Artificial muscle soft driver bionic control method based on pulse neural network

Abstract

The invention discloses an artificial muscle soft driver bionic control method based on a spiking neural network, and the method employs a feedforward controller based on a rough inverse model as coarse adjustment, employs a cerebellar-like controller based on the spiking neural network as fine adjustment, and compensates the modeling uncertainty and environment interaction uncertainty in real time through an online learning method. According to the invention, a representative'artificial muscle 'soft driver-pneumatic artificial muscle is used, a pulse neural network simulating a working mechanism of a cerebellar nerve of a mammal is realized, and efficient, robust and adaptive control of the highly nonlinear and time-varying'artificial muscle' soft driver is realized.

Description

Technical Field

[0001] The present invention relates to the field of control science and engineering, and particularly to a bionic control method for an "artificial muscle" soft actuator based on a spiking neural network. Background Art

[0002] "Artificial muscle" soft actuators have advantages such as high energy density, fast response speed, and strong environmental compliance, and can better meet the requirements of compliance, safety, and comfort in human-robot interaction, thus receiving extensive attention in the field of robotics research. However, the inherent characteristics of "artificial muscle" soft actuators also pose many challenges for control in robot applications. On the one hand, the unique material properties of the soft material itself and the actuation method of the soft actuator make the system exhibit strong time-varying and non-linear characteristics. Especially in the high-speed and large-deformation motion state, viscoelasticity, elastic hysteresis, creep, and mechanical friction are more significant, which makes it extremely difficult to mathematically describe the controlled object and design the corresponding controller. On the other hand, the characteristics of soft actuators make classical control methods lack in terms of both control accuracy and system compliance. Although some researchers have proposed methods such as force / position hybrid control and impedance control to deal with compliant force control problems, they are more oriented towards traditional "hard drive" robots. "Artificial muscle" soft actuators are highly similar to biological muscles in terms of actuation mechanism, material properties, etc. However, organisms can effectively coordinate and control their skeletal muscles through the action of the nervous system to achieve complex and dexterous movements that are currently difficult to achieve in artificial control systems, and at the same time have excellent environmental adaptability. Therefore, some researchers have been inspired by this and tried to explore the control method of "artificial muscle" soft actuators from the perspective of life science. Summary of the Invention

[0003] The purpose of the present invention is to provide a bionic control method for an "artificial muscle" soft actuator based on a spiking neural network in view of the deficiencies of the existing control methods for "artificial muscle" soft actuators.

[0004] The purpose of the present invention is achieved through the following technical solutions: A bionic control method for an "artificial muscle" soft actuator based on a spiking neural network, comprising:

[0005] Determining the system model of the soft actuator robot based on the pneumatic characteristic equation of a single PAM "artificial muscle" soft actuator and the system model of the "artificial muscle" soft actuator system;

[0006] Deriving a feedforward controller based on the system model of the soft actuator robot;

[0007] Designing a cerebellum-like controller based on a spiking neural network based on the mammalian cerebellar neural circuit;

[0008] Control is carried out by combining the control instructions output by the feedforward controller and the cerebellum-like controller.

[0009] Furthermore, the construction of the pneumatic characteristic equation of a single PAM "artificial muscle" soft actuator includes:

[0010] After filling the inner tube of the PAM with gas, the input energy obtained by the actuator system is:

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

[0012] Among them, dW in is the input energy microelement, P and P0 are the inner tube air pressure and the ambient atmospheric pressure (1 atm ≈ 1.01×10 5 Pa) respectively, P′ is the relative air pressure of the inner tube, I is the total inner tube space, dl is the inner tube displacement microelement, ds is the inner tube cross-sectional area microelement, and dV is the inner tube volume microelement; according to the definition, the output energy of the actuator is:

[0013] dW out = -Fdl (2)

[0014] Among them, dW out is the output energy microelement, and F is the output axial driving force; ignoring the frictional energy consumption and energy storage of the PAM system, so according to the law of conservation of energy, the input energy is equal to the output energy, that is:

[0015] dW in = dW out (3)

[0016] Substituting equations (1) and (2) gives:

[0017] P′dV = -Fdl

[0018]

[0019] In order to calculate the differential dV / dl of the PAM volume with respect to the PAM length, the PAM is regarded as a cylinder with a length of l and a bottom diameter of d; since the surface of the PAM uses a nylon woven mesh and the elasticity of each woven wire is very small, it can be assumed that the woven wire has a fixed length b. α represents the angle between the woven wire and the PAM axis, and n is the number of turns of the woven wire winding. According to the geometric relationship, the geometric parameters of the PAM can be obtained as:

[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 the initial state, the driving force F output by it is 0. Therefore, let Equation (8) be 0, and the initial α, that is, α0, satisfies:

[0027]

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

[0029]

[0030] Substituting Equation (5) into Equation (10) can obtain the final pneumatic characteristic equation:

[0031]

[0032] Among them, l0 is the initial length of PAM.

[0033] Furthermore, the construction of the system model of the soft actuator robot includes:

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

[0035] For the length l1 of the flexor PAM1, according to the cosine theorem, we can get:

[0036]

[0037] The length l2 of the extensor PAM2 will increase by the length increment Δl2 caused by the elbow flexion movement on the basis of the original length, that is:

[0038]

[0039] Equations (15) and (16) are the kinematic equations of the robot system. D0, D1, D2, D3, D4 are the structural parameters of the actuator system, and θ is the angular position. According to this equation, the lengths that the two PAMs should have at the expected angular position can be calculated;

[0040] Next, the dynamic equation of the robot system will be derived: Since the robot mechanism is driven by a PAM antagonistic pair, the rotational movement of the active segment is determined by the combined action of the driving forces of the two PAMs; According to the rigid body fixed-axis rotation law, the dynamic equation of the overall antagonistic mechanism is:

[0041]

[0042] Among them, J is the moment of inertia of the movable segment of the robot skeleton, τ is the equivalent external torque generated by the two PAMs together, and τ g is the gravitational torque, is the internal and external uncertainty jointly generated by external environmental disturbances and model simplification assumptions.

[0043] Regarding the movable segment of the robot skeleton as a homogeneous slender rod, the moment of inertia of the movable segment can be estimated as follows:

[0044]

[0045] Among them, m is the mass of the movable segment, and L r is the length of the movable segment. When the angular position of the movable segment is θ, the gravitational torque τ g can be directly obtained according to the definition of torque, that is:

[0046]

[0047] If the force arms corresponding to the driving forces f1 and f2 generated by PAM1 and PAM2 are denoted as r1 and r2 respectively, then the combined driving torque generated by the PAM antagonistic pair is:

[0048] τ = f1r1 - f2r2 (20)

[0049] According to Equation (11), we have:

[0050]

[0051] Among them, K is the proportionality constant, P1 is the internal tube air pressure of the flexor PAM1, P2 is the internal tube air pressure of the extensor PAM2, and l 10 = D1 + D2 is the initial length of the flexor PAM1, and l 20 = D3 + D4 is the initial length of the extensor PAM2; according to the geometric relationship of the robot mechanism, we get:

[0052]

[0053] Substituting Equation (21) into Equation (20) and regarding the force arms of the two driving forces as constants R, we obtain the driving torque expressed in matrix form, that is:

[0054]

[0055] Among them

[0056]

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

[0058] Define the joint stiffness \(s\) as the differential of the driving torque with respect to the angular position, i.e.:

[0059]

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

[0061] \(s = -\Gamma\) θ (\(\theta\))\(P\) (27)

[0062] where

[0063]

[0064] Combining Equation (23) and Equation (27), we get the system model of the soft actuator robot, i.e.:

[0065]

[0066] Furthermore, the feed - forward controller is the system model of the soft actuator robot. It calculates the control command required to achieve the target motion, i.e., the air pressure \(P\) of the PAM inner tube model, based on the current desired angular position \(\theta\) d and the desired joint stiffness \(s\) d .

