A soft robot arm control method and control system based on optical fiber multi-mode sensing

Abstract

The application discloses a kind of soft robot arm control method and control system based on optical fiber multi-mode perception.It includes: obtaining the desired state of current time, input feedforward model to obtain feedforward power distribution;Collect multiple real-time pose signals, get current state through attitude perception neural network model, obtain outer loop state feedback error distribution according to current and desired state, obtain feedback temperature distribution by outer loop fuzzy PID controller;Collect the real-time temperature signal of each helical artificial muscle, obtain current temperature distribution, obtain inner loop temperature feedback error distribution according to current temperature distribution and feedback temperature distribution, obtain feedback power distribution by inner loop PID controller;Combine feedforward and feedback power distribution to obtain predicted power distribution, adjust the helical artificial muscle input voltage distribution of next time according to predicted power distribution;Repeat the above closed-loop control process until the work is finished.The application can ensure the reliability of soft robot arm in complex tasks and environment.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, specifically to a soft robotic arm control method and control system based on fiber optic multi-mode perception. Background Technology

[0002] Soft robots are mostly made of soft materials. Due to the inherent properties of soft materials, they can be used to achieve various biomimetic movements, such as grasping, crawling, jumping, and swimming. In terms of environmental adaptability and operational safety, they are superior to traditional rigid robots that require complex structures and control algorithms. To realize the posture control system of such robots, it is necessary to find a body perception method that can adapt to the body's compliance, achieve real-time perception of the body's state, monitor soft deformation, and incorporate algorithmic logic that enables precise control during operation.

[0003] In terms of ontological perception methods, many studies use external visual 3D tracking systems to collect geometric information of soft robots. This method is simple and easy to deploy, but it occupies a large space. Piezoelectric tensile sensor arrays can achieve posture perception with a small footprint, but their rigid structure makes them difficult to integrate well with soft structures, hindering the utilization of the material's flexibility. In recent years, rapidly developing fiber optic technology has provided another solution for transmitting strain, temperature, and force. An existing published patent (CN 115847430 A) uses directly measurable high-intensity fiber Bragg gratings (FBGS) as shape sensing sensors for soft arms, which can calculate position coordinates for key feature points on the robotic arm. However, strain compensation becomes very complex when encountering non-uniform strain fields.

[0004] Foreign scholars, such as Sareh S et al., developed a tensile sensor based on highly flexible optical fibers, which senses the bending angle of a soft robotic arm by detecting changes in the macro-bending radius of the fiber. Meanwhile, scholars like Zhao et al. fabricated optical fibers based on soft materials for tactile sensing. While these methods can be combined with the flexible characteristics of soft structures, their sensing data typically relies on direct experimental fitting to establish a mapping relationship. However, due to inherent errors in the light generation circuit, fiber structure, light receiving circuit, and subsequent processing circuits involved in the fabrication of fiber optic sensors, this method struggles to capture coupled nonlinear information in the system in real time, leading to a decline in the state perception performance of the soft robot.

[0005] Meanwhile, existing soft robot control methods often employ direct open-loop control or closed-loop control that does not consider the mechanical model of flexible materials. While these methods are simple, they cannot effectively capture the nonlinear characteristics of soft materials, especially since soft materials exhibit hysteresis in their response, which limits the dynamic performance of the control. To address the issues of dynamic response, accuracy, and stability in soft robot control, innovative control strategies are urgently needed to improve its performance and adaptability. Summary of the Invention

[0006] To address the problems existing in the background technology, this invention provides a soft robotic arm control method and control system based on fiber optic multi-mode sensing. The control method of this invention is based on position and force multi-mode sensing data processing methods using macro-bent optical fibers and soft optical fibers, constructing a soft robotic arm control method based on multi-mode sensing information. This control method combines outer-loop state feedback control, inner-loop temperature feedback control, and feedforward control, thereby achieving high-precision, fast, and stable control of the soft robotic arm.

[0007] The technical solution adopted in this invention is:

[0008] I. A control method for a soft robotic arm based on fiber optic multi-mode sensing

[0009] The soft robotic arm control method includes the following steps:

[0010] S1. For a soft robotic arm driven by multiple electrothermal-driven spiral artificial muscles, construct the basic coordinate system of the soft robotic arm.

[0011] Preferably, the state of the soft robotic arm is represented by the two-dimensional coordinates of the end feature point in the basic coordinate system or by a combination of the bending angle and swing angle of the central axis in the basic coordinate system.

[0012] S2. Obtain the desired state of the soft robotic arm at the current moment, process the desired state using the feedforward model of the soft robotic arm, and obtain the feedforward power distribution.

[0013] In step S2, the feedforward model of the soft robotic arm includes a kinematic model, a mechanical model, a muscle drive model, and a muscle thermoelectric model. The kinematic model is a model of the relationship between the muscle length distribution and the state of the soft robotic arm, established based on the piecewise constant curvature assumption. The mechanical model is a model of the relationship between the driving force distribution and the state of the soft robotic arm, established based on a third-order Yeoh constitutive model. The muscle drive model is a model of the relationship between the muscle length, driving force, and temperature of the helical artificial muscle, established based on a spring-damped temperature model. The muscle thermoelectric model is a model of the relationship between the temperature and input power of the helical artificial muscle, established based on a thermoelectric model.

[0014] Step S2 specifically involves:

[0015] S2.1 Obtain the desired state at the current moment based on the preset temporal desired state, periodic desired state, or real-time input desired state;

[0016] S2.2 Using the aforementioned robotic arm kinematic model, obtain the corresponding desired muscle length distribution based on the desired state;

[0017] S2.3 Using the mechanical model of the robotic arm, obtain the corresponding desired driving force distribution according to the desired state;

[0018] S2.4 Using the muscle drive model, based on the desired muscle length distribution obtained in step S2.2 and the desired driving force distribution obtained in step S2.3, obtain the desired temperature distribution;

[0019] S2.5. Using the muscle thermoelectric model, obtain the feedforward power distribution based on the desired temperature distribution obtained in step S2.4.

[0020] S3. Use a stretch sensor to collect multiple real-time pose signals. Based on each real-time pose signal, obtain the current state through a pose perception neural network model. Obtain the outer loop state feedback error distribution based on the current state and the desired state. Use an outer loop fuzzy PID controller to process the outer loop state feedback error and obtain the feedback temperature distribution.

[0021] Step S3 specifically involves:

[0022] S3.1. Use a stretch sensor to collect multiple real-time pose signals. The stretch sensor is mainly composed of several macro-bending optical fibers arranged at intervals along the circumference of the soft robotic arm. The output optical signal of each macro-bending optical fiber is a real-time pose signal.

[0023] S3.2. Through photoelectric conversion and analog-to-digital conversion, each real-time pose signal is converted into the corresponding current pose voltage. All current pose voltages are input into the pose perception neural network model, and the pose perception neural network model outputs the current state.

