Bionic control method for soft actuator of artificial muscle based on pulse neural network

Abstract

The application discloses a kind of " artificial muscle " soft driver bionic control method based on pulse neural network, the method is with rough adjustment based on feedforward controller of rough inverse model, with fine adjustment based on cerebellum controller of pulse neural network, real-time compensation modeling uncertainty and environmental interaction uncertainty is realized through online learning method.The application uses representative " artificial muscle " soft driver-pneumatic artificial muscle, realizes a kind of pulse neural network that imitates mammal cerebellum neural working mechanism, realizes the efficient, robust, adaptive control for highly nonlinear, time-varying " artificial muscle " soft driver.

Description

Technical Field

[0001] This invention relates to the field of control science and engineering, and in particular to a biomimetic control method for a soft actuator of an "artificial muscle" based on a spiking neural network. Background Technology

[0002] Soft actuators, often referred to as "artificial muscles," possess advantages such as high energy density, fast response speed, and strong environmental compliance, effectively meeting the requirements of compliance, safety, and comfort in human-computer interaction, thus attracting widespread attention in the field of robotics research. However, the inherent characteristics of soft actuators also bring many challenges to control applications in robotics. On the one hand, the unique material properties of soft materials and the actuation method of soft actuators cause the system to exhibit strong time-varying and nonlinear characteristics, especially under high-speed, large-deformation motion, where viscoelasticity, elastic hysteresis, creep, and mechanical friction are more pronounced, making the mathematical description of the controlled object and the design of the corresponding controller extremely difficult. On the other hand, the characteristics of soft actuators make classical control methods somewhat lacking in balancing control accuracy and system compliance. Although researchers have proposed methods such as force / position hybrid control and impedance control to handle compliant force control problems, these are mostly geared towards traditional "hard-actuated" robots. Artificial muscle soft actuators share high similarities with biological muscles in terms of actuation mechanisms and material properties. However, living organisms can effectively coordinate and control their skeletal muscles through the nervous system to achieve complex and dexterous movements that are currently difficult for artificial control systems to achieve, while also exhibiting excellent environmental adaptability. Therefore, some researchers have been inspired by this to explore control methods for artificial muscle soft actuators from a life science perspective. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing soft actuator control methods for "artificial muscles" by providing a biomimetic control method for soft actuators based on spiking neural networks.

[0004] The objective of this invention is achieved through the following technical solution: a biomimetic control method for a soft actuator of an "artificial muscle" based on a spiking neural network, comprising:

[0005] Based on the aerodynamic characteristic equations of a single PAM "artificial muscle" soft actuator and the "artificial muscle" soft actuator system model, the system model of the soft actuator robot is determined.

[0006] Derive the feedforward controller from the system model of the soft actuator robot;

[0007] Based on mammalian cerebellar neural circuits, a cerebellar-like controller based on spiking neural networks was designed.

[0008] Control is achieved by combining the control commands output by the feedforward controller and the cerebellum-like controller.

[0009] Furthermore, the construction of the aerodynamic characteristic equations for a single PAM "artificial muscle" soft actuator includes:

[0010] When gas is introduced into the PAM inner tube, the input energy obtained by the drive system is:

[0011] dW in =∫ I (P-P0)dl·ds=(P-P0)∫ I dl·ds=P′dV (1)

[0012] Among them, dW in Let P and P0 be the input energy micro-element, respectively, the internal tube pressure and the ambient atmospheric pressure (1 atm ≈ 1.01 × 10⁻⁶). 5 Pa), P′ is the relative air pressure inside the tube, I is the total space inside the tube, dl is the infinitesimal displacement element of the inner tube, ds is the infinitesimal cross-sectional area element of the inner tube, and dV is the infinitesimal volume element of the inner tube; according to the definition, the output energy of the driver is:

[0013] dW out =-Fdl (2)

[0014] Among them, dW out Let F be the output energy infinitesimal element, and F be the output axial driving force. Ignoring frictional energy loss and energy storage in the PAM system, according to the law of conservation of energy, the input energy equals the output energy, i.e.:

[0015] dW in =dW out (3)

