Control method and device for a flexible actuator, computing device, cluster and robot

Abstract

The application discloses a control method and device of a flexible actuator, a computing device and cluster and a robot, and belongs to the technical field of flexible actuator control. The method comprises the following steps: converting a target motion instruction into a pulse sequence; inputting the pulse sequence into a pulse neural network to cooperatively generate a multi-channel pulse control signal; converting the multi-channel pulse control signal into a driving signal and applying the driving signal to a plurality of flexible driving units to make the flexible driving units produce cooperative deformation; collecting deformation state information and converting the deformation state information into a feedback pulse sequence, and inputting the feedback pulse sequence into the pulse neural network to close-loop adjust the multi-channel pulse control signal. The application takes pulses as a unified information format to run through the control of the whole link, realizes multi-unit dynamic cooperation and closed-loop accurate control, effectively improves the flexibility, stability and integration of the flexible actuating system, and can be widely applied to the fields of soft robots, medical rehabilitation and wearable devices.

Description

Technical Field

[0001] This application relates to the field of flexible actuator control technology, specifically to a control method and device for a flexible actuator, a computing device and cluster, and a robot. Background Technology

[0002] Flexible actuators are novel devices made of flexible materials that can undergo continuous deformation under external stimuli. They can be applied in fields such as soft robots, bionic muscles, medical assistive devices, and flexible grasping devices. Unlike traditional rigid joints, flexible actuators typically contain multiple independently actuated flexible drive units. These units achieve various complex movements such as bending, twisting, and extension through coordinated deformation. To precisely control the motion of the flexible actuator, multiple drive signals need to be generated based on the target motion command, and the deformation state needs to be sensed in real time for closed-loop adjustment.

[0003] However, existing control methods for flexible actuators often employ general-purpose microcontrollers (MCUs) or digital signal processors (DSPs) as the control core. Their architecture, based on continuous numerical processing, struggles to accurately handle the timing relationships between multi-channel signals. This results in a lack of unified coordination mechanisms among the deformation units during motion, making it difficult to precisely match the timing and amplitude of actions, thus affecting the overall motion accuracy and smoothness. To generate the required drive signals, existing solutions often use discrete components to build multi-stage amplifier circuits. As the number of drive units increases, the circuit size significantly increases, and consistency between channels becomes difficult to guarantee, limiting the system's integration and reliability. Existing feedback mechanisms typically only collect a single physical quantity (such as displacement at a point), while the deformation state of a flexible actuator manifests as distributed spatial morphological changes. A single physical quantity cannot fully reflect the actual deformation, preventing the control system from effectively correcting the drive signal based on the accurate deformation state. Existing controllers are usually separated from the flexible actuators, resulting in long signal transmission paths that are susceptible to interference. Furthermore, controllers are often rigid structures, making it difficult to maintain good mechanical compatibility during deformation. Summary of the Invention

[0004] This application provides a control method and apparatus for a flexible actuator, a computing device and cluster, and a robot, aiming to at least solve the prior art problems described in the background art.

[0005] In a first aspect, this application provides a control method for a flexible actuator, the flexible actuator comprising multiple independently drivable flexible drive units, comprising: converting a target motion command into a pulse sequence; inputting the pulse sequence into a spiking neural network, wherein the spiking neural network collaboratively generates a multi-channel pulse control signal; converting the multi-channel pulse control signal into a drive signal and applying it to the multiple flexible drive units to cause the multiple flexible drive units to undergo coordinated deformation; acquiring deformation state information of the flexible actuator and converting it into a feedback pulse sequence; inputting the feedback pulse sequence into the spiking neural network to adjust the multi-channel pulse control signal in a closed loop, thereby realizing the motion output indicated by the target motion command.

[0006] In one possible implementation, the spiking neural network collaboratively generates multi-channel pulse control signals, including: the spiking neural network uniformly determines the generation strategy of each channel pulse control signal according to the target motion command, so that the pulse control signals of each channel are correlated in terms of timing, frequency or pulse density, so as to jointly realize the motion output indicated by the target motion command.

[0007] In one possible implementation, the pulse control signals of each channel are correlated in terms of timing, frequency, or pulse density, including: the multi-channel pulse control signals have a preset cooperative relationship, the cooperative relationship including at least one of the following: there is a predetermined relative timing relationship between the pulse control signals of each channel; the pulse frequency ratio of the pulse control signals of each channel is a preset value; the pulse density distribution of the pulse control signals of each channel conforms to a preset cooperative mode.

[0008] In one possible implementation, the spiking neural network also dynamically adjusts the relative timing relationship based on real-time feedback signals to form dynamic antagonism among the plurality of flexible drive units, thereby regulating the compliance and stability of the cooperative deformation.

[0009] In one possible implementation, the spiking neural network includes a cooperative control layer comprising: multiple input neurons for receiving the pulse sequence; multiple interneurons forming excitatory and / or inhibitory connections with the multiple input neurons; and multiple output neurons connected to the multiple interneurons, wherein the pulse activity of the multiple output neurons corresponds to the multi-channel pulse control signal; wherein the interneurons are configured to cooperatively drive the multiple output neurons to generate pulses according to the pulse activity of the multiple input neurons, thereby forming a preset cooperative relationship between the multi-channel pulse control signals.

[0010] In one possible implementation, the spiking neural network collaboratively generates multi-channel pulse control signals, including: the spiking neural network dynamically adjusts the collaborative relationship between the multi-channel pulse control signals according to the feedback pulse sequence to adapt to the deformation state changes of the flexible actuator.