[0024] S3.3 In the basic coordinate system, the current state and the desired state are mapped to the direction vectors of each spiral artificial muscle to obtain the current muscle mapping state distribution and the desired muscle mapping state distribution. The difference between the current muscle mapping state distribution and the desired muscle mapping state distribution is used to obtain the muscle mapping state error distribution as the outer loop state feedback error distribution. The outer loop state feedback error distribution is composed of the outer loop state feedback error corresponding to each spiral artificial muscle.

[0025] S3.4. Use the outer loop fuzzy PID controller corresponding to each spiral artificial muscle to process the corresponding outer loop state feedback error and obtain the corresponding feedback temperature. The feedback temperatures corresponding to each spiral artificial muscle constitute the feedback temperature distribution.

[0026] S4. Use temperature sensors to collect real-time temperature signals of each spiral artificial muscle, obtain the current temperature distribution based on the real-time temperature signals of each spiral artificial muscle, and obtain the inner loop temperature feedback error distribution based on the current temperature distribution and the feedback temperature distribution; use an inner loop PID controller to process the inner loop temperature feedback error and obtain the feedback power distribution.

[0027] Specifically, using an inner-loop PID controller to handle the inner-loop temperature feedback error means using the inner-loop PID controller corresponding to each helical artificial muscle to handle the inner-loop temperature feedback error corresponding to that helical artificial muscle, obtaining the corresponding feedback power, and the feedback power corresponding to each helical artificial muscle constitutes the feedback power distribution.

[0028] S5. Combining the feedforward power distribution obtained in step S2 and the feedback power distribution obtained in step S4, a predicted power distribution is obtained. Based on the predicted power distribution, the input voltage of each spiral artificial muscle is adjusted at the next moment.

[0029] Step S5 specifically involves:

[0030] S5.1. Weighted summation of the feedforward power distribution and the feedback power distribution yields the predicted power distribution before saturation-limited amplitude signal processing.

[0031] S5.2. Compare the predicted input power of each helical artificial muscle before saturation limiting amplitude signal processing with the input power threshold in the predicted power distribution before saturation limiting amplitude signal processing. If the predicted input power of a certain helical artificial muscle before saturation limiting amplitude signal processing is greater than the input power threshold, then set the input power threshold to the predicted input power of this helical artificial muscle. Otherwise, keep the predicted input power unchanged to obtain the predicted power distribution.

[0032] S5.3. Convert the predicted power distribution into voltage control signals for each spiral artificial muscle at the next moment, and then adjust the input voltage of each spiral artificial muscle at the next moment, so as to control the state of the soft robotic arm at the next moment by adjusting the length of each spiral artificial muscle at the next moment.

[0033] S6. Repeat steps S2 to S5 until the soft robotic arm finishes its work.

[0034] The soft robotic arm control method further includes the following steps: using a tactile sensor to collect contact force signals in real time, wherein the tactile sensor adopts a soft optical fiber, and the output optical signal of the soft optical fiber is used as the contact force signal; through photoelectric conversion and analog-to-digital conversion processing, the real-time contact force signal is converted into the current contact voltage, the current contact voltage is input into the contact force sensing neural network model, and the contact force sensing neural network model outputs the current contact force.

[0035] For the posture-aware neural network model and the contact force-aware neural network model, each neural network model mainly consists of an input layer, three hidden layers, and an output layer connected in series. The activation function is the LeakyReLU function. During training, the Adam optimizer is used, and the mean squared error loss function is used for evaluation.

[0036] II. A soft robotic arm control system applied to the above-mentioned soft robotic arm control method

[0037] The soft robotic arm control system includes:

[0038] The soft robotic arm includes a robotic arm body, multiple helical artificial muscles for driving the soft robotic arm, multiple macro-bending optical fibers for use as tension sensors, multiple temperature sensors for collecting the temperature of the corresponding helical artificial muscles, and a soft optical fiber for use as a tactile sensor.

[0039] The feedforward control module is used to obtain the feedforward power distribution based on the desired state at the current moment;

[0040] The outer loop state feedback control module is used to obtain the feedback temperature distribution based on the desired state at the current moment and the real-time pose signal;

[0041] The inner loop temperature feedback control module is used to obtain the feedback power distribution based on the current feedback temperature distribution and the real-time temperature signal;

[0042] The saturation and signal processing module is used to obtain the input voltage of each spiral artificial muscle at the next moment based on the feedforward power distribution and feedback power distribution at the current moment.

[0043] Preferably, in the soft robotic arm, each spiral artificial muscle is arranged at intervals along the circumference of the robotic arm body, and each spiral artificial muscle is embedded inside the robotic arm body. Each macro-bent optical fiber is arranged at intervals along the circumference of the robotic arm body, and each macro-bent optical fiber is arranged on the outside of the robotic arm body. The number of spiral artificial muscles and macro-bent optical fibers are three, and they are arranged symmetrically in a one-to-one correspondence. Each spiral artificial muscle has a power monitoring chip integrated in its driving circuit as a temperature sensor. The soft optical fiber is coaxially arranged inside the soft robotic arm. The spiral artificial muscles are made by spiral twisting nylon polymer fibers and metal wires together.

[0044] The beneficial effects of this invention are:

[0045] This invention combines multi-mode sensing technology based on macro-bent optical fibers and soft optical fibers to achieve high-precision, rapid, and stable control of soft robotic arms in dynamic environments. The method learns and extracts signals from fiber optic sensors through a multi-layer neural network, establishing an accurate state-aware model that effectively solves the problem of traditional sensing methods failing to capture nonlinear coupling information. Simultaneously, a multi-level control strategy employing feedforward control, outer-loop state feedback control, and inner-loop temperature feedback control further enhances the system's adaptability and robustness, ensuring the reliability of the soft robotic arm in complex tasks and environments. By precisely adjusting the temperature and state of the actuators, the stability and safety of the soft robotic arm during long-term operation are guaranteed, significantly improving its ability to perform complex tasks. These innovative control strategies enable the soft robotic arm to efficiently perform various tasks and have broad application prospects. Attached Figure Description

[0046] Figure 1 A schematic diagram of the soft robotic arm control method based on fiber optic multi-mode sensing provided by the present invention;

[0047] Figure 2 This refers to the soft robotic arm used in the embodiments of the present invention;

[0048] Figure 3 This is a diagram showing the training process results of the posture-aware neural network model in an embodiment of the present invention;

[0049] Figure 4 This is a schematic diagram illustrating the effect of the soft robotic arm in tracking circles, spirals, hexagons, and pentagons in an embodiment of the present invention.

[0050] In the diagram: 100, main body of the robotic arm; 200, macro-bending optical fiber; 300, soft optical fiber. Detailed Implementation

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] The first aspect of this invention provides a control method for a soft robotic arm based on fiber optic multi-mode sensing. This invention provides a method for processing position and force multi-mode sensing data based on macro-bent optical fibers and soft optical fibers, and further constructs a control method for a soft robotic arm based on multi-mode sensing information. This control method combines outer-loop state feedback control, inner-loop temperature feedback control, and feedforward control, thereby achieving high-precision, fast, and stable control of the soft robotic arm.