[0067] Furthermore, the cerebellum - like controller includes the MF layer, IO layer, GC layer, PC layer and DCN layer. First, the MF layer performs neural pulse encoding on the input sensing signals, i.e., the actual angular position \(\theta\) and the desired angular position \(\theta\) d , and then maps the pulse sequence to the discrete excitation state of the neurons in the GC layer. Then, the nerve pulses emitted by the neurons in the GC layer will be transmitted to the PC layer via PF. The synaptic weights of PF - PC are updated online with the help of the learning signal input from the IO layer, i.e., the control error \(e\). After integrating the motion sensing state from the GC layer in the PC layer, the nerve pulses it emits will inhibit the neurons in the DCN layer. At the same time, the DCN layer is also affected by the activation from the MF layer. Finally, the DCN layer will generate the excitation state of the neurons under the combined action of inhibition and activation, and then emit output nerve pulses encoding the current compensation signal. Decoding this pulse can obtain the output compensation command of the cerebellum - like controller, i.e., the compensation air pressure \(P\) c .

[0068] Furthermore, the cerebellum - like controller optimizes its own synaptic weights through online learning based on the current control error \(e\), so as to fit the modeling uncertainty and external environmental disturbances, and generate the optimal compensation air pressure \(P\) in the current state c .

[0069] Furthermore, optimize its synaptic weights through online learning, including:

[0070] The update rule of the MSTDPET learning rule for the network weights is as follows:

[0071]

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

[0073] where w ij is the network weight, t is the time, δ is the network pulse, the subscripts GR and IO represent the corresponding neuron layers, and E ij represents the eligibility trace of synapse ij, and η + and η - are the learning rates of LTP and LTD respectively; in the initial training stage, the learning rate η is set to the maximum value η max , and decays exponentially as learning progresses, and finally reaches a stable learning rate η min , that is:

[0074]

[0075] where τ0 is the time constant for the decay of the learning rate. The traces of pre- and post-synaptic neural activities are updated according to the following rule:

[0076]

[0077] where τ + and τ - are the time constants for the decay of the traces, and t0 is the initial time. The update of the eligibility trace can be achieved by integrating the previous state, that is:

[0078]

[0079] where τ1 and τ2 are the decay constants of the eligibility trace.

[0080] Furthermore, the control command output by the feedforward controller is the air pressure in the inner tube of the PAM, the control command output by the cerebellum-like controller is the compensation air pressure, and the combined control command output by the feedforward controller and the cerebellum-like controller is the sum of the air pressure in the inner tube of the PAM and the compensation air pressure. The sum of the air pressure in the inner tube of the PAM and the compensation air pressure is the desired air pressure P d .

[0081] Further, the PID method is adopted for the underlying control of each PAM. It calculates the control voltage u of the solenoid valve opening according to the air pressure deviation, and then adjusts the actual air pressure P of the corresponding PAM to the desired air pressure P d , so as to output the driving force F, which then acts together on the robot skeleton to generate the actual angular motion θ.

[0082] The present invention also provides a bionic control device for an "artificial muscle" soft actuator based on a spiking neural network, including:

[0083] A model construction module, which is used to determine the system model of the soft actuator robot based on the pneumatic characteristic equation of a single PAM "artificial muscle" soft actuator and the system model of the "artificial muscle" soft actuator system;

[0084] A first control instruction module, which is used to derive a feedforward controller based on the system model of the soft actuator robot; the input of the feedforward controller is the desired angular position θ d and the desired joint stiffness s d , and the output is the air pressure P inside the PAM;

[0085] A second control instruction module, which is used to design a cerebellum-like controller based on a spiking neural network based on the mammalian cerebellar neural circuit; the input of the cerebellum-like controller is the actual angular position θ and the desired angular position θ d , and the output is the compensation air pressure P c ;

[0086] A control module, which is used to perform control by combining the control instructions output by the feedforward controller and the cerebellum-like controller.