[0011] In one possible implementation, the multi-channel pulse control signal is amplified and / or its energy form is converted to generate a drive signal having sufficient voltage, field strength, or energy to drive the corresponding flexible drive unit to deform. The drive signal is at least one of an electric field signal, an optical signal, a thermal signal, or a magnetic field signal.

[0012] In one possible implementation, the conversion of the multi-channel pulse control signal into a drive signal is performed by a plurality of drive circuits corresponding one-to-one with each channel, the plurality of drive circuits being physically distributed on or near the flexible actuator body.

[0013] In one possible implementation, converting the multi-channel pulse control signal into a drive signal includes: maintaining the relative timing relationship between the multi-channel pulse control signals, and applying the drive signal maintaining the relative timing relationship to the plurality of flexible drive units so that the plurality of flexible drive units deform in a coordinated manner according to a predetermined timing.

[0014] In one possible implementation, during the process of converting the multi-channel pulse control signal into the drive signal, the output parameters of each drive signal corresponding to the channel are dynamically adjusted; wherein the output parameters include at least one of the following: amplitude, pulse width, or duration; the output parameters of the channel are associated with the deformation or deformation rate required by the corresponding flexible drive unit.

[0015] In one possible implementation, applying the driving signal to the plurality of flexible driving units includes: coupling the driving signal to the corresponding flexible driving unit in one of the following ways: electric field coupling, causing the flexible driving unit to undergo electro-induced deformation under electric field excitation; magnetic field coupling, causing the flexible driving unit to undergo magneto-induced deformation under magnetic field excitation; thermal field coupling, causing the flexible driving unit to undergo thermally induced deformation under thermal field excitation; and optical field coupling, causing the flexible driving unit to undergo photo-induced deformation under optical field excitation.

[0016] In one possible implementation, the plurality of flexible drive units include a first drive unit and a second drive unit; when the drive signal is applied to the corresponding flexible drive unit, the first drive unit generates a first deformation, and the second drive unit generates a second deformation; wherein, the first deformation includes axial extension and contraction, and the second deformation includes bending.

[0017] In one possible implementation, the axial stiffness of the first driving unit is higher than that of the second driving unit, such that under the same electric field excitation, the first type of driving unit produces a first deformation and the second type of driving unit produces a second deformation.

[0018] In one possible implementation, converting the target motion command into a pulse sequence includes: generating a corresponding pulse sequence by employing at least one of rate coding, time coding, or group coding based on at least one of the motion direction, motion amplitude, motion speed, or motion trajectory of the target motion command.

[0019] In one possible implementation, the deformation state information includes at least one of the following: local curvature information of the flexible actuator; surface strain distribution information of the flexible actuator; spatial pose information of the flexible actuator; deformation information of at least one of the plurality of flexible drive units; and contact force information between the flexible actuator and the external environment.

[0020] In one possible implementation, the closed-loop adjustment of the multi-channel pulse control signal includes: comparing the feedback pulse sequence with the pulse sequence corresponding to the target motion command, and adjusting the network parameters of the spiking neural network according to the comparison result to update the generation strategy of the multi-channel pulse control signal.

[0021] In one possible implementation, the network parameters include at least one of the following: synaptic weights between neurons; impulse firing thresholds of neurons; membrane potential time constants of neurons; and connection topology between neurons.

[0022] In one possible implementation, the closed-loop adjustment of the multi-channel pulse control signal includes: dynamically adjusting the gain, bias, or timing relationship of the multi-channel pulse control signal based on the error between the deformation state information and the target deformation state, so as to reduce the error.

[0023] Secondly, this application also provides a control device for a flexible actuator, the flexible actuator comprising multiple independently drivable flexible drive units, including: an instruction conversion module for converting a target motion instruction into a pulse sequence; a first processing module for inputting the pulse sequence into a spiking neural network, wherein the spiking neural network collaboratively generates a multi-channel pulse control signal; a second processing module for converting the multi-channel pulse control signal into a drive signal and applying it to the multiple flexible drive units to cause the multiple flexible drive units to undergo coordinated deformation; an acquisition module for acquiring deformation state information of the flexible actuator and converting it into a feedback pulse sequence; and a third processing module for inputting the feedback pulse sequence into the spiking neural network to adjust the multi-channel pulse control signal in a closed loop, thereby realizing the motion output indicated by the target motion instruction.

[0024] Thirdly, this application also provides a computing device, the computing device including a processor, the processor being configured to execute program code, causing the computing device to perform the control method for the flexible actuator provided in the first aspect.

[0025] Fourthly, this application also provides a computing device cluster including multiple computing devices, including a processor, the processor being configured to execute program code, causing the computing device cluster to perform the control method of the flexible actuator as provided in the first aspect.

[0026] Fifthly, this application also provides a computer program product containing instructions that, when run by a cluster of computing devices, cause the cluster of computing devices to perform a control method for a flexible actuator as provided in the first aspect.

[0027] In a sixth aspect, this application also provides a computer-readable storage medium including computer program instructions, which, when executed by a computing device, enable the computing device to perform a control method for a flexible actuator as provided in the first aspect.

[0028] In a seventh aspect, this application also provides a robot, including the computing device provided in the third aspect or the cluster of computing devices provided in the fourth aspect.