[0016] Substituting equations (1) and (2) into the equations, we get:

[0017] P′dV=-Fdl

[0018]

[0019] To calculate the differential of PAM volume with respect to PAM length, dV / dl, we consider PAM as a cylinder with length l and base diameter d. Since the PAM surface is made of a nylon mesh, the elasticity of each braid is very small; therefore, we can assume the braid is of fixed length b. α represents the angle between the braid and the PAM axis, and n is the number of turns of the braid. Based on geometric relationships, the geometric parameters of PAM can be obtained as follows:

[0020] l=bcosα (5)

[0021]

[0022] Therefore, the volume of PAM is:

[0023]

[0024] Therefore, equation (4) can be further written in the following form:

[0025]

[0026] When PAM is in its initial state, its output driving force F is 0, so let equation (8) be 0, and the initial α, i.e. α0, satisfies:

[0027]

[0028] Therefore, equation (8) can be further written as:

[0029]

[0030] Substituting equation (5) into equation (10) yields the final aerodynamic characteristic equation:

[0031]

[0032] Where l0 is the initial length of PAM.

[0033] Furthermore, the construction of the system model for the soft-actuated robot includes:

[0034] The "artificial muscle" soft actuator system model includes the flexor muscle PAM1 and the extensor muscle PAM2;

[0035] For the length l1 of the flexor muscle PAM1, we can obtain the following from the law of cosines:

[0036]

[0037] The length l2 of the extensor PAM2 will be increased by the length increment Δl2 caused by elbow flexion, on top of the original length, that is:

[0038]

[0039] Equations (15) and (16) are the kinematic equations of the robot system. D0, D1, D2, D3, and D4 are the structural parameters of the actuator system, and θ is the angular position. Based on these equations, the lengths of the two PAM blocks at the desired angular position can be calculated.

[0040] Next, the dynamic equations of the robot system will be derived: Since the robot mechanism is driven by the PAM antagonistic pair, the rotational motion of the moving segment is determined by the combined effect of the two PAM driving forces; according to the rigid body rotation law, the dynamic equation of the overall antagonistic mechanism is:

[0041]

[0042] Where J is the moment of inertia of the moving segment of the robot skeleton, τ is the equivalent external torque generated by the two PAMs, and τ g For gravitational torque, These are internal and external uncertainties arising from both external environmental disturbances and model simplification assumptions.

[0043] Treating the moving segments of the robot skeleton as uniform thin rods, the moment of inertia of the moving segments can be estimated as follows:

[0044]

[0045] Where m is the mass of the active segment, L r Let θ be the length of the moving segment. When the angular position of the moving segment is θ, the gravitational torque τ can be directly obtained according to the definition of torque. g ,Right now:

[0046]

[0047] If we denote the lever arms corresponding to the driving forces f1 and f2 generated by PAM1 and PAM2 as r1 and r2 respectively, then the resultant driving torque generated by the PAM antagonistic pair is:

[0048] τ=f1r1-f2r2 (20)

[0049] According to equation (11):

[0050]

[0051] Where K is a proportionality constant, P1 is the air pressure inside the flexor PAM1 tube, P2 is the air pressure inside the extensor PAM2 tube, and l 10 =D1+D2 is the initial length of the flexor muscle PAM1, l 20 =D3+D4 is the initial length of the extensor muscle PAM2; based on the geometric relationship of the robot mechanism, we obtain:

[0052]

[0053] Substituting equation (21) into equation (20), and treating the lever arms of the two driving forces as constants R, we obtain the driving torque expressed in matrix form, namely:

[0054]

[0055] in

[0056]

[0057] P = [P1 P2] T (25)

[0058] The joint stiffness s is defined as the differential of the driving torque at the diagonal position, i.e.:

[0059]

[0060] Substituting equation (23) into equation (26), according to the matrix differentiation rule, we have:

[0061] s=-Γ θ (θ)P (27)

[0062] in

[0063]

[0064] Combining equations (23) and (27), we have the system model for the soft-actuated robot:

[0065]

[0066] Furthermore, the feedforward controller is a system model for a soft-actuated robot, which determines the current desired angular position θ. d With the desired joint stiffness s d Calculate the control commands required to achieve the target motion, i.e., the air pressure P in the PAM inner tube model.