[0053] Multi-mode sensing methods based on optical fibers include state (position information) sensing methods based on macro-bent optical fibers and contact force sensing methods based on soft optical fibers. Key features in the sensing data are learned and extracted using neural network methods, establishing a correlation between the signals obtained by the optical fiber sensors and the state of the mission space.

[0054] The closed-loop control method combines outer-loop state feedback control, inner-loop temperature feedback control, and feedforward control to achieve high-precision, fast, and stable control of the soft robotic arm. This invention enables efficient and stable state control of the soft robotic arm, adapting to the execution of various complex tasks and significantly improving the performance and reliability of soft robots in practical applications.

[0055] The method of this invention is applicable to soft robotic arms driven by multiple helical artificial muscles. In the following description, unless otherwise specified, muscle length distribution refers to the set of muscle lengths of all helical artificial muscles in the soft robotic arm; power distribution refers to the set of power of all helical artificial muscles in the soft robotic arm; temperature distribution refers to the set of temperature of all helical artificial muscles in the soft robotic arm; driving force distribution refers to the set of driving forces of all helical artificial muscles in the soft robotic arm; and muscle mapping state distribution refers to the set of muscle mapping states of all helical artificial muscles in the soft robotic arm.

[0056] The following is combined Figure 1 The control method of the present invention is described in detail, and the control method of the present invention specifically includes the following steps:

[0057] S1. In the computer, for a soft robotic arm driven by multiple electrothermal-driven spiral artificial muscles, construct the basic coordinate system of the soft robotic arm, obtain the direction vector of each spiral artificial muscle, obtain the two-dimensional coordinates of the end of the soft robotic arm, and set the initial state of the soft robotic arm.

[0058] Optionally, the state of the soft robotic arm (including the initial state, desired state, and current state, etc.) is represented by the two-dimensional coordinates of its end feature point in the basic coordinate system or by a combination of the bending angle and swing angle of its central axis in the basic coordinate system.

[0059] The direction vector of each helical artificial muscle is typically along the axial direction, pointing from the tip of the soft robotic arm to the end.

[0060] Step S1 is as follows: For the soft robotic arm, a basic coordinate system is established with the geometric center of the head end as the origin and the head end face as the XY plane. Based on the position of each spiral artificial muscle in the soft robotic arm, the direction vector of each spiral artificial muscle is obtained. With the geometric center of the end of the soft robotic arm as the relative origin, the two-dimensional coordinates of the end of the soft robotic arm relative to the basic coordinate system are obtained. The two-dimensional coordinates consist of coordinates in the x and y directions.

[0061] S2. Feedforward Control: Obtain the desired state of the soft robotic arm at the current moment, process the desired state using the soft robotic arm's feedforward model, and obtain the feedforward power distribution. The feedforward control process significantly improves the dynamic performance of the control system and reduces the rise time and overshoot of the control system.

[0062] The feedforward model of a soft robotic arm includes a kinematic model of the robotic arm, a mechanical model of the robotic arm, a muscle-driven model, and a muscle thermoelectric model.

[0063] Step S2 is as follows:

[0064] S2.1 Obtain the desired state at the current moment based on the preset temporal desired state, periodic desired state, or real-time input desired state;

[0065] Optionally, a temporal or periodic desired state can be generated based on a preset motion curve;

[0066] Furthermore, for the preset temporal expected state, the next target coordinate can be iterated after the current coordinate falls within the two-dimensional spatial absolute state error (APE) threshold.

[0067] S2.2 Using the kinematic model of the robotic arm, obtain the corresponding expected muscle length distribution based on the desired state;

[0068] S2.3. Using the mechanical model of the robotic arm, obtain the corresponding desired driving force distribution based on the desired state;

[0069] S2.4. Using the muscle-driven model, the desired temperature distribution is obtained based on the desired muscle length distribution obtained in step S2.2 and the desired driving force distribution obtained in step S2.3. The input of the muscle-driven model is the muscle length and driving force of each helical artificial muscle, and the output is the temperature of the helical artificial muscle.

[0070] S2.5. Using the muscle thermoelectric model, obtain the feedforward power distribution based on the desired temperature distribution obtained in step S2.4; wherein, the input of the muscle thermoelectric model is the temperature of each helical artificial muscle, and the output is the input power of the helical artificial muscle.

[0071] Preferably, the kinematic model of the robotic arm is a model of the relationship between the muscle length distribution of the soft robotic arm and the state of the soft robotic arm, which is established based on the assumption of piecewise constant curvature.

[0072] Preferably, the mechanical model of the robotic arm is a model relating the driving force distribution of the soft robotic arm to the state of the soft robotic arm, and is established based on the third-order Yeoh constitutive model.

[0073] Preferably, the muscle driving model is a model of the relationship between muscle length, driving force and temperature of a spiral artificial muscle, which is established based on a spring-damped temperature model.

[0074] Preferably, the muscle thermoelectric model is a model of the relationship between the temperature and input power of the spiral artificial muscle, and is established based on the thermoelectric model.

[0075] S3. Outer Loop State Feedback Control: Multiple real-time pose signals are collected using a stretching sensor. Based on each real-time pose signal, the current state is obtained through an attitude perception neural network model. The outer loop state feedback error distribution is obtained based on the current state and the desired state. The outer loop fuzzy PID controller is used to process the outer loop state feedback error and obtain the feedback temperature distribution.

[0076] Among them, the outer-loop fuzzy PID controller refers to a PID controller with a fuzzy controller connected in parallel in the outer-loop state feedback control process. By using the fuzzy controller to dynamically adjust the gain of the PID controller, the adaptability of the system is improved, while ensuring that the outer closed-loop controller can accurately track the controlled state target.

[0077] Each spiral artificial muscle actuator corresponds to an outer-loop fuzzy PID controller.

[0078] Step S3 is as follows:

[0079] S3.1. Multiple real-time pose signals are acquired using a tension sensor. The tension sensor mainly consists of several macro-bending optical fibers 200 evenly spaced along the circumference of the robotic arm body 100. Each macro-bending optical fiber 200 is arranged along the axial direction of the robotic arm body 100 and is connected to the robotic arm body 100. The number of macro-bending optical fibers 200 is the same as that of the spiral artificial muscles and they are arranged symmetrically in one-to-one correspondence. The output light signal of each macro-bending optical fiber 200 serves as a real-time pose signal. By monitoring the optical power loss of the pose signal in real time, the pose state of the soft robotic arm can be monitored in real time. The pose state includes displacement and bending state.

[0080] S3.2. Through photoelectric conversion and analog-to-digital conversion, each real-time pose signal is converted into the corresponding current pose voltage. The current pose voltage corresponding to all real-time pose signals is input into the attitude perception neural network model, and the attitude perception neural network model outputs the current state.

[0081] S3.3 In the basic coordinate system, the current state and the desired state are mapped to the direction vectors of each spiral artificial muscle to obtain the current muscle mapping state distribution and the desired muscle mapping state distribution. The difference between the current muscle mapping state distribution and the desired muscle mapping state distribution is used as the muscle mapping state error distribution as the outer loop state feedback error distribution.