[0029] The above technical solution converts the target motion command into a pulse sequence and inputs it into a spiking neural network. The spiking neural network then collaboratively generates multi-channel pulse control signals based on this pulse sequence, ensuring that the signals in each channel are correlated in terms of timing, frequency, or pulse density. These multi-channel signals are then converted into driving signals and applied to multiple flexible drive units of the flexible actuator, causing each unit to produce coordinated deformation according to a preset coordination relationship. Simultaneously, the distributed deformation state information of the flexible actuator is collected and converted into a feedback pulse sequence, which is then re-input into the spiking neural network to adjust the multi-channel pulse control signals in a closed loop. Through the collaborative generation mechanism of the spiking neural network, the flexible drive units are precisely coordinated in terms of timing and amplitude, solving the problems of difficult handling of timing relationships between multi-channel signals and limited coordinated motion control effects. By utilizing the feedback pulse closed loop based on distributed deformation state information, the control system can correct control deviations in real time based on comprehensive deformation information, avoiding information loss and additional signal conversion losses caused by feedback from a single physical quantity, significantly improving motion accuracy and anti-interference capability. Furthermore, the entire control process is completed within the pulse domain, reducing delay and power consumption, thus enabling the flexible actuator to achieve smooth, stable, and adaptive coordinated motion output. Attached Figure Description

[0030] Figure 1 A schematic flowchart illustrating a control method for a flexible actuator provided in one embodiment of this application;

[0031] Figure 2 A schematic diagram of the structure of a control device for a flexible actuator provided in an embodiment of this application;

[0032] Figure 3 This is a schematic diagram of the structure of a computing device provided in one embodiment of this application;

[0033] Figure 4 This is a schematic diagram of the structure of a computing device cluster provided in one embodiment of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following description is intended to demonstrate how to implement multiple specific solutions of this application to fully support the scope of protection claimed in the claims, and does not constitute a limitation on the scope of protection.

[0035] This application provides a control method for a flexible actuator. The method first converts the target motion command into a pulse sequence and inputs it into a pulse neural network. This network then collaboratively generates multi-channel pulse control signals based on a unified control objective. Specifically, the signals in each channel are correlated in terms of timing, frequency, or pulse density, rather than being generated independently. Subsequently, these multi-channel signals are converted into driving signals and applied to multiple flexible drive units of the flexible actuator, causing each unit to undergo coordinated deformation according to a predetermined cooperation relationship. Simultaneously, the distributed deformation state information of the flexible actuator is collected and converted into a feedback pulse sequence, which is then re-input into the pulse neural network to form a closed-loop adjustment mechanism. This dynamically corrects the multi-channel pulse control signals, ultimately enabling the flexible actuator to accurately output the motion indicated by the target motion command.

[0036] The specific implementation steps of the control method for the flexible actuator provided in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0037] Figure 1 This is a flowchart illustrating a control method for a flexible actuator provided in one embodiment of this application.

[0038] Reference Figure 1 As shown, the control method may include the following steps:

[0039] S1: Convert the target motion command into a pulse sequence.

[0040] In some embodiments, after receiving a target motion command, the pulse coding module maps the command into a corresponding pulse sequence according to the motion intention information contained in the command and through a preset conversion rule. The pulse density, pulse interval, and total number of pulses of the pulse sequence jointly encode the speed, amplitude, and other characteristics of the target motion, which serve as the input stimulus for the subsequent spiking neural network processing unit.

[0041] In some embodiments, converting a target motion command into a pulse sequence includes: generating a corresponding pulse sequence by employing at least one of rate coding, time coding, or group coding based on at least one of the motion direction, motion amplitude, motion speed, or motion trajectory of the target motion command.

[0042] In a specific implementation, when the pulse coding module converts the target motion command into a pulse sequence, it uses different coding strategies to generate the corresponding pulse sequence according to the parameters such as motion direction, motion amplitude, motion speed or motion trajectory contained in the command. Specifically: For motion direction, a population coding method is used to map the direction angle to the firing ratio of a group of direction-selecting neurons. For example, 0° (directly forward) corresponds to 100% firing rate of the first group of neurons and 0% firing rate of the second group, and 90° (directly to the right) corresponds to 50% firing rate of the first group and 50% firing rate of the second group, so that the direction information is represented by the parallel pulse activity of multiple neurons. For motion amplitude (such as bending angle, extension length), a rate coding method is used to map the amplitude value to the pulse frequency linearly or nonlinearly. For example, 0° corresponds to 0Hz, 30° corresponds to 3kHz, and 60° corresponds to 6kHz. The larger the amplitude, the higher the pulse frequency. For motion speed, a time coding method is used to map the speed value to the total duration of the pulse train or the pulse interval. For example, at low speed, the pulse train lasts for 200ms and the pulse interval is sparse, while at high speed, the pulse train lasts for 50ms and the pulse interval is close. For motion trajectory (such as S-curve, uniform straight line), time coding and population coding are used in combination to discretize the trajectory into a series of instantaneous directions and amplitudes at a series of time points, which are expressed by the time pattern of the pulse sequence. This method of encoding multidimensional parameters separately enables the pulse sequence to carry complete motion commands with high information density and low redundancy, while maintaining natural compatibility with the pulse domain processing logic inside the spiking neural network processing unit, avoiding the delay and accuracy loss caused by digital-to-analog conversion.