[0087] The beneficial effect of the present invention is that the present invention uses a cerebellum-like controller based on a spiking neural network to achieve efficient, robust, and adaptive control of the "artificial muscle" soft actuator, solves the problems of strong nonlinearity and time-variation faced by existing methods, and improves the environmental adaptability of the control system on the premise of taking into account the control accuracy. Description of the Drawings

[0088] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0089] Figure 1 It is a geometric characteristic diagram of the "artificial muscle" soft actuator in the deformed state.

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

[0091] Figure 3 It is the overall bionic control architecture diagram.

[0092] Figure 4 It is a schematic diagram of a cerebellum-like controller based on a spiking neural network. Specific implementation manners

[0093] The present invention will be described in detail below with reference to the accompanying drawings. Without conflict, the features in the following embodiments and implementation manners can be combined with each other.

[0094] A bionic control method for an "artificial muscle" soft actuator based on a spiking neural network according to the present invention includes the following steps:

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

[0096] As Figure 1 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] where dW in is the input energy microelement, P and P0 are the inner tube air pressure and the ambient atmospheric pressure (1 atm ≈ 1.01×10 5 Pa) respectively, P′ is the relative air pressure of the inner tube, I is the total inner tube space, dl is the inner tube displacement microelement, ds is the inner tube cross-sectional area microelement, and dV is the inner tube volume microelement. According to the definition, the output energy of the actuator is:

[0099] dW out = -Fdl (2)

[0100] where dW out is the output energy microelement, and F is the output axial driving force. In the static or quasi-static motion state, the friction energy consumption and energy storage of the PAM system can be ignored. Therefore, according to the law of conservation of energy, the input energy is equal to the output energy, that is:

[0101] dW in = dW out (3)

[0102] Substituting equations (1) and (2) gives:

[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 a cylinder as shown in Figure 1 Figure [not provided], with a length of l and a base diameter of d. Since the PAM surface is made of a nylon braided mesh and the elasticity of each braided wire is very small, it can be assumed that the braided wire has a fixed length b. α represents the angle between the braided wire and the PAM axis, and n is the number of turns of the braided wire. According to the geometric relationship, the geometric parameters of the PAM can be obtained as:

[0106] l = bcosα (5)

[0107]

[0108] Therefore, the PAM volume is:

[0109]

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

[0111]

[0112] When the PAM is in the initial state, the driving force F it outputs is 0. Therefore, 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) gives the final pneumatic characteristic equation:

[0117]

[0118] l0 is the initial length of the PAM;

[0119] As shown in Figure 2 Figure [not provided], for the length l1 of the flexor PAM1, by the cosine theorem, we have:

[0120]

[0121] The length l2 of the extensor PAM2 will increase by the length increment Δl2 caused by the elbow flexion movement on the basis of the original length, i.e.:

[0122]

[0123] Equations (15) and (16) are the kinematic equations of the robot system. D0, D1, D2, D3, D4 are the structural parameters of the drive system, and θ is the angular position. According to this equation, the lengths of the two PAMs at the desired angular position can be calculated. It should be noted that this kinematic equation only considers the axial deformation of the PAM. In actual situations, the change in air pressure in the inner tube will also cause a certain radial deformation of the PAM, manifested as an increase or decrease in the muscle diameter. For the convenience of problem research, the radial deformation is not considered in this step, but regarded as modeling uncertainty and compensated by the robust control method in the subsequent steps.

[0124] Next, the dynamic equation of the robot system will be derived. Since the robot mechanism is driven by the PAM antagonistic pair, the rotational motion of the active segment is determined by the combined action of the driving forces of the two PAMs. According to the law of fixed-axis rotation of a rigid body, the dynamic equation of the overall antagonistic mechanism is:

[0125]

[0126] where J is the moment of inertia of the active segment of the robot skeleton, τ is the equivalent external torque generated by the two PAMs, τ g is the gravitational torque, and is the internal and external uncertainty jointly generated by external environmental disturbances and model simplification assumptions.