[0082] S3.4. The corresponding outer-loop fuzzy PID controller for each helical artificial muscle processes the corresponding outer-loop state feedback error to obtain the corresponding feedback temperature. The feedback temperatures for each helical artificial muscle constitute the feedback temperature distribution. This process can dynamically adjust the controller parameters based on the outer-loop state feedback error and the error change rate to ensure that the soft robotic arm maintains efficient and precise motion control in complex tasks.

[0083] S4. Inner Loop Temperature Feedback Control: A temperature sensor is used to collect the real-time temperature signal of each spiral artificial muscle. The current temperature distribution is obtained based on the real-time temperature signal of each spiral artificial muscle. The inner loop temperature feedback error distribution is obtained based on the current temperature distribution and the feedback temperature distribution. The inner loop PID controller is used to process the inner loop temperature feedback error and obtain the feedback power distribution.

[0084] Specifically, using an inner-loop PID controller to handle the inner-loop temperature feedback error means using the inner-loop PID controller corresponding to each helical artificial muscle to handle the inner-loop temperature feedback error corresponding to that helical artificial muscle, obtaining the corresponding feedback power, and the feedback power corresponding to each helical artificial muscle constitutes the feedback power distribution.

[0085] In this embodiment of the invention, the temperature sensor employs a power monitoring chip. Each helical artificial muscle's driving circuit integrates a power monitoring chip, and the output voltage and current signals of each power monitoring chip serve as the temperature signals for the corresponding helical artificial muscle. Based on the output voltage and current signals of the power monitoring chip corresponding to each helical artificial muscle, the resistance of the helical artificial muscle can be calculated, and its temperature can be calculated based on the resistance.

[0086] S5. Combining the feedforward power distribution obtained in step S2 and the feedback power distribution obtained in step S4, a predicted power distribution is obtained. Based on the predicted power distribution, the input voltage of each spiral artificial muscle is adjusted at the next moment.

[0087] Step S5 is as follows:

[0088] S5.1. Weighted summation of the feedforward power distribution and the feedback power distribution yields the predicted power distribution before saturation-limited amplitude signal processing.

[0089] S5.2. Compare the predicted input power of each helical artificial muscle before saturation limiting amplitude signal processing with the input power threshold in the predicted power distribution before saturation limiting amplitude signal processing. If the predicted input power of a certain helical artificial muscle before saturation limiting amplitude signal processing is greater than the input power threshold, then the input power threshold is set to the predicted input power of this helical artificial muscle. Otherwise, the predicted input power is kept unchanged to obtain the predicted power distribution. The purpose of saturation limiting amplitude signal processing is to ensure that the helical artificial muscle is stable within the safe operating temperature range.

[0090] S5.3. Convert the predicted power distribution into voltage control signals for each spiral artificial muscle at the next moment, and then adjust the input voltage of each spiral artificial muscle at the next moment, so as to control the state of the soft robotic arm at the next moment by adjusting the length of each spiral artificial muscle at the next moment.

[0091] Preferably, in step S5.1, the feedforward power distribution and the feedback power distribution are weighted and summed according to preset weights; in the first three motion cycles, the weighted values ​​of the feedforward power distribution and the feedback power distribution are 0.8 and 0.2, respectively, and in the subsequent motion cycles, the weighted values ​​of the feedforward power distribution and the feedback power distribution are 0.2 and 0.8, respectively.

[0092] Optionally, in step S5.2, a microcontroller is used to convert the predicted power distribution into voltage control signals for each spiral artificial muscle at the next moment using pulse width modulation technology.

[0093] S6. Repeat steps S2 to S5 until the soft robotic arm finishes its work. Specifically, within the time step after the current moment ends, steps S2 to S5 are executed sequentially to dynamically adjust the input voltage of each helical artificial muscle at the next moment based on the desired state at the current moment, the real-time pose signal, and the real-time temperature signal. In this way, the soft robotic arm can move according to a preset temporal desired state or a real-time input desired state.

[0094] Furthermore, the control method of the present invention also includes the following processes:

[0095] The tactile sensor collects the contact force signal in real time. The tactile sensor uses a soft optical fiber, which is coaxially inserted into the center of the robotic arm body 100. The output light signal of the soft optical fiber serves as the contact force signal. By monitoring the real-time light intensity of the contact force signal, the contact force at the end of the soft robotic arm can be monitored in real time. Through photoelectric conversion and analog-to-digital conversion, the real-time contact force signal is converted into the current contact voltage. The current contact voltage is input into the contact force sensing neural network model, and the contact force sensing neural network model outputs the current contact force.

[0096] Preferably, for the attitude-aware neural network model and the contact force-aware neural network model: each neural network model mainly consists of an input layer, three hidden layers and an output layer connected in series. The activation function is the Leaky ReLU function. During training, the Adam optimizer is used and the mean squared error loss function is used for evaluation.

[0097] A second aspect of the present invention provides a soft robotic arm control system based on fiber optic multi-mode sensing.

[0098] The soft robotic arm control system of this invention includes:

[0099] The soft robotic arm includes multiple helical artificial muscles for driving the soft robotic arm, multiple macro-bending optical fibers for use as tension sensors, multiple temperature sensors for collecting the temperature of the corresponding helical artificial muscles, and a soft optical fiber.

[0100] The feedforward control module is used to obtain the feedforward power distribution based on the desired state at the current moment;

[0101] The outer loop state feedback control module is used to obtain the feedback temperature distribution based on the desired state at the current moment and the real-time pose signal;

[0102] The inner loop temperature feedback control module is used to obtain the feedback power distribution based on the current feedback temperature distribution and the real-time temperature signal;

[0103] The saturation and signal processing module is used to obtain the input voltage of each spiral artificial muscle at the next moment based on the feedforward power distribution and feedback power distribution at the current moment.

[0104] Specifically, in the soft robotic arm, each spiral artificial muscle is evenly arranged at circumferential intervals along the main body 100 of the robotic arm, and each spiral artificial muscle is embedded inside the main body 100 of the robotic arm. Each macro-bent optical fiber 200 is evenly arranged at circumferential intervals along the main body 100 of the robotic arm, and each macro-bent optical fiber 200 is arranged on the outside of the main body 100 of the robotic arm. The number of spiral artificial muscles and macro-bent optical fibers 200 are three, and they are arranged symmetrically in a one-to-one correspondence. That is, each spiral artificial muscle and its corresponding macro-bent optical fiber 200 are arranged opposite each other on both sides of the main body 100 of the robotic arm. Each spiral artificial muscle's drive circuit integrates a power monitoring chip as a temperature sensor. The soft optical fiber 300 is coaxially arranged inside the soft robotic arm.

[0105] Specifically, the helical artificial muscle is made by twisting nylon polymer fibers and metal wires together in a spiral. This drive unit, under the influence of an electrical signal, can cause temperature changes, resulting in a change in its length and providing flexible and efficient driving force.