[0043] For example, when the target motion instruction is "bend 30 degrees to the right, at a moderate speed, and move along an S-shaped trajectory", the pulse coding module first outputs the direction "right" through population coding, with the first group of neurons firing at 100% and the second group firing at 0%; it sets the pulse frequency to 3kHz for the amplitude "30 degrees" through rate coding; it sets the pulse duration to 100ms for the speed "moderate" through time coding; and it discretizes the "S-shaped trajectory" into 50 time segments through time coding, with each segment using population coding to represent the instantaneous direction (gradually changing from right to left and then back to right), ultimately generating a time-varying pulse sequence that is input into the spiking neural network processing unit, enabling it to accurately understand and execute complex continuous trajectory motion.

[0044] S2: Input the pulse sequence into the spiking neural network, which then generates multi-channel pulse control signals in a coordinated manner.

[0045] In some embodiments, after the pulse sequence is fed into the spiking neural network processing unit as input excitation, the integrated array of neuron circuits within the unit updates the values ​​in their respective membrane potential state registers through hardware accumulators based on the received pulse signals. When the membrane potential value of a neuron circuit exceeds the firing threshold stored in its threshold register, the neuron circuit immediately outputs a pulse and transmits the pulse signal to the input accumulator of the downstream neuron circuit through a hardware routing network. The connection relationships between the neuron circuits (determined by the weight values ​​in the synaptic weight memory) and the network topology (defined by the connection routing table) together constitute the hard-wired evolution rules of the network, enabling the processing unit to "convert" a single input pulse sequence into multiple output pulse sequences with specific spatiotemporal coordination relationships, i.e., multi-channel pulse control signals. The membrane potential update, threshold comparison, and pulse routing of the neuron circuit are all executed in parallel by dedicated hardware, which can complete the parallel state update of all neurons within the same clock cycle, ensuring that the multi-channel output signals maintain the preset timing coordination relationship with nanosecond-level accuracy. The cooperative characteristics of the output signals of each channel are directly determined by the fixed or configurable weights and connection structure of the hardware circuit, avoiding the delay and power consumption caused by real-time calculation. At the same time, the fully hardware-based pulse domain processing eliminates the uncertainty of software interruption and task scheduling, significantly improving the real-time performance and reliability of the control system.

[0046] For example, the spiking neural network processing unit integrates two sets of neuronal circuits, each containing 50 neurons. The two sets are configured with cross-inhibition connections through a hardware routing network. When the input pulse sequence arrives, the left neuronal circuit accumulates membrane potential and outputs a pulse first due to its connection weight advantage. This pulse reduces the membrane potential of the right neuronal circuit through inhibitory synapses. After several cycles, the two sets of neuronal circuits autonomously form a stable 180-degree phase difference output under hardware timing control, which drives the drive units on both sides of the flexible actuator to alternately contract and extend. The entire evolution process does not require any software intervention and is completed automatically by the hardware circuit.

[0047] In some embodiments, the spiking neural network collaboratively generates multi-channel pulse control signals, including: the spiking neural network uniformly determines the generation strategy of each channel pulse control signal according to the target motion command, so that the pulse control signals of each channel are correlated in terms of timing, frequency or pulse density, so as to jointly realize the motion output indicated by the target motion command.

[0048] In a specific implementation, after receiving the pulse sequence input from the pulse encoding module, the spiking neural network processing unit does not generate control signals independently for each channel. Instead, it uniformly determines the generation strategy for pulse control signals for all channels based on parameters such as motion direction, motion amplitude, motion speed, or motion trajectory contained in the target motion command. Specifically, the processing unit internally includes a collaborative decision-making layer. This layer first parses the command parameters represented by the input pulse sequence: for example, it calculates the motion speed through the frequency of the input pulses (higher frequency means faster speed), the motion direction through the group encoding ratio of the input pulses (e.g., the firing ratio of the left and right neuron groups determines the bending direction), and the motion amplitude through the cumulative number or pulse density of the input pulses (the total number of pulses is proportional to the bending angle). Subsequently, the collaborative decision-making layer simultaneously maps these parameters to each output channel, independently calculating the pulse start time, pulse frequency, number of pulses, and phase offset from other channels for each channel. This unified decision-making mechanism avoids timing conflicts or action incoordination that may result from independent decision-making by each channel, and ensures that multiple flexible drive units coordinate their actions according to the preset cooperation relationship from the beginning of the motion, thereby significantly improving the synchronization and compliance of the overall motion of the flexible actuator.

[0049] For example, when the target motion command is "bend 30 degrees to the right at a moderate speed," the collaborative decision layer resolves that: the direction is "right" (characterized by a firing rate of 100% for the left neuron group and 0% for the right), the amplitude of 30 degrees corresponds to a total of 300 pulses, and the moderate speed corresponds to a pulse frequency of 3kHz and a pulse train duration of 100ms. Based on this, the collaborative decision layer uniformly determines that: the extensor channel (responsible for contraction to generate bending) outputs 300 pulses at a frequency of 3kHz at the 0ms start time, and the flexor channel (responsible for providing damping) starts 5ms later, outputting 100 pulses at a frequency of 1kHz (generating reverse damping force). The pulse start phase difference between the left and right channels is set to 180°, so that the extensor muscles gradually contract and the flexor muscles gradually relax during the bending process, ultimately achieving a smooth, shock-free 30-degree bend, rather than initial jitter or terminal overshoot caused by independent responses of each channel.

[0050] In some embodiments, the pulse control signals of each channel are correlated with each other in terms of timing, frequency, or pulse density, including: the multi-channel pulse control signals have a preset cooperative relationship, the cooperative relationship including at least one of the following: there is a predetermined relative timing relationship between the pulse control signals of each channel; the pulse frequency ratio of the pulse control signals of each channel is a preset value; the pulse density distribution of the pulse control signals of each channel conforms to a preset cooperative mode.