[0067] Furthermore, the cerebellar controller comprises an MF layer, an IO layer, a GC layer, a PC layer, and a DCN layer; firstly, the MF layer processes the input sensing signals, i.e., the actual angular position θ and the desired angular position θ. d The process involves encoding neural pulses and mapping the pulse sequence to discrete excitation states of neurons in the GC layer. The neural pulses emitted by the GC layer neurons are then transmitted to the PC layer via the PF layer. The synaptic weights of the PF-PC layer are updated online using the learning signal from the IO layer, i.e., the control error e. After the PC layer integrates the motion sensing states from the GC layer, its emitted neural pulses will inhibit neurons in the DCN layer. Simultaneously, the DCN layer is activated by the MF layer. Finally, under the combined effects of inhibition and activation, the DCN layer generates an excitation state of neurons, thereby emitting output neural pulses encoded with the current compensation signal. Decoding these pulses yields the output compensation command of the cerebellar controller, i.e., the compensation pressure P. c .

[0068] Furthermore, the cerebellar-like controller optimizes its synaptic weights through online learning based on the current control error e, thereby fitting and modeling uncertainties and external environmental disturbances to generate the optimal compensating air pressure P under the current state. c .

[0069] Furthermore, optimizing synaptic weights through online learning includes:

[0070] The MSTDPET learning algorithm applies the following rules to network weight updates:

[0071]

[0072] LTD ij (t)=η - ∫δ IO,j (t)E ij (t)dt (32)

[0073] Among them, w ij Here, δ represents the network weights, t represents time, δ represents the network impulse, and the subscripts GR and IO indicate the corresponding neuron layers. ij η represents the qualification trace of synapses ij. + and η - These are the learning rates for LTP and LTD, respectively; during the initial training phase, the learning rate η is set to its maximum value η. max Furthermore, the learning rate decays exponentially as learning progresses, eventually reaching a stable learning rate η. min ,Right now:

[0074]

[0075] Where τ0 is the time constant for learning rate decay. The traces of presynaptic and postsynaptic neural activity are updated according to the following rules:

[0076]

[0077] Where, τ + and τ - Let be the time constant for trace decay, and t0 be the initial time. The eligibility trace can be updated by integrating over the previous states, i.e.:

[0078]

[0079] Where τ1 and τ2 are the decay constants of the qualification trace.

[0080] Furthermore, the control command output by the feedforward controller is the PAM inner tube pressure, and the control command output by the cerebellar controller is the compensation pressure. Combining the control commands output by the feedforward controller and the cerebellar controller results in the sum of the PAM inner tube pressure and the compensation pressure. The sum of the PAM inner tube pressure and the compensation pressure is the desired PAM pressure P. d .

[0081] Furthermore, a PID method is used for the underlying control of each PAM, which calculates the control voltage u of the solenoid valve opening based on the air pressure deviation, and then adjusts the actual air pressure P of the corresponding PAM to the desired air pressure P. d This outputs a driving force F, which in turn acts on the robot skeleton to generate an actual angular motion θ.

[0082] This invention also provides a biomimetic control device for an "artificial muscle" soft actuator based on a spiking neural network, comprising:

[0083] The model building module is used to determine the system model of the soft actuator robot based on the aerodynamic characteristic equations of a single PAM "artificial muscle" soft actuator and the "artificial muscle" soft actuator system model.

[0084] The first control command module is used to derive a feedforward controller based on the system model of the soft-actuated robot; the input of the feedforward controller is the desired angular position θ. d With the desired joint stiffness s d The output is the PAM inner tube pressure P;

[0085] The second control command module is used to design a cerebellar-like controller based on a spiking neural network, based on the cerebellar neural circuit of mammals; the inputs of the cerebellar-like controller are the actual angular position θ and the desired angular position θ. d The output is the compensation air pressure P. c ;

[0086] The control module is used to combine the control commands output by the feedforward controller and the cerebellum-like controller for control.