[0127] Since the design of the subsequent control method only requires a rough system model and has a large tolerance for modeling uncertainty, the active segment of the robot skeleton can be approximated as a homogeneous thin rod. Based on this, the moment of inertia of the active segment can be estimated as:

[0128]

[0129] where m is the mass of the active segment, and L r is the length of the active segment. When the angular position of the active segment is θ, the gravitational torque τ g can be directly obtained according to the definition of torque, that is:

[0130]

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

[0132] τ = f1r1 - f2r2 (20)

[0133] In Step 1, the pneumatic characteristic equation of a single PAM has been obtained. According to Equation (11), we have:

[0134]

[0135] where K is a proportionality constant, P1 is the relative air pressure in the inner tube of the flexor PAM1, P2 is the relative air pressure in the inner tube of the extensor PAM2, and l 10 = D1 + D2 is the initial length of the flexor PAM1, and l 20 = D3 + D4 is the initial length of the extensor PAM2. According to the geometric relationship of the robotic mechanism, it is easy to obtain:

[0136]

[0137] Substituting Equation (21) into Equation (20) and approximating the moment arms of the two driving forces as a constant R, the driving torque expressed in matrix form can be obtained, that is:

[0138]

[0139] where

[0140]

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

[0142] The antagonistic PAM configuration enables the robotic system to synchronously increase or decrease the driving force of a single PAM in a certain proportion while keeping the total driving torque unchanged, thereby obtaining joint stiffness suitable for different task environments. Therefore, in order to facilitate subsequent control of joint stiffness, it is necessary to establish its mathematical model. Define the joint stiffness s as the differential of the driving torque with respect to the angular position, that is:

[0143]

[0144] Substituting Equation (23) into Equation (26), according to the matrix derivative rule:

[0145] s = -Γ θ (θ)P (27)

[0146] where

[0147]

[0148] Combining Equation (23) and Equation (27), that is, the system model of the soft actuator robot:

[0149]

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

[0151] As Figure 3 shown, the feedforward controller is the system model of the soft actuator robot established in Step 1, which calculates according to the current desired angular position θ dWith the desired joint stiffness s d Roughly calculate the approximate control instructions required to achieve this movement, that is, the air pressure P of the PAM inner tube model; the cerebellum-like controller optimizes its own synaptic weights based on the current control error e through online learning, so as to fit the modeling uncertainty and external environmental disturbances, and generate the optimal compensation air pressure P under the current state c , The sum of the air pressure of the PAM inner tube model and the compensation air pressure will be used as the desired air pressure P of the corresponding PAM d . The bottom layer control of each PAM adopts the PID method, which calculates the control voltage u of the solenoid valve opening according to 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 the excitation air pressure (controlled by the control voltage u) of the two PAMs are certain, they will output the driving force F, and then jointly act on the robot skeleton to generate the actual angular movement θ

[0152] As Figure 4 shown, for the cerebellum-like controller based on the spiking neural network designed by this method, the topological structure of the network and the naming of each neural layer are all borrowed from the corresponding cerebellar neural circuit entities. The two input layers of the cerebellum-like neural network are MF and IO respectively, corresponding to the input and output signals. The MF layer inputs the sensing signals related to the current movement state, and the IO layer inputs the learning signals. The only output layer of the network is DCN, which is responsible for outputting the compensation instructions calculated by the controller. Briefly speaking, first, the MF layer performs neural pulse coding on the input sensing signals (including the actual angular position θ and the desired angular position θ d ), and then maps the pulse sequence to the discrete excitation state of the GC layer neurons; then the neural pulses emitted by the GC layer neurons will be transmitted to the PC layer via PF, and the synaptic weights of PF-PC are updated online with the learning signal e input from the IO layer; after integrating the movement sensing state from the GC layer in the PC layer, the neural pulses it emits will have an inhibitory effect on the DCN layer neurons. At the same time, the DCN layer is also affected by the activation from the MF layer; finally, the DCN layer will generate the excitation state of the neurons under the combined action of inhibition and activation, and then emit the output neural pulses encoding the current compensation signal. Decoding this pulse can obtain the output compensation instructions of the cerebellum-like controller