[0106] Preferably, the metal wire is a nickel wire.

[0107] Specific embodiments of the present invention are as follows:

[0108] Example

[0109] This embodiment is based on, as follows Figure 1 The control system shown is implemented.

[0110] Figure 1 The meanings of each parameter are as follows:

[0111] (xd ,y d () represents the desired state at the current moment;

[0112] (x c ,y c This represents the current state.

[0113] n Id This represents the desired muscle mapping state decomposed to the direction of the I-th helical artificial muscle. The desired muscle mapping states of the three helical artificial muscles constitute the desired muscle mapping state distribution.

[0114] n Ic This represents the real-time muscle mapping state decomposed to the direction of the I-th helical artificial muscle. The current muscle mapping states of the three helical artificial muscles constitute the current muscle mapping state distribution.

[0115] E nI The outer loop state feedback error of the outer loop fuzzy PID controller corresponding to the I-th spiral artificial muscle is represented by the current outer loop state feedback error distribution of the three spiral artificial muscles.

[0116] dE nI / dt represents the rate of change of the outer loop state feedback error of the outer loop controller corresponding to the I-th spiral artificial muscle;

[0117] K pI K iI K dI These represent the proportional gain, integral gain, and derivative gain of the outer loop fuzzy PID controller channel corresponding to the I-th helical artificial muscle, respectively.

[0118] T Id This represents the feedback temperature signal output by the outer loop fuzzy PID controller of the I-th helical artificial muscle. The feedback temperature signals of the three helical artificial muscles constitute the feedback temperature distribution.

[0119] T Ic This represents the current temperature signal of the I-th spiral artificial muscle obtained by the temperature calculation module. The current temperature signals of the three spiral artificial muscles constitute the current temperature distribution.

[0120] P Ifb This represents the feedback power signal of the I-th spiral artificial muscle output by the inner loop PID controller. The feedback power signals of the three spiral artificial muscles constitute the feedback power distribution.

[0121] P Iff This represents the feedforward power signal of the I-th spiral artificial muscle output by the feedforward controller. The feedforward power signals of the three spiral artificial muscles constitute the feedforward power distribution.

[0122] P I The predicted power signal before saturation limiting amplitude signal processing for the I-th helical artificial muscle is represented by the predicted power signal before saturation limiting amplitude signal processing for the three helical artificial muscles.

[0123] u I This represents the input voltage applied to the I-th spiral artificial muscle after saturation treatment. The input voltages of the three spiral artificial muscles constitute the input voltage distribution.

[0124] R I This represents the resistance of the I-th spiral artificial muscle;

[0125] U SI This represents the current pose voltage output by the I-th macro-bending tension sensor;

[0126] l I This represents the theoretical length corresponding to the final voltage control signal of the I-th helical artificial muscle;

[0127] In the control system of this embodiment, the structure of the soft robotic arm is as follows: Figure 2As shown. The soft robotic arm used in this embodiment includes three helical artificial muscles for driving the soft robotic arm, three macro-bending optical fibers for serving as tension sensors, three temperature sensors for collecting the temperature of the corresponding helical artificial muscles, and one soft optical fiber. Specifically: the soft robotic arm includes a robotic arm body 100, a tension sensor 200, and a tactile sensor 300. The robotic arm body 100 is made of silicone material and has three helical artificial muscles embedded inside. The three helical artificial muscles are evenly distributed circumferentially and are powered by Joule heating to achieve high-precision two-degree-of-freedom bending motion. The helical artificial muscles are made of nylon polymer fibers and nickel wires through a common helical twisting method. The tension sensor 200 includes three macro-bending optical fibers arranged on the outside of the robotic arm body 100. The macro-bending optical fibers correspond one-to-one with the helical artificial muscles, with each macro-bending optical fiber arranged on the opposite side of the corresponding helical artificial muscle. The tactile sensor 300 includes one soft optical fiber arranged in the middle of the robotic arm body 100. The temperature sensor employs a power monitoring chip; each helical artificial muscle's drive circuit integrates a power monitoring chip, and the output voltage and current signals of each chip serve as the temperature signal for the corresponding helical artificial muscle. Utilizing proprioception technology, the soft robotic arm can not only monitor its own posture changes in real time but also instantly report modulus differences and pressure changes in external contact objects. This enables the soft robotic arm to perform tasks more accurately and stably, improving operational reliability and flexibility. Furthermore, by reusing position closed-loop control code and adjusting relevant models, closed-loop control of contact forces can be achieved. The flexibility of this control strategy allows the soft robotic arm to adapt to various complex environments and provide efficient and precise control performance according to different task requirements.

[0128] like Figure 1 As shown, in the control system of this embodiment, the state of the soft robotic arm is represented by the two-dimensional coordinates of its end feature point on the xy plane in the basic coordinate system.

[0129] like Figure 1 As shown, in this embodiment of the control system, the feedforward control module includes a feedforward controller, and the input of the feedforward controller is the desired state (x). d ,y d The output is a feedforward power distribution (P) 1ff ,P 2ff ,P 3ff ).

[0130] like Figure 1As shown, in this embodiment of the control system, the outer-loop state feedback control module includes an attitude-aware neural network model, two direction adjusters, and three outer-loop fuzzy PID controllers. The direction adjusters are equipped with the basic coordinate system of the soft robotic arm and are used to map the end-effector state of the soft robotic arm to the direction vector of each helical artificial muscle, thus obtaining the muscle mapping state of the helical artificial muscle. The outer-loop fuzzy PID controllers refer to PID controllers connected in parallel with fuzzy controllers; each outer-loop fuzzy PID controller corresponds to one helical artificial muscle. The input to the attitude-aware neural network model is the current pose voltage distribution (U... S1 U S2 U S2 The output is the current state (x) c ,y c The input to the direction adjuster_1 is the desired state (x). d ,y d The output is the desired muscle mapping state distribution (n). 1d ,n 2d ,n 3d The input to the direction adjuster_2 is the current state (x). c ,y c The output is the current muscle mapping state distribution (n 1c ,n 2c ,n 3c The input to the outer-loop fuzzy PID controller_I is the outer-loop state feedback error E of the helical artificial muscle_I. nI The output is the feedback temperature T. Id .

[0131] like Figure 1 As shown, in this embodiment of the control system, the inner-loop temperature feedback control module includes three inner-loop PID controllers, each corresponding to a helical artificial muscle. The input of the inner-loop PID controller_I is the inner-loop temperature feedback error of the helical artificial muscle_I, and the output is the feedback power P. Ifb .

[0132] like Figure 1 As shown, in the control system of this embodiment, three weighted summation calculation units are set between the inner loop temperature feedback control module and the saturation and signal processing module, each corresponding to one of the three helical artificial muscles. The input of the weighted summation calculation unit_I is the feedforward power P of the helical artificial muscle_I. Iff and feedback power P Ifb The output is the predicted power distribution P of the saturation-limited amplitude signal of the helical artificial muscle_I before signal processing. I .