[0087] The beneficial effects of this invention are that it uses a cerebellar-like controller based on a spiking neural network to achieve efficient, robust, and adaptive control of the soft actuator of "artificial muscle". It solves the problems of strong nonlinearity and time-varying nature faced by existing methods, and improves the environmental adaptability of the control system while taking into account the control accuracy. Attached Figure Description

[0088] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0089] Figure 1 This is a geometric characteristic diagram of the soft actuator of the "artificial muscle" under deformation.

[0090] Figure 2 This is a diagram of an antagonistic "artificial muscle" soft actuator system.

[0091] Figure 3 This is the overall biomimetic control architecture diagram.

[0092] Figure 4 This is a simplified diagram of a cerebellar controller based on a spiking neural network. Detailed Implementation

[0093] The present invention will now be described in detail with reference to the accompanying drawings. Unless otherwise specified, the features of the following embodiments and implementations can be combined with each other.

[0094] The present invention provides a biomimetic control method for a soft actuator of an "artificial muscle" based on a spiking neural network, comprising the following steps:

[0095] Step 1: System modeling of the soft-actuated robot.

[0096] like Figure 1 As shown, when a certain amount of gas is filled into the inner tube of the PAM (pneumatic artificial muscle), the input energy obtained by the actuator system is:

[0097] dW in =∫ I (P-P0)dl·ds=(P-P0)∫ I dl·ds=P′dV (1)

[0098] Among them, dW in Let P and P0 be the input energy micro-element, respectively, the internal tube pressure and the ambient atmospheric pressure (1 atm ≈ 1.01 × 10⁻⁶). 5 Pa), P′ is the relative air pressure inside the tube, I is the total space inside the tube, dl is the infinitesimal displacement element of the inside tube, ds is the infinitesimal cross-sectional area element of the inside tube, and dV is the infinitesimal volume element of the inside tube. According to the definition, the driver output energy is:

[0099] dW out =-Fdl (2)

[0100] Among them, dW out Let F be the output energy element, and F be the output axial driving force. In static or quasi-static motion states, frictional energy loss and energy storage in the PAM system can be neglected. Therefore, according to the law of conservation of energy, the input energy equals the output energy, i.e.:

[0101] dW in =dW out (3)

[0102] Substituting equations (1) and (2) into the equations, we get:

[0103] P′dV=-Fdl

[0104]

[0105] To calculate the differential dV / dl of the PAM volume with respect to the PAM length, the PAM can be approximated as follows: Figure 1 The cylinder shown has a length of l and a base diameter of d. Since the PAM surface is made of a nylon mesh, the elasticity of each braid is very small; therefore, it can be assumed that the braid is of a fixed length b. α represents the angle between the braid and the PAM axis, and n is the number of turns of the braid. Based on geometric relationships, the geometric parameters of the PAM can be obtained as follows:

[0106] l=bcosα (5)

[0107]

[0108] Therefore, the volume of PAM is:

[0109]

[0110] Therefore, equation (4) can be further written in the following form:

[0111]

[0112] When PAM is in its initial state, its output driving force F is 0, so let equation (8) be 0, and the initial α, i.e. α0, satisfies:

[0113]

[0114] Therefore, equation (8) can be further written as:

[0115]

[0116] Substituting equation (5) into equation (10) yields the final aerodynamic characteristic equation:

[0117]

[0118] l0 is the initial length of PAM;

[0119] like Figure 2 As shown, for the length l1 of the flexor muscle PAM1, we can obtain the following from the law of cosines:

[0120]

[0121] The length l2 of the extensor PAM2 will be increased by the length increment Δl2 caused by elbow flexion, on top of the original length, that is:

[0122]

[0123] Equations (15) and (16) are the kinematic equations of the robot system. D0, D1, D2, D3, and D4 are the structural parameters of the actuator system, and θ is the angular position. Based on these equations, the lengths of the two PAMs at the desired angular position can be calculated. It should be noted that these kinematic equations only consider the axial deformation of the PAMs. In reality, changes in air pressure in the inner tube will also cause a certain radial deformation of the PAMs, which manifests as an increase or decrease in the muscle diameter. For the sake of convenience, radial deformation is not considered in this step, but is regarded as a modeling uncertainty and compensated for by a robust control method in subsequent steps.