[0051] In a specific implementation, based on the aforementioned unified decision-making, the pulse control signals of each channel can be configured to have a preset cooperative relationship, specifically including at least one of the following: a predetermined relative temporal relationship exists between the pulse control signals of each channel; the pulse frequency ratio of the pulse control signals of each channel is a preset value; or the pulse density distribution of the pulse control signals of each channel conforms to a preset cooperative mode. In terms of implementation, the spiking neural network processing unit presets the connection polarity (excitability or inhibition) and connection strength between output neurons through a hardware routing table or a configurable synaptic weight memory. Regarding the relative temporal relationship: by setting cross-inhibition connections, the pulses of the two sets of output neurons are staggered in time. For example, if the delay of channel A inhibiting channel B is set to 5ms, then channel B is inhibited within 5ms after the pulse of channel A is emitted, thus forming a fixed sequence. Regarding the pulse frequency ratio: by setting a fixed synaptic weight ratio, the pulse frequencies of the two channels are maintained at a constant ratio. For example, if the excitatory synaptic weight ratio is set to 3:1, then the frequency of the main channel is three times that of the auxiliary channel. Regarding pulse density distribution: By setting the spatial pattern of the feedforward connection, the pulse density of each channel conforms to a preset distribution. For example, if the synaptic weights of the eight-channel output neurons are set to a sinusoidal distribution, the output pulse density will exhibit a periodic spatial variation pattern. Under this preset cooperative relationship, the cooperative law is fixed by the hardware connection, eliminating the need for real-time calculation. This ensures the stability and repeatability of multi-channel signals during long-term operation, while significantly reducing the computational requirements of the processor.

[0052] For example, in a biomimetic peristaltic application controlling four-channel drive, the spiking neural network processing unit configures the connection between the four output channels to be cyclically inhibited through hardware configuration: channel 1 inhibits channel 2 by a 10ms delay, channel 2 inhibits channel 3 by a 10ms delay, channel 3 inhibits channel 4 by a 10ms delay, and channel 4 inhibits channel 1 by a 10ms delay, forming a clockwise inhibition loop; simultaneously, the synaptic weight ratio of each channel is set to 1:1:1:1, so that the four sets of output pulses automatically form a phase sequence with a sequential lag of 10ms. When the input pulse frequency is 100Hz, the four channel output pulses are emitted sequentially with phases of 0ms, 10ms, 20ms, and 30ms, respectively, driving the four drive units of the biomimetic muscle to contract sequentially, generating continuous peristaltic waves, and driving the flexible actuator to move forward. This synergistic relationship, solidified in the hardware, ensures that the four channels always maintain a fixed phase difference of 10ms regardless of changes in the input pulse frequency, eliminating the need for the controller to calculate the switching timing of each drive unit in real time.

[0053] In some embodiments, the spiking neural network also dynamically adjusts the relative timing relationship based on real-time feedback signals to form dynamic antagonism among multiple flexible drive units, thereby regulating the compliance and stability of cooperative deformation.

[0054] In specific implementations, to enhance the compliance and stability of coordinated deformation, the spiking neural network dynamically adjusts the relative timing relationship between the output pulses of each channel based on real-time feedback signals, forming dynamic antagonism among multiple flexible actuators. Specifically, the feedback fusion circuit within the processing unit receives the pulse sequence from the feedback interface. Through pulse timing-dependent plasticity (STDP) or a dedicated comparison circuit, it monitors the time difference between the current output pulse sequence and the feedback pulse sequence in real time, and dynamically fine-tunes the relative delay or synaptic weights between output neurons based on this time difference. For example, when the flexible actuator encounters external resistance during bending, the pulse frequency representing the actual deformation speed in the feedback pulse sequence decreases, deviating from the target pulse frequency. After detecting this deviation, the feedback fusion circuit automatically adjusts the relative timing between the extensor and flexor channels through the STDP mechanism: if the actual bending speed is slower than the target speed, the advance of the extensor channel pulse relative to the flexor channel is increased (e.g., from 5ms to 10ms), allowing the extensor to exert force earlier to overcome the resistance; if the actual bending speed is faster than the target speed, the advance is decreased, or even the flexor channel exerts force earlier to provide braking. This dynamic antagonistic mechanism enables flexible actuators to adjust their coordination strategy in real time according to actual load changes, material fatigue, or external disturbances, maintaining the smoothness and stability of motion and avoiding deformation overshoot or oscillation caused by fixed timing.

[0055] For example, in a scenario where a flexible gripper grasps objects of different weights, when the gripper grasps a lightweight egg, the feedback interface detects a low frequency of contact force feedback pulses (indicating a small contact force). The dynamic antagonistic mechanism adjusts the relative timing of the extensor channel (closed) and the flexor channel (opened) so that the extensor contracts slowly first, and the flexor begins to provide slight damping 5ms after the extensor contraction, allowing the gripper to gently touch the egg surface and then gradually tighten. When the gripper grasps a heavy water bottle, the feedback interface detects a sudden increase in the frequency of contact force feedback pulses (indicating an instantaneous increase in contact force). The dynamic antagonistic mechanism immediately adjusts the timing so that the extensor and flexor muscles exert force almost simultaneously, and the flexor provides strong damping only 1ms after the extensor contraction, allowing the gripper to quickly lock the object and prevent it from slipping. When the gripper is closed without any object, the feedback interface detects no contact force feedback. The dynamic antagonistic mechanism maintains a moderate timing difference between the extensor and flexor muscles, allowing the gripper to quickly close to its minimum position and then automatically stop, avoiding collision damage. The entire process is automatically timed by a spiking neural network based on changes in the frequency of feedback pulses, without the need for real-time intervention from an external controller or mode switching.