[0153] The optimization of its own synaptic weights by the way of online learning includes:

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

[0155]

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

[0157] where w ij is the network weight, t is time, δ is the network spike, the subscripts GR and IO represent the corresponding neuron layers, and E ij represents the eligibility trace of synapse ij, and η + and η - are the learning rates of LTP (long-term potentiation) and LTD (Long-Term Depression), respectively. To prevent overfitting during the learning process, in the initial training stage, the learning rate η is set to the maximum value η max , and it decays exponentially as learning progresses and finally reaches the stable learning rate η min , that is:

[0158]

[0159] where τ0 is the time constant for the decay of the learning rate. The traces of pre- and post-synaptic neural activities are updated according to the following rules:

[0160]

[0161] where τ + and τ - are the time constants for trace decay, and t0 is the initial time. The update of the eligibility trace can be achieved by integrating the previous state, that is:

[0162]

[0163] where τ1 and τ2 are the decay constants of the eligibility trace.

[0164] The present invention also provides a bionic control device for an "artificial muscle" soft actuator based on a spiking neural network, including:

[0165] A model construction module for determining the system model of the soft actuator robot based on the pneumatic characteristic equation of a single PAM "artificial muscle" soft actuator and the system model of the "artificial muscle" soft actuator system;

[0166] A first control instruction module for deriving a feedforward controller based on the system model of the soft actuator robot; the input of the feedforward controller is the desired angular position θ d and the desired joint stiffness s d , and the output is the inner tube air pressure P of the PAM;

[0167] A second control instruction module for designing a cerebellum-like controller based on a spiking neural network based on the mammalian cerebellar neural circuit; the input of the cerebellum-like controller is the actual angular position θ and the desired angular position θ d , and the output is the compensation air pressure P c ;

[0168] A control module for controlling by combining the control instructions output by the feedforward controller and the cerebellum-like controller.

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

[0170] The above embodiments are only used to illustrate the design idea and characteristics 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 according to the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

[0171] Those skilled in the art will readily think of other implementation schemes of the present application after considering the specification and practicing the content disclosed herein. The present application aims to cover any variations, uses, or adaptive changes of the present application, and these variations, uses, or adaptive changes follow the general principles of the present application and include the common general knowledge or conventional technical means in the technical field not disclosed in the present application. The specification and embodiments are only regarded as exemplary.

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

Claims

1. A bionic control method for an "artificial muscle" soft actuator based on a pulse neural network, characterized in that: include: Based on the aerodynamic characteristic equation 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; A feedforward controller is derived based on the system model of the soft actuator robot; the input of the feedforward controller is the desired angular position θ d and the desired joint stiffness s d , the output is the PAM inner tube pressure P; Based on the mammalian cerebellar neural circuit, a cerebellum-like controller based on a spiking neural network is designed; the input of the cerebellum-like controller is the actual angular position θ and the desired angular position θ d , the output is the compensation air pressure P c ; The control is performed by combining the control instructions output by the feedforward controller and the cerebellum-like controller.