[0133] like Figure 1As shown, in the control system of this embodiment, the input of the saturation and signal processing module is the predicted power distribution (P1, P2, P3) before saturation limiting amplitude signal processing, and the output is the input voltage signal (u1, u2, u3) of each helical artificial muscle at the next moment. By adjusting the input voltage of each helical artificial muscle at the next moment, the length of each helical artificial muscle at the next moment can be adjusted, thereby controlling the state of the soft robotic arm at the next moment.

[0134] In this embodiment, the control method specifically proceeds as follows:

[0135] S1. For a soft robotic arm driven by multiple helical artificial muscles, construct a basic coordinate system for the soft robotic arm and set its initial state. In this embodiment, the state of the soft robotic arm is represented by the two-dimensional coordinates (x, y) of its end-effector feature points in the basic coordinate system.

[0136] S2. Obtain the desired state at the current moment, process the desired state using the feedforward model of the soft robotic arm, and obtain the feedforward power distribution. The feedforward model of the soft robotic arm includes the kinematic model of the robotic arm, the mechanical model of the robotic arm, the muscle drive model, and the muscle thermoelectric model.

[0137] In this embodiment, the feedforward model is constructed through the following process:

[0138] First, based on the piecewise constant curvature assumption, the geometric relationship between the end-effector posture of the soft arm in the task space and the length of the artificial muscle actuator in the actuation space is derived, and a kinematic model of the robotic arm is established based on this relationship.

[0139] Next, a static model of the soft robotic arm structure is performed to obtain the mechanical model of the robotic arm. During the modeling process, it is assumed that the bending deformation of the soft arm follows the Euler-Bernoulli beam assumption, neglecting the influence of shear deformation on the overall stress state. This assumption simplifies the calculation process and helps to accurately predict the bending behavior of the soft robotic arm. Since the soft robotic arm substrate is made of a hyperelastic material (such as silicone), its mechanical properties can be characterized using a third-order Yeoh constitutive model. During the modeling process, by simultaneously determining the stress state on any cross-section of the soft robotic arm during deformation, and combining this with the force distribution under the target bending state, the stress and force state on the muscle distribution circle corresponding to each target bending angle are determined. Furthermore, this force state is decomposed into muscle reaction force distributions in three directions to obtain the required driving force in each muscle direction. The mechanical model of the robotic arm provides an accurate reference for feedforward control, enabling effective adjustment of the output of each actuator, thereby achieving precise control of the soft robotic arm.

[0140] Subsequently, a quasi-static muscle drive model was established based on the classic spring-damped temperature model. The muscle drive model is expressed by the following formula:

[0141]

[0142] In the formula, F represents the driving force of the helical artificial muscle, k represents the stiffness coefficient of the helical artificial muscle, b represents the damping coefficient, c represents the temperature coefficient, l represents the desired muscle length of the helical artificial muscle, and l0 represents the initial muscle length of the helical artificial muscle. The value represents the rate of change of length of the helical artificial muscle, T represents the desired temperature of the helical artificial muscle, and T0 represents the initial temperature of the helical artificial muscle. The stiffness coefficient k, damping coefficient b, and temperature coefficient c of the helical artificial muscle can be obtained through isothermal stretching, isotonic heating, and free oscillation experiments.

[0143] Finally, a muscle thermoelectric model was established using the classical thermoelectric model.

[0144] S3. Use a stretch sensor to collect multiple real-time pose signals. Based on each real-time pose signal, obtain the current state through a pose perception neural network model. Obtain the outer loop state feedback error distribution based on the current state and the desired state. Use an outer loop fuzzy PID controller to process the outer loop state feedback error and obtain the feedback temperature distribution.

[0145] S4. Use a temperature sensor to collect the real-time temperature signal of each spiral artificial muscle, obtain the current temperature distribution based on the real-time temperature signal of each spiral artificial muscle, and obtain the inner loop temperature feedback error distribution based on the current temperature distribution and the feedback temperature distribution; use an inner loop PID controller to process the inner loop temperature feedback error and obtain the feedback power distribution.

[0146] S5. Combining the feedforward power distribution obtained in step S2 and the feedback power distribution obtained in step S4, a predicted power distribution is obtained. The input voltage of each spiral artificial muscle at the next moment is adjusted according to the predicted power distribution.

[0147] S6. Repeat steps S2 to S5 until the soft arm reaches the desired state. During this process, the system makes adaptive adjustments based on feedback data to ensure efficient and stable task execution.

[0148] Figure 3 The training process of the neural network model in this embodiment is demonstrated, specifically showing that the mean squared error decreases rapidly with each training iteration. The following details the training process of the position kinematics model. For a 2-DOF soft robotic arm, kinematics can be represented as the actuator drive space (i.e., the actuator length, l = (l1, l2, l3), where l1, l2, and l3 are the lengths of the three helical artificial muscles), to the task space state (s = (s0, s1)). T(i.e., the mapping of the two degrees of freedom s0 and s1 of the soft arm, which can be the two positions of the end effector or the bending angle and the swing angle) The task space position information used in this invention is the electrical signal (U) output by the tension sensor based on macro-bending optical fiber installed on the opposite side of the drive unit (helical artificial muscle). s =(U s0 U s1 U s2 ) T (U) S0 U S1 and U S2 These are obtained from the voltage output of each of the tension sensor components.

[0149] Because the macro-bending state of the optical fiber in three directions during the deformation of the soft robotic arm involves multiple highly coupled motion patterns, simple stretching calibration cannot accurately obtain information about the task space. However, neural networks, through multi-layered structures and nonlinear activation functions, can learn and extract important features, effectively capturing complex nonlinear relationships. This allows for a good connection between the driving space signal obtained from the macro-bending optical fiber stretching sensor and the task space state. Therefore, neural networks can be used to obtain kinematic values, which can be represented by the kinematic expression of the soft robotic arm as follows:

[0150]

[0151] In the formula, For the estimation of the current spatial state, U s This is the electrical signal output by a tension sensor based on macro-bending optical fiber.

[0152] In this embodiment, firstly, under different input voltage distributions of the helical artificial muscle, the corresponding soft arm end state and voltage information converted from the output optical signals of each macro-bent fiber were collected experimentally. All voltage information under the same helical artificial muscle input voltage distribution was used as input, and the soft arm end state as output, constructing a sample pair. All sample pairs were combined into a dataset. The dataset was used to train the posture-aware neural network model. During sampling, the input voltage range of each helical artificial muscle was divided into three segments according to the maximum safe voltage value. Every 10 seconds, zero to two drivers were randomly selected in each segment and assigned random PWM values. Finally, the coverage of the dataset was checked, and the PWM value was manually assigned to control the soft arm end to move in the direction where the data (soft arm end state) was sparse. Data monitoring and acquisition were achieved by real-time camera capture of the soft arm end's identification code. Fiber optic sensor data and soft arm tip position information were monitored and recorded every 5 seconds. A total of 7512 sample datasets were collected for offline learning to obtain higher neural network fitting accuracy. During training, the sample dataset was divided into a training set (80%) and a validation set (20%).