[0124] Next, the dynamic equations of the robot system will be derived. Since the robot mechanism is driven by the antagonistic pair of PAMs, the rotational motion of the moving segment is determined by the combined effect of the driving forces of the two PAMs. According to the rigid body rotational law, the dynamic equations of the overall antagonistic mechanism are:

[0125]

[0126] Where J is the moment of inertia of the moving segment of the robot skeleton, τ is the equivalent external torque generated by the two PAMs, and τ g For gravitational torque, These are internal and external uncertainties arising from both external environmental disturbances and model simplification assumptions.

[0127] Since the design of subsequent control methods only requires a coarse system model and has a high tolerance for modeling uncertainties, the moving segments of the robot skeleton can be approximated as uniform thin rods. Based on this, the rotational inertia of the moving segments can be estimated as follows:

[0128]

[0129] Where m is the mass of the active segment, L r Let θ be the length of the moving segment. When the angular position of the moving segment is θ, the gravitational torque τ can be directly obtained according to the definition of torque. g ,Right now:

[0130]

[0131] If we denote the lever arms corresponding to the driving forces f1 and f2 generated by PAM1 and PAM2 as r1 and r2 respectively, then the resultant driving torque generated by the PAM antagonistic pair is:

[0132] τ=f1r1-f2r2 (20)

[0133] Step one has yielded the aerodynamic characteristic equations for a single PAM, and according to equation (11):

[0134]

[0135] Where K is a proportionality constant, P1 is the relative air pressure inside the flexor PAM1 tube, P2 is the relative air pressure inside the extensor PAM2 tube, and l 10 =D1+D2 is the initial length of the flexor muscle PAM1, l 20 =D3+D4 is the initial length of the extensor muscle PAM2. Based on the geometry of the robot mechanism, it is easy to obtain:

[0136]

[0137] Substituting equation (21) into equation (20), and approximating the lever arms of the two driving forces as constants R, we can obtain the driving torque expressed in matrix form, namely:

[0138]

[0139] in

[0140]

[0141] P = [P1 P2] T (25)

[0142] The antagonistic PAM configuration allows the robot system to synchronously increase or decrease the driving force of individual PAMs proportionally while maintaining a constant total driving torque, thereby obtaining joint stiffness suitable for different task environments. Therefore, to facilitate subsequent control of joint stiffness, a mathematical model needs to be established. The joint stiffness *s* is defined as the differential of the driving torque at the diagonal position, i.e.:

[0143]

[0144] Substituting equation (23) into equation (26), according to the matrix differentiation rule, we have:

[0145] s=-Γ θ (θ)P (27)

[0146] in

[0147]

[0148] Combining equations (23) and (27), we have the system model for the soft-actuated robot:

[0149]

[0150] Step 2: Design of a bionic controller based on a spiking neural network.

[0151] like Figure 3 As shown, the feedforward controller is the system model of the soft-actuated robot established in step 1, which is based on the current desired angular position θ. dWith the desired joint stiffness s d The approximate control command required to achieve this movement is roughly calculated, namely the PAM tube model air pressure P. The cerebellar controller optimizes its synaptic weights through online learning based on the current control error e, thereby fitting and modeling uncertainties and external environmental disturbances to generate the optimal compensating air pressure P for the current state. c The sum of the PAM inner tube model pressure and the compensation pressure will be used as the expected pressure P of the corresponding PAM. d The underlying control of each PAM employs a PID method, which calculates the control voltage u for the solenoid valve opening based on the air pressure deviation, and then adjusts the actual air pressure P of the corresponding PAM to the desired air pressure P. d When the muscle length and excitation air pressure (controlled by the control voltage u) are constant, the two PAMs will output a driving force F, which will then act together on the robot skeleton to generate actual angular motion θ.