2. The method according to claim 1, characterized in that: The construction of the aerodynamic characteristic equation of a single PAM "artificial muscle" soft actuator includes: When the gas is filled into the PAM inner tube, the input energy obtained by the drive system is: dW in =∫ I (P-P0)dl·ds=(P-P0)∫ I dl·ds=P′dV (1) Among them, dW in is the input energy element, P and P0 are the inner tube pressure and the ambient atmospheric pressure (1atm≈1.01×10 5 Pa), P' is the relative air pressure of the inner tube, I is the total inner tube space, dl is the inner tube displacement microelement, ds is the inner tube cross-sectional area microelement, dV is the inner tube volume microelement; according to the definition, the output energy of the driver is: dW out =-Fdl (2) Among them, dW out is the output energy element, and F is the output axial driving force; ignoring the friction energy consumption and energy storage of the PAM system, according to the law of conservation of energy, the input energy is equal to the output energy, that is: dW in =dW out (3) Substituting equation (1) and equation (2) into the equation, we can obtain: P′dV=-Fdl In order to calculate the differential of PAM volume to PAM length dV / dl, PAM is regarded as a cylinder with a length of l and a bottom diameter of d. Since the surface of PAM is made of nylon mesh, the elasticity of each braided wire is very small, so the braided wire can be assumed to be a fixed length b. α represents the angle between the braided wire and the axis of PAM, and n is the number of winding turns of the braided wire. According to the geometric relationship, the geometric parameters of PAM are: l=bcosα (5) Therefore, the volume of PAM is: Therefore, formula (4) can be further written as follows: When PAM is in the initial state, its output driving force F is 0, so let equation (8) be 0, and the initial α, i.e. α0, satisfies: Therefore, formula (8) can be further written as: Substituting equation (5) into equation (10) we can get the final aerodynamic characteristic equation: Among them, l0 is the initial length of PAM.

3. The method according to claim 2, characterized in that The construction of the system model of the soft actuator robot includes: The "artificial muscle" soft actuator system model includes flexor PAM1 and extensor PAM2; For the length l1 of the flexor muscle PAM1, the cosine theorem gives: The length l2 of the extensor muscle PAM2 will be increased by the length increment Δl2 caused by the elbow flexion movement on the basis of the original length, that is: Formula (15) and Formula (16) are the kinematic equations of the robot system. D0, D1, D2, D3, and D4 are the structural parameters of the drive system, and θ is the angular position. According to the equation, the lengths of the two PAMs at the desired angular position can be calculated. Next, the dynamic equation of the robot system will be derived: Since the robot mechanism is driven by the PAM antagonistic pair, the rotation of the active segment is determined by the cooperation of the two PAM driving forces; according to the law of rigid body fixed axis rotation, the dynamic equation of the overall antagonistic mechanism is: Where J is the moment of inertia of the active segment of the robot skeleton, τ is the equivalent external torque generated by the two PAMs, and τ g is the gravitational moment, Internal and external uncertainties are caused by external environmental disturbances and simplified assumptions of the model; The active segment of the robot skeleton is regarded as a homogeneous thin rod, and the moment of inertia of the active segment can be estimated as: Where m is the mass of the active segment, L r is the length of the active segment; when the angular position of the active segment is θ, the gravity moment τ can be directly obtained according to the definition of moment g ,Right now: If the force arms corresponding to the driving forces f1 and f2 generated by PAM1 and PAM2 are recorded as r1 and r2 respectively, the combined driving torque generated by the PAM antagonistic pair is: τ=f1r1-f2r2 (20) According to formula (11), we have: Where K is the proportional constant, P1 is the air pressure in the inner tube of the flexor PAM1, P2 is the air pressure in the inner tube of the extensor PAM2, l 10 =D1+D2 is the initial length of flexor muscle PAM1, l 20 =D3+D4 is the initial length of the extensor muscle PAM2; according to the geometric relationship of the robot mechanism, we get: Substituting equation (21) into equation (20), and regarding the force arms of the two driving forces as a constant R, the driving torque expressed in matrix form is obtained, namely: in P=[P1 P2] T (25) The joint stiffness s is defined as the differential of the driving torque with respect to the angular position, that is: Substituting formula (23) into formula (26), according to the matrix derivation rule, we have: s=-Γ θ (i)P (27) in Combining equations (23) and (27), we have the system model of the soft actuator robot:

4. The method according to claim 1, characterized in that: The feedforward controller is a system model of the soft actuator robot, which is based on the current desired angular position θ d and the desired joint stiffness s d Calculate the control instructions required to achieve the target motion, that is, the air pressure P of the PAM inner tube model.