[0153] The training and computation environment for the neural network is Keras, a high-level API based on PyTorch developed in Python 3.10. The neural network consists of five layers, including three hidden layers with varying numbers of neurons. The number of neurons in the hidden layers is determined based on the learning outcome. The input layer has three neurons (U... s1 U s2 U s3 The output voltage value of the macro-bending fiber optic stretch sensor is represented by , and the output layer neurons correspond to the end-effector kinematic state of the soft arm. The activation function used is the Leaky ReLU function. This allows for a small gradient even when the input is negative, effectively mitigating the vanishing gradient problem inherent in ReLU in the negative region, which can cause some neurons to become inactive during training. Furthermore, compared to other activation functions (such as sigmoid or tanh), Leaky ReLU is simpler and more computationally efficient.

[0154] Model fitting performance was evaluated using the mean squared error (MSE) loss function, and the model parameters were optimized using the Adam optimizer with a learning rate of 0.001. To prevent overfitting, an early stopping method was employed: training was stopped when the MSE on the validation set was below 1% for 50 consecutive cycles. The weights and biases with the lowest MSE values ​​were selected as the optimal training parameters. In the regression analysis evaluation, the coefficient of determination (R-squared) score is more intuitive than metrics such as mean absolute error (MAE) and mean squared error (MSE). Therefore, the R-squared index was used to evaluate the kinematic accuracy of the neural network model. The R-squared values ​​for the training and validation sets were 0.9743 and 0.9788, respectively, indicating a close relationship between the input and output variables.

[0155] The above methods for learning and extracting location information can also be used to learn and extract the relationship between the electrical signals output by the soft fiber optic contact sensors installed inside the soft arm and the contact information of the soft arm, and to construct a contact force sensing neural network model.

[0156] Figure 4This paper demonstrates experiments applying the proposed control method to a soft arm. This embodiment includes experiments on circular trajectory dynamic tracking, hexagonal trajectory tracking, star trajectory tracking, spiral trajectory tracking, repetitive point localization, and tactile perception. The specific process is as follows: the desired soft arm tip position is written to a file and transmitted to a computer. The computer calculates the required muscle voltage value using a closed-loop control algorithm and transmits it to a lower-level computer (microcontroller) via serial communication. The lower-level computer adjusts the duty cycle of the output PWM wave by adjusting the CCR (Capture / Compare Register) register value and the ARR (Auto-reload Register) register value to make the output PWM frequency 1kHz. In the control timing, each drive cycle (280ms) uses the first 260ms to generate a square wave voltage drive with a controllable duty cycle, and the remaining time generates a square wave low voltage with a fixed duty cycle of 10% for resistance temperature measurement. A macro-bending fiber optic tension sensor system measures the actual arm tip state at a frequency of 50Hz and provides status feedback to the computer at a frequency of 20Hz after filtering.

[0157] Because the dynamic response of thermally driven systems is relatively slow, the control update frequency is set to 10Hz to ensure that the controller can effectively track system dynamics and avoid excessive computational load. In the trajectory tracking experiment, to shorten the task time, the next target coordinate is iterated only when the current coordinate falls within the absolute state error (APE) threshold in two-dimensional space. The absolute state error threshold is set to 1.5mm.

[0158] The dynamic circular trajectory tracking experiment aims to evaluate the comprehensive performance of the soft arm in terms of precision control, compliance, and robustness. The experimental task is to track a circular trajectory with a radius of 40 mm counterclockwise. We divided the trajectory into 360 target coordinate points. The experimental results show that the soft arm, which relies on a fiber optic macro-bending and stretching sensor and a neural network to achieve closed-loop position control, achieves a motion trajectory that is close to the ideal circular trajectory throughout the process. The mean absolute position error (MAE) of this method is 1.06 ± 0.46 mm, and the maximum positioning error (MaxE) is 3.03 mm.

[0159] The spiral trajectory tracking experiment aims to evaluate the motion capability of a soft arm for complex trajectories with gradually increasing amplitude and radius. The experimental task is to track an Archimedean spiral with an initial radius of 0 and a spiral spacing of 3, whose polar coordinate equation is:

[0160]

[0161] The trajectory starts from the center and rotates three times around the origin, with the polar angle changing from 0 to 6π radians. The radius of the spiral increases by 3 mm with each rotation, forming a uniformly spaced spiral shape. We divide the trajectory into 360 target coordinate points. Calculations show that when tracking the spiral trajectory, the mean absolute position error (MAE) of this method is 1.17 ± 0.57 mm, and the maximum positioning error (MaxE) is 3.37 mm.

[0162] The hexagonal trajectory tracking experiment aimed to evaluate the tracking capability of a soft arm for complex geometric trajectories. The experimental task was to track a hexagonal trajectory counterclockwise, with an circumscribed circle radius of 40 mm. The hexagonal trajectory was divided into 200 target coordinate points. Calculations showed that the mean absolute position error (MAE) of this method was 1.12 ± 0.46 mm, and the maximum positioning error (MaxE) was 2.86 mm. Throughout the experiment, the soft arm was able to approach the target hexagonal trajectory well, demonstrating strong trajectory tracking capabilities.

[0163] The star-shaped trajectory tracking experiment aims to evaluate the tracking accuracy of the soft arm on complex polygonal paths with sharp corners, and further examine the adaptability of the control algorithm, the compliance of the system, and its dynamic response performance. The controller tracks a pentagonal star-shaped trajectory with an outer circle radius of 30 mm with constant target accuracy. We divide this trajectory into 200 target coordinate points. When tracking this trajectory, the mean absolute position error (MAE) of this method is 1.28 ± 0.66 mm, and the maximum positioning error (MaxE) is 3.08 mm.

[0164] The experimental results above show that this control method performs excellently in achieving precise position tracking and can effectively track trajectories of different shapes and complexities.

[0165] In summary, the method of this invention ensures the stability and safety of the soft robotic arm during long-term operation by precisely adjusting the temperature and state of the actuator, significantly improving its ability to perform complex tasks. These innovative control strategies enable the soft robotic arm to efficiently perform various tasks and have broad application prospects.

[0166] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other structures that conform to changes in materials, thickness, etc., and any other changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A control method for a soft robotic arm based on fiber optic multi-mode sensing, characterized in that, Includes the following steps: S1. For a soft robotic arm driven by multiple electrothermal-driven spiral artificial muscles, construct the basic coordinate system of the soft robotic arm; S2. Obtain the desired state of the soft robotic arm at the current moment, process the desired state using the feedforward model of the soft robotic arm, and obtain the feedforward power distribution. The state of the soft robotic arm is represented by the two-dimensional coordinates of the end feature point in the basic coordinate system or by a combination of the bending angle and the swing angle of the central axis in the basic coordinate system; S3. Use a stretch sensor to collect multiple real-time pose signals. Based on each real-time pose signal, obtain the current state through a pose perception neural network model. Obtain the outer loop state feedback error distribution based on the current state and the desired state. Use an outer loop fuzzy PID controller to process the outer loop state feedback error and obtain the feedback temperature distribution. S4. Use a temperature sensor to collect the real-time temperature signal of each spiral artificial muscle, obtain the current temperature distribution based on the real-time temperature signal of each spiral artificial muscle, and obtain the inner loop temperature feedback error distribution based on the current temperature distribution and the feedback temperature distribution; use an inner loop PID controller to process the inner loop temperature feedback error and obtain the feedback power distribution. S5. Combining the feedforward power distribution obtained in step S2 and the feedback power distribution obtained in step S4, a predicted power distribution is obtained. The input voltage of each spiral artificial muscle at the next moment is adjusted according to the predicted power distribution. Step S5 specifically involves: S5.