[0152] like Figure 4 As shown, the cerebellar controller based on a spiking neural network designed in this method borrows its network topology and the naming of each neural layer from the corresponding cerebellar neural circuit entities. The two input layers of the cerebellar neural network are MF and IO, corresponding to the input and output signals. The MF layer receives sensor signals related to the current motion state, and the IO layer receives learning signals. The network's only output layer is DCN, responsible for outputting the compensation commands calculated by the controller. In short, the MF layer first processes the input sensor signals (including the actual angular position θ and the desired angular position θ)... d The process involves encoding neural pulses and mapping the pulse sequence to discrete excitation states of neurons in the GC layer. The neural pulses emitted by the GC layer neurons are then transmitted to the PC layer via the PF layer. The synaptic weights of the PF-PC layer are updated online using the learning signal e input from the IO layer. After the PC layer integrates the motion sensing states from the GC layer, its emitted neural pulses inhibit neurons in the DCN layer. Simultaneously, the DCN layer is activated by the MF layer. Finally, under the combined effects of inhibition and activation, the DCN layer generates an excitation state of neurons, which then emits output neural pulses encoding the current compensation signal. Decoding these pulses yields the output compensation instructions of the cerebellar controller.

[0153] The optimization of synaptic weights through online learning includes:

[0154] This method employs Modulated Spike-Timing-Dependent-Plasticity with Eligibility Trace (MSTDPET) as the learning rule for PF-PC synapses. The introduction of the eligibility trace allows the MSTDPET rule to better record temporal dynamics information compared to the traditional STDP rule. Therefore, even with a significant time delay in the feedback control error, the cerebellar network can still improve compensation accuracy under its supervision. The update law of the network weights under the MSTDPET learning rule is as follows:

[0155]

[0156] LTD ij (t)=η - ∫δ IO,j (t)E ij (t)dt (32)

[0157] Among them, w ij Here, δ represents the network weights, t represents time, δ represents the network impulse, and the subscripts GR and IO indicate the corresponding neuron layers. ij η represents the qualification trace of synapses ij. + and η - These are the learning rates for LTP (long-term potentiation) and LTD (long-term depression), respectively. To prevent overfitting during the learning process, the learning rate η is set to its maximum value η during the initial training phase. max Furthermore, the learning rate decays exponentially as learning progresses, eventually reaching a stable learning rate η. min ,Right now:

[0158]

[0159] Where τ0 is the time constant for learning rate decay. The traces of presynaptic and postsynaptic neural activity are updated according to the following rules:

[0160]

[0161] Where, τ + and τ - Let be the time constant for trace decay, and t0 be the initial time. The eligibility trace can be updated by integrating over the previous states, i.e.:

[0162]

[0163] Where τ1 and τ2 are the decay constants of the qualification trace.

[0164] This invention also provides a biomimetic control device for an "artificial muscle" soft actuator based on a spiking neural network, comprising:

[0165] The model building module is used to determine the system model of the soft actuator robot based on the aerodynamic characteristic equations of a single PAM "artificial muscle" soft actuator and the "artificial muscle" soft actuator system model.

[0166] The first control command module is used to derive a feedforward controller based on the system model of the soft-actuated robot; the input of the feedforward controller is the desired angular position θ. d With the desired joint stiffness s d The output is the PAM inner tube pressure P;

[0167] The second control command module is used to design a cerebellar-like controller based on a spiking neural network, based on the cerebellar neural circuit of mammals; the inputs of the cerebellar-like controller are the actual angular position θ and the desired angular position θ. d The output is the compensation air pressure P. c ;

[0168] The control module is used to combine the control commands output by the feedforward controller and the cerebellum-like controller for control.

[0169] It should be noted that the device embodiment shown in this embodiment matches the content of the above method embodiment, and the content of the above method embodiment can be referred to, and will not be repeated here.

[0170] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

[0171] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only.