[0056] In some embodiments, the spiking neural network includes a cooperative control layer, which includes: multiple input neurons for receiving pulse sequences; multiple interneurons forming excitatory and / or inhibitory connections with the multiple input neurons; and multiple output neurons forming connections with the multiple interneurons, wherein the pulse activity of the multiple output neurons corresponds to a multi-channel pulse control signal; wherein the interneurons are configured to cooperatively drive the multiple output neurons to generate pulses according to the pulse activity of the multiple input neurons, so that a preset cooperative relationship is formed between the multi-channel pulse control signals.

[0057] In a specific implementation, to achieve the aforementioned collaborative generation function, the spiking neural network is internally configured with a collaborative control layer. This layer consists of multiple input neurons, intermediate neurons, and output neurons connected according to a specific topological structure. Specifically, the input neurons employ an integrator-fire circuit. Each input neuron receives a pulse sequence output by the pulse coding module and transmits it in parallel to multiple intermediate neurons. The intermediate neurons are divided into several functional groups. Within each group, excitatory synaptic connections form positive feedback loops to maintain oscillations (e.g., enhancing input through output feedback to ensure continuous firing). Intergroups are mutually constrained through inhibitory synaptic connections to generate phase differences (e.g., when neurons in group A fire, the membrane potential of neurons in group B is reduced through inhibitory synapses). Intermediate neurons and output neurons are connected through configurable synaptic weights. Each output neuron can receive input from multiple intermediate neurons and determines its firing mode based on a weighted sum. The intermediate neurons are configured to collaboratively drive multiple output neurons to generate pulses based on the pulse activity of the input neurons, thus establishing a preset collaborative relationship between the multi-channel pulse control signals. This structural design embeds the complexity of collaborative control into the hardware connection, enabling parallelization of control logic, low-latency processing, and flexible expansion of multi-channel output.

[0058] For example, in a bionic muscle application controlling an eight-channel drive, the collaborative control layer is configured as follows: two input neurons receive a "start" pulse and a "velocity regulation" pulse, respectively; four interneurons form a first group of oscillators, which generate synchronous oscillations of 50Hz through mutual excitation; another four interneurons form a second group of oscillators, which form a 90° phase difference with the first group through cross-inhibition; each interneuron then drives an output neuron through a one-to-one excitatory connection, while each output neuron also receives weak inhibitory connections from two adjacent interneurons. Specifically, the interneurons A1, A2, A3, and A4 in the first group maintain synchronous oscillations through excitatory connections (weight 0.8); the interneurons B1, B2, B3, and B4 in the second group form a phase lag with the first group through cross-inhibition connections (weight -0.5); output neuron O1 receives excitation from A1 (weight 1.0) and inhibition from B2 (weight -0.2), O2 receives excitation from A2 and inhibition from B3, and so on. When the input neurons receive the "start" pulse, the first group of interneurons fire simultaneously at 50Hz oscillations, driving the output neurons O1-O4 to fire synchronously through excitatory connections. After a phase lag, the second group of interneurons fires 10ms later, driving the output neurons O5-O8 to fire synchronously. Simultaneously, weak inhibition of adjacent channels causes the pulses of each output neuron to attenuate slightly at the edges. Ultimately, the eight output pulses form a complex coordinated pattern: the first group of four channels is synchronous, the second group of four channels is synchronous, the two groups are 10ms apart, and the pulse intensity of the edge channels within each group is slightly weaker. This drives the bionic muscle to produce a peristaltic wave-like movement that propagates from one end to the other. All coordinated relationships are automatically achieved by hardware connections, without the need for software intervention or real-time calculations.

[0059] In some embodiments, the collaborative control layer consists of a three-level circuitry: an input neuron array, an intermediate neuron grid, and output neurons. All neurons are implemented using integrator-discharge hardware circuitry. Each neuron in the input neuron array receives an external pulse sequence, converts it into an integral membrane potential signal, and broadcasts it to the intermediate neuron grid via a parallel bus. The intermediate neuron grid configures synaptic weights according to a preset excitation / inhibition connection routing table: excitatory connections transmit membrane potential increments through positive weights, while inhibitory connections reduce the membrane potential of the target neuron through negative weights. Neurons within the grid form mutually stimulating or mutually inhibiting topologies, such as symmetrical cross-inhibition loops or cascaded delay chains. The output neurons receive weighted pulse inputs from specific nodes in the intermediate neuron grid and generate the final multi-channel pulse control signal. The entire routing configuration can be fixed in hardware (e.g., read-only memory) or programmed online via registers, thereby enabling flexible switching between different collaborative modes (e.g., fixed phase difference, frequency ratio, or timing misalignment).

[0060] In some embodiments, the spiking neural network collaboratively generates multi-channel pulse control signals, including: the spiking neural network dynamically adjusts the collaborative relationship between the multi-channel pulse control signals according to the feedback pulse sequence to adapt to the deformation state changes of the flexible actuator.