1. Weighted summation of the feedforward power distribution and the feedback power distribution yields the predicted power distribution before saturation-limited amplitude signal processing. S5.

2. Compare the predicted input power of each helical artificial muscle before saturation limiting amplitude signal processing with the input power threshold in the predicted power distribution before saturation limiting amplitude signal processing. If the predicted input power of a certain helical artificial muscle before saturation limiting amplitude signal processing is greater than the input power threshold, then set the input power threshold to the predicted input power of this helical artificial muscle. Otherwise, keep the predicted input power unchanged to obtain the predicted power distribution. S5.

3. Convert the predicted power distribution into voltage control signals for each spiral artificial muscle at the next moment, and then adjust the input voltage of each spiral artificial muscle at the next moment to control the state of the soft robotic arm at the next moment by adjusting the length of each spiral artificial muscle at the next moment. S6. Repeat steps S2 to S5 until the soft robotic arm finishes its work.

2. The soft robotic arm control method based on fiber optic multi-mode sensing according to claim 1, characterized in that: In step S2, the feedforward model of the soft robotic arm includes a kinematic model of the robotic arm, a mechanical model of the robotic arm, a muscle drive model, and a muscle thermoelectric model. The kinematic model of the robotic arm is a model relating the muscle length distribution of the soft robotic arm to the state of the soft robotic arm, and is established based on the assumption of piecewise constant curvature. The mechanical model of the robotic arm is a model relating the driving force distribution of the soft robotic arm to the state of the soft robotic arm, and is established based on the third-order Yeoh constitutive model. The muscle driving model is a model of the relationship between muscle length, driving force and temperature of a spiral artificial muscle, which is based on a spring-damped temperature model. The muscle thermoelectric model is a model of the relationship between temperature and input power of spiral artificial muscle, and is established based on the thermoelectric model.

3. The soft robotic arm control method based on fiber optic multi-mode sensing according to claim 2, characterized in that: Step S2 specifically involves: S2.1 Obtain the desired state at the current moment based on the preset temporal desired state, periodic desired state, or real-time input desired state; S2.2 Using the aforementioned robotic arm kinematic model, obtain the corresponding desired muscle length distribution based on the desired state; S2.3 Using the mechanical model of the robotic arm, obtain the corresponding desired driving force distribution according to the desired state; S2.4 Using the muscle drive model, based on the desired muscle length distribution obtained in step S2.2 and the desired driving force distribution obtained in step S2.3, obtain the desired temperature distribution; S2.

5. Using the muscle thermoelectric model, obtain the feedforward power distribution based on the desired temperature distribution obtained in step S2.

4.

4. The soft robotic arm control method based on fiber optic multi-mode sensing according to claim 2, characterized in that: Step S3 specifically involves: S3.

1. Use a stretch sensor to collect multiple real-time pose signals. The stretch sensor consists of several macro-bending optical fibers (200) arranged at intervals along the circumference of the soft robotic arm. The output signal of each macro-bending optical fiber (200) is a real-time pose signal. S3.

2. Through photoelectric conversion and analog-to-digital conversion, each real-time pose signal is converted into the corresponding current pose voltage. All current pose voltages are input into the pose perception neural network model, and the pose perception neural network model outputs the current state. S3.3 In the basic coordinate system, the current state and the desired state are mapped to the direction vectors of each spiral artificial muscle to obtain the current muscle mapping state distribution and the desired muscle mapping state distribution. The difference between the current muscle mapping state distribution and the desired muscle mapping state distribution is obtained to obtain the outer loop state feedback error distribution. S3.

4. Use the outer loop fuzzy PID controller corresponding to each spiral artificial muscle to process the corresponding outer loop state feedback error and obtain the corresponding feedback temperature. The feedback temperatures corresponding to all spiral artificial muscles constitute the feedback temperature distribution.

5. The soft robotic arm control method based on fiber optic multi-mode sensing according to claim 1, characterized in that, It also includes the following steps: A tactile sensor is used to collect contact force signals in real time. The tactile sensor uses a soft optical fiber, and the output signal of the soft optical fiber is used as the contact force signal. The real-time contact force signal is converted into the current contact voltage through photoelectric conversion and analog-to-digital conversion. The current contact voltage is input into the contact force sensing neural network model, and the contact force sensing neural network model outputs the current contact force.

6. The soft robotic arm control method based on fiber optic multi-mode sensing according to claim 1, 4, or 5, characterized in that: The neural network models are mainly composed of an input layer, three hidden layers, and an output layer connected in series. The activation function is the LeakyReLU function. During training, the Adam optimizer is used, and the mean squared error loss function is used for evaluation.

7. A soft robotic arm control system applied to the soft robotic arm control method as described in any one of claims 1 to 6, characterized in that, include: The soft robotic arm includes a robotic arm body (100), multiple helical artificial muscles for driving the soft robotic arm, multiple macro-bending optical fibers (200) for use as tension sensors, multiple temperature sensors for collecting the temperature of the corresponding helical artificial muscles, and a soft optical fiber (300) for use as a tactile sensor. The feedforward control module is used to obtain the feedforward power distribution based on the desired state at the current moment; The outer loop state feedback control module is used to obtain the feedback temperature distribution based on the desired state at the current moment and the real-time pose signal; The inner loop temperature feedback control module is used to obtain the feedback power distribution based on the current feedback temperature distribution and the real-time temperature signal; The saturation and signal processing module is used to obtain the input voltage of each spiral artificial muscle at the next moment based on the feedforward power distribution and feedback power distribution at the current moment.

8. The soft robotic arm control system according to claim 7, characterized in that: Each spiral artificial muscle is arranged at intervals along the circumference of the main body (100) of the robotic arm. Each spiral artificial muscle is embedded inside the main body (100). Each macro-bent optical fiber (200) is arranged at intervals along the circumference of the main body (100). There are three spiral artificial muscles and three macro-bent optical fibers (200), and they are arranged symmetrically in a one-to-one correspondence. Each spiral artificial muscle has a power monitoring chip integrated in its drive circuit as a temperature sensor. The soft optical fiber (300) is coaxially arranged inside the soft robotic arm. The spiral artificial muscles are made by spiral twisting nylon polymer fibers and metal wires together.

Images