[0172] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A biomimetic control method for a soft actuator of "artificial muscle" based on a spiking neural network, characterized in that, include: Based on the aerodynamic characteristic equations of a single PAM "artificial muscle" soft actuator and the "artificial muscle" soft actuator system model, the system model of the soft actuator robot is determined. Derive the feedforward controller from the system model of the soft actuator robot; The input to the feedforward controller is the desired angular position. With desired joint stiffness The output is the PAM inner tube pressure. ; Based on mammalian cerebellar neural circuits, a cerebellar-like controller based on a spiking neural network is designed; the input of the cerebellar-like controller is the actual angular position. and the expected angular position The output is the compensation air pressure. ; Control is achieved by combining control commands output from the feedforward controller and the cerebellar-like controller. The cerebellar-like controller is based on the current control error. By optimizing its synaptic weights through online learning, it can fit and model uncertainties and external environmental disturbances, thereby generating the optimal compensating pressure for the current state. ; Optimize your synaptic weights through online learning, including: The MSTDPET learning algorithm applies the following rules to network weight updates: (30) (31) (32) in, For network weights, For time, For network pulses, the subscripts GR and IO represent the corresponding neuron layers. Indicates synapse The qualification traces, and The learning rates are for LTP and LTD, respectively; during the initial training phase, the learning rate... Set to maximum value Furthermore, the learning rate decays exponentially as learning progresses, eventually reaching a stable learning rate. ,Right now: (33) in, The time constant for learning rate decay; the traces of pre- and post-synaptic neural activity are updated according to the following rules: (34) in, and The time constant of trace decay, The initial time is used; the qualification trace can be updated by integrating the previous state, that is: (35) (36) in, and is the decay constant of the qualification trace.

2. The method according to claim 1, characterized in that, The construction of the aerodynamic characteristic equations for a single PAM "artificial muscle" soft actuator includes: When gas is introduced into the PAM inner tube, the input energy obtained by the drive system is: (1) in, For input energy micro-elements, , These are the internal air pressure and ambient atmospheric pressure (1 atm ≈ Pa), The relative air pressure inside the tube. For the main internal space, For the displacement of the inner tube, Let the cross-sectional area of ​​the inner tube be a infinitesimal element. Let the inner tube volume be a infinitesimal element; according to the definition, the driver output energy is: (2) in, To output energy micro-elements, The output axial driving force; neglecting the frictional energy loss and energy storage of the PAM system, according to the law of conservation of energy, the input energy equals the output energy, that is: (3) Substituting equations (1) and (2) into the equations, we get: (4) To calculate the differential of the PAM volume with respect to the PAM length If PAM is considered as a cylinder with a length of... The diameter of the base is Since the PAM surface uses a nylon braided mesh, the elasticity of each braided thread is very small; therefore, it can be assumed that the braided threads are of a fixed length. ; This indicates the angle between the braided yarn and the PAM axis. The number of turns of the braided yarn; based on the geometric relationship, the geometric parameters of PAM can be obtained as follows: (5) (6) Therefore, the volume of PAM is: (7) Therefore, equation (4) can be further written in the following form: (8) When PAM is in its initial state, its output driving force Since the initial value is 0, let equation (8) be 0, and initially... Right now satisfy: (9) Therefore, equation (8) can be further written as: (10) Substituting equation (5) into equation (10) yields the final aerodynamic characteristic equation: (11) in, This is the initial length of the PAM.

