initial reconstructed contour; and The global shape reconstruction result is obtained by cubic spline interpolation.
19. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the deformation reconstruction method of the multidimensional deformation sensing soft robot as described in any one of claims 11 to 18.
20. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the deformation reconstruction method for a multidimensional deformation-sensing soft robot as described in any one of claims 11 to 18.