[0061] In a specific implementation, to further enhance the system's adaptability, the spiking neural network dynamically adjusts the collaborative relationship between multi-channel pulse control signals based on the feedback pulse sequence to adapt to changes in the deformation state of the flexible actuator during operation. Specifically, a feedback adjustment loop is incorporated within the processing unit: after the feedback pulse sequence is input, it is compared with the current output pulse sequence in the time domain. A dedicated adaptive circuit (such as a synaptic weight adjuster or timing deviation corrector) calculates the deviation, and the synaptic weights, connection polarities, or timing offset parameters in the collaborative control layer are updated online based on the deviation. For example, when the feedback pulse indicates that the actual output frequency of a certain channel is lower than expected, the adaptive circuit increases the synaptic gain between the input neuron and the intermediate neuron of that channel (e.g., from 0.8 to 1.2), or reduces the inhibitory connection weight of the output neuron of that channel, restoring the output pulse frequency to the target value. This feedback adaptive mechanism enables the controller to automatically compensate for performance drift caused by factors such as material aging, temperature changes, and load fluctuations in the flexible actuator, without manual intervention or recalibration, significantly improving the system's long-term stability and robustness.

[0062] For example, in a long-term bionic arm application, the dielectric elastomer material of the bionic muscle fatigues after 10,000 bending cycles, with the deformation decreasing from an initial 30 degrees to 24 degrees under the same driving voltage. The feedback interface detects the actual bending angle in real time via a flexible strain sensor attached to the muscle surface and converts it into a feedback pulse sequence (10 pulses per degree of deformation). The adaptive circuit within the processing unit detects that the target pulse count should be 300 (corresponding to 30 degrees), but the actual feedback pulse count is only 240 (corresponding to 24 degrees), a deviation of 60 pulses. The adaptive circuit first calculates the deviation rate (20%), and then, through a synaptic weight regulator, gradually increases the synaptic weight of the extensor output neurons from the initial value of 1.0 to 1.25 (an increase of 25%), while simultaneously increasing the synaptic weight of the flexor output neurons from the initial value of 0.5 to 0.6 (an increase of 20%), thereby increasing the output pulse frequency of the extensor channel by 25% and the output pulse frequency of the flexor channel by 20%. After three consecutive cycles of adaptive adjustment, the feedback pulses recovered to 300, and the actual bending angle returned to 30 degrees. When the ambient temperature rose from 20℃ to 40℃, the response speed of the dielectric elastomer increased, and the feedback pulses arrived earlier. The adaptive circuit detected the timing delay and automatically adjusted the relative timing of the extensor and flexor muscles from the initial 5ms lead to 3ms, thus stabilizing the bending speed. Throughout the entire 10,000 cycles, the spiking neural network continuously monitored the frequency, timing, and amplitude deviations between the feedback pulses and the target pulses, adjusting the synaptic weights and timing relationships of each channel in real time. This ensured that the bionic arm maintained the same grasping force, bending angle, and movement speed as in the initial state without any manual calibration or software parameter adjustments.

[0063] S3: Convert the multi-channel pulse control signal into a drive signal and apply it to multiple flexible drive units so that the multiple flexible drive units can generate coordinated deformation.

[0064] In some embodiments, the pulse control signal of each channel is amplified or converted into a signal form by the driving circuit to generate a driving signal with sufficient energy (such as voltage, current, field strength, or heat), and then applied to the corresponding flexible driving unit through independent conductive paths. Since the pulse control signals of each channel are pre-correlated in terms of timing, frequency, or pulse density by the pulse neural network, the converted driving signal also retains these cooperative characteristics. For example, parameters such as the start time, duration, and intensity change rate of the driving signal of each channel are output strictly according to the preset coordination relationship. When these driving signals with cooperative characteristics act on multiple flexible driving units simultaneously, each unit generates corresponding contraction, extension, torsion, or bending deformation according to its own driving signal, and the deformations cooperate with each other in time and space to synthesize the expected overall motion. By directly converting the coordinated control signal in the pulse domain into driving energy and applying it in parallel to each flexible drive unit, the timing deviation and information loss caused by converting it into a continuous signal and then driving it separately in the traditional scheme are avoided. This ensures the synchronization and coordination of the deformation of multiple flexible drive units with millisecond-level accuracy. At the same time, the coordination of the deformation of each flexible drive unit is determined by the inherent characteristics of the driving signal, without the need for additional real-time synchronization control. This reduces the computational burden and response delay of the control system while achieving smooth and compliant overall motion.

[0065] In some embodiments, the multi-channel pulse control signal is amplified and / or its energy form is converted to generate a drive signal having a voltage, field strength, or energy sufficient to drive the corresponding flexible drive unit to produce deformation. The drive signal is at least one of an electric field signal, an optical signal, a thermal signal, or a magnetic field signal.

[0066] In a specific implementation, the pulse control signal of each channel is first amplified by a pre-drive stage to enhance its driving capability; then, different processing paths are selected according to the type of flexible drive unit. It should be noted that the types of flexible drive units include electro-actuated, thermal-actuated, optical-actuated, and magnetic-actuated types, and the corresponding drive signals must be adapted to their driving principles.

[0067] For electrically actuated units (such as dielectric elastomers and piezoelectric ceramics), a charge pump, transformer, or switched capacitor boost circuit is used to amplify the voltage by tens to hundreds of times to generate a high-voltage electric field signal.

[0068] For photo-actuated units (such as photoresponsive hydrogels), electrical pulse signals are converted into light pulses of corresponding wavelengths through light-emitting diodes or laser diodes, and optical power is amplified by modulating the drive current during the conversion process.

[0069] For thermally actuated units (such as shape memory alloys and thermally expanding polymers), amplified pulse currents drive micro heating elements (such as metal thin film resistors and Joule heating wires) to convert electrical energy into heat energy, raising the temperature to the phase transition threshold.