3. The method according to claim 2, characterized in that, The construction of the system model for the soft-actuated robot includes: The "artificial muscle" soft actuator system model includes the flexor muscle PAM1 and the extensor muscle PAM2; Regarding the length of flexor PAM1 By the Law of Cosines, we can obtain: (15) Length of extensor PAM2 The length increment caused by elbow flexion will be added to the original length. ,Right now: (16) Equations (15) and (16) are the kinematic equations of the robot system. For the structural parameters of the driver system, Given the angular position, the equation can be used to calculate the required lengths of the two PAM blocks at the desired angular position. Next, the dynamic equations of the robot system will be derived: Since the robot mechanism is driven by the PAM antagonistic pair, the rotational motion of the moving segment is determined by the combined effect of the two PAM driving forces; according to the rigid body rotation law, the dynamic equation of the overall antagonistic mechanism is: (17) in, Let the moment of inertia be the rotational inertia of the moving segments of the robot skeleton. The equivalent external torque generated by the two PAMs together. For gravitational torque, These are internal and external uncertainties arising from both external environmental disturbances and model simplification assumptions. Treating the moving segments of the robot skeleton as uniform thin rods, the moment of inertia of the moving segments can be estimated as follows: (18) in, For the quality of the active segment, The length of the active segment; when the angular position of the active segment is... At that time, the gravitational torque can be directly obtained according to the definition of torque. ,Right now: (19) If the driving forces generated by PAM1 and PAM2 are combined and The corresponding lever arms are denoted as follows: and The resultant driving torque generated by the PAM antagonistic pair is: (20) According to equation (11): (21) in It is a proportionality constant. The air pressure inside the PAM1 tube of the flexor muscles. This refers to the air pressure inside the PAM2 tube of the extensor muscle. This represents the initial length of the flexor PAM1 muscle. Let PAM2 be the initial length of the extensor muscle; based on the geometric relationship of the robot mechanism, we obtain: (22) Substituting equation (21) into equation (20), and treating the lever arms of the two driving forces as constants. This yields the driving torque expressed in matrix form, namely: (23) in (24) (25) Define joint stiffness The derivative of the driving torque at the diagonal position is: (26) Substituting equation (23) into equation (26), according to the matrix differentiation rule, we have: (27) in (28) Combining equations (23) and (27), we have the system model for the soft-actuated robot: (29)。 4. The method according to claim 1, characterized in that, The feedforward controller is a system model for a soft-actuated robot, which determines the desired angular position based on the current desired position. With desired joint stiffness Calculate the control commands required to achieve the target motion, i.e., the air pressure of the PAM inner tube model. .

5. The method according to claim 1, characterized in that, The cerebellar controller comprises an MF layer, an IO layer, a GC layer, a PC layer, and a DCN layer. First, the MF layer processes the input sensor signal, i.e., the actual angular position. and the expected angular position Neural pulse encoding is performed, and the pulse sequence is then mapped to the discrete excitation states of neurons in the GC layer. The neural pulses fired by the GC layer neurons are then transmitted to the PC layer via the PF layer. The synaptic weights of the PF-PC layer utilize the learning signal from the I / O layer, i.e., the control error. Conduct online learning updates; After the PC layer integrates the motion sensing state from the GC layer, its emitted neural impulses will inhibit neurons in the DCN layer. Simultaneously, the DCN layer is activated by neurons from the MF layer. Ultimately, under the combined effects of inhibition and activation, the DCN layer generates an excited neuronal state, which then emits output neural impulses encoding the current compensation signal. Decoding these impulses yields the output compensation command of the cerebellar controller, namely, the compensation air pressure. .

6. The method according to claim 1, characterized in that, The control command output by the feedforward controller is the PAM inner tube pressure, and the control command output by the cerebellar controller is the compensation pressure. Combining the control commands output by the feedforward controller and the cerebellar controller yields the sum of the PAM inner tube pressure and the compensation pressure. This sum is the desired PAM pressure. .

7. The method according to claim 1, characterized in that, The underlying control of each PAM uses the PID method, which calculates the control voltage for the solenoid valve opening based on the air pressure deviation. This allows for the adjustment of the actual air pressure of the corresponding PAM. To the desired air pressure To output driving force These forces, in turn, act together on the robot's skeleton to generate actual angular motion. .

8. A biomimetic control device for an "artificial muscle" soft actuator based on a spiking neural network, characterized in that, To implement the method according to any one of claims 1-7, comprising: The model building module is used to determine the system model of the soft actuator robot based on the aerodynamic characteristic equations of a single PAM "artificial muscle" soft actuator and the "artificial muscle" soft actuator system model. The first control command module is used to derive a feedforward controller based on the system model of the soft-actuated robot; the input of the feedforward controller is the desired angular position. With desired joint stiffness The output is the PAM inner tube pressure. ; The second control command module is used to design a cerebellar-like controller based on a spiking neural network, based on the cerebellar neural circuits of mammals; the input of the cerebellar-like controller is the actual angular position. and the expected angular position The output is the compensation air pressure. ; The control module is used to combine the control commands output by the feedforward controller and the cerebellum-like controller for control.

Images