5. The method according to claim 1, characterized in that The cerebellum-like controller includes MF layer, IO layer, GC layer, PC layer and DCN layer. First, the MF layer processes the input sensor signal, i.e., the actual angular position θ and the expected angular position θ. d Perform neural pulse encoding, and then map the pulse sequence into the discrete excitation state of the GC layer neurons; then the neural pulses emitted by the GC layer neurons will be transmitted to the PC layer via the PF, and the PF-PC synaptic weights will be updated online with the help of the learning signal input from the IO layer, that is, the control error e; After the PC layer integrates the motion sensing state from the GC layer, the neural pulses it releases will inhibit the neurons in the DCN layer. At the same time, the DCN layer is also activated by the MF layer. Finally, the DCN layer will produce an excited state of neurons under the combined action of inhibition and activation, and then release output neural pulses encoding the current compensation signal. By decoding this pulse, the output compensation instruction of the cerebellar-like controller can be obtained, that is, the compensation air pressure P c .

6. The method according to claim 1, characterized in that The cerebellum-like controller optimizes its own synaptic weights through online learning based on the current control error e, thereby fitting the modeling uncertainty and external environmental disturbances to generate the optimal compensation air pressure P in the current state. c .

7. The method according to claim 6, characterized in that Optimize your own synaptic weights through online learning, including: The update rules of the MSTDPET learning rule for network weights are as follows: LTD ij (t)=η - ∫δ IO,j (t)E ij (t)dt (32) Among them, w ij is the network weight, t is the time, δ is the network pulse, subscripts GR and IO represent the corresponding neuron layers, E ij represents the qualification trace of synapse ij, η + and η - are the learning rates of LTP and LTD respectively; in the initial training stage, the learning rate η is set to the maximum value η max , and it decays exponentially as learning progresses, eventually reaching a stable learning rate η min ,Right now: Among them, τ0 is the time constant of learning rate decay; the traces of previous and next synaptic neural activities are updated according to the following rules: Among them, τ + and τ - is the time constant of trace decay, t0 is the initial time; the update of the qualification trace can be achieved by integrating the previous state, that is: Where τ1 and τ2 are the decay constants of the eligibility trace.

8. The method according to claim 1, characterized in that: The control command output by the feedforward controller is the PAM inner tube air pressure, and the control command output by the cerebellum-like controller is the compensation air pressure. The control command output by the feedforward controller and the cerebellum-like controller is the sum of the PAM inner tube air pressure and the compensation air pressure. The sum of the PAM inner tube air pressure and the compensation air pressure is the desired air pressure P of the PAM. d .

9. The method according to claim 1, characterized in that: The PID method is used for the underlying control of each PAM, which calculates the control voltage u of the solenoid valve opening according to the air pressure deviation, and then adjusts the actual air pressure P of the corresponding PAM to the expected air pressure P d , to output a driving force F, which then act together on the robot skeleton to produce actual angular motion θ.

10. An "artificial muscle" soft actuator bionic control device based on a pulse neural network, characterized in that: include: A model building module is used to determine the system model of the soft actuator robot based on the pneumatic characteristic equation of a single PAM "artificial muscle" soft actuator and the "artificial muscle" soft actuator system model; The first control instruction module is used to derive a feedforward controller based on the system model of the soft actuator robot; the input of the feedforward controller is the desired angular position θ d and the desired joint stiffness s d , the output is the PAM inner tube pressure P; The second control instruction module is used to design a cerebellum-like controller based on a pulse neural network based on the mammalian cerebellum neural circuit; the input of the cerebellum-like controller is the actual angular position θ and the desired angular position θ d , the output is the compensation air pressure P c ; The control module is used for controlling by combining the control instructions output by the feedforward controller and the cerebellum-like controller.

Images