[0070] For magneto-actuated units (such as magnetostrictive materials), the pulsed current is amplified and converted into a pulsed magnetic field through coils or electromagnets.

[0071] The above implementation ensures that the drive signal has sufficient energy to drive the load through power amplification, avoiding deformation failure due to insufficient drive capability; through energy form conversion, the same control method can be compatible with flexible drive units based on various physical principles, greatly expanding the versatility of the controller; at the same time, the configurability of voltage, field strength or energy allows the controller to flexibly adjust the output parameters according to the threshold characteristics of different drive units, achieving precise matching between drive energy and load requirements.

[0072] For example, when the flexible driving unit is a dielectric elastomer biomimetic muscle, the 3.3V pulse control signal is amplified (current increased to 100mA), and then the voltage is boosted to 800V through a cascaded charge pump to generate a high-voltage electric field signal applied to the dielectric elastomer electrode, causing the dielectric elastomer to contract rapidly. When the driving unit is a shape memory alloy wire, the same pulse control signal is amplified (pulse current increased from 10mA to 2A), and electrical energy is converted into Joule heat through a micro nickel-chromium alloy heating wire, causing the alloy wire temperature to rise from room temperature to 70℃ phase transition temperature within 50ms, resulting in contraction deformation. When the driving unit is a photoresponsive hydrogel, the pulse control signal is amplified to drive a high-power blue LED, outputting a light pulse with a wavelength of 450nm and a light intensity of 50mW / cm², causing local volume contraction of the hydrogel and achieving bending motion.

[0073] In some embodiments, the multi-channel pulse control signal is converted into a drive signal and executed by a plurality of drive circuits corresponding one-to-one with each channel. The plurality of drive circuits are physically distributed on or near the flexible actuator body.

[0074] In specific implementations, each drive circuit, in the form of a bare chip or packaged chip, is directly fixed to the surface, inner wall, or embedded in the elastic material of the flexible actuator body through surface mounting or embedded processes. Each drive circuit is responsible for driving only one or a few flexible drive units closest to it. The drive circuits are connected to a common pulse neural network processing unit through flexible wires or conductive ink lines. This distributed layout significantly shortens the transmission path of the high-voltage drive signal, reduces electromagnetic interference, signal attenuation, and timing distortion caused by long-line transmission, and ensures that the cooperative relationship in the multi-channel pulse control signal can be transmitted to each drive unit with high fidelity. At the same time, the distributed circuit and the flexible actuator body form a mechanically compatible integrated structure, avoiding the traction constraint of the centralized drive circuit on the movement of the flexible body, and improving the integration and reliability of the system.

[0075] It should also be noted that this application adopts a one-to-one localized layout for the distributed driving circuits of photoactuated and thermally actuated units. For photoactuated units (e.g., photoresponsive hydrogels, photoflexible polymers), the corresponding driving circuit is a micro LED driver chip. This chip is integrated with the light-emitting diode (or laser diode) in the same flexible package, and is attached to the surface of the photoactuated unit or embedded in its substrate via a flexible circuit board or conductive ink lines. The pulse control signal output by the pulse neural network directly drives the chip, and the light intensity and exposure time of the LED are controlled by adjusting the pulse width or amplitude, so that the light pulse irradiates the driving unit with minimal transmission loss. For thermally actuated units (e.g., shape memory alloy wires, thermally expanding polymers), the corresponding driving circuit is a micro heating driver chip (including a power MOSFET and a current-limiting resistor), which can be integrated on a flexible circuit board and attached near the shape memory alloy wire, or directly bonded to the alloy wire with thermally conductive adhesive. The pulse control signal is amplified into a pulse current by the chip, which flows through the alloy wire to generate Joule heat. Because the driving circuit is in close contact with the alloy wire, the heat loss is minimal and the response is rapid. The advantages of this distributed layout are: it avoids light attenuation and line loss caused by long-distance transmission of optical signals or high current, improving energy conversion efficiency and response speed; at the same time, each driving unit is independently configured with a driving circuit, which facilitates local closed-loop temperature or light intensity control, improving the accuracy and consistency of multi-unit coordinated motion. For example, in a biomimetic soft crawling robot, multiple photo-bending units are arranged along both sides of the body, and each unit is attached with a micro-module integrating an LED and a driving chip. The pulse neural network independently controls the light emission sequence of each module, causing the units on both sides to bend alternately to generate peristaltic waves; in a shape memory alloy driven biomimetic hand, a micro-driving chip is welded near each alloy wire. The chip outputs a pulse current according to the control signal, causing the alloy wire to heat and contract rapidly, realizing the independent bending of the finger.

[0076] In some embodiments, for electrically actuated flexible drive units (such as dielectric elastomers) that require high-voltage electric field driving, the aforementioned distributed drive circuit may specifically include a boost circuit for amplifying the low-voltage pulse control signal output by the pulse neural network processing unit to generate a high-voltage electric field signal sufficient to drive the corresponding drive unit to deform.

[0077] Because the bionic muscles of different parts of a bionic robot (such as fingers, wrists, and upper arms) have significantly different requirements for driving voltage, output power, size, and compliance, a single boost circuit cannot meet all these needs. Therefore, this application provides three different sizes of boost circuits based on different application scenarios: micro-embedded, small flexible integrated, and compact modular. These are described below.

[0078] A miniature embedded boost circuit is configured to be fully embedded within the flexible inner wall of the distal end (such as fingers or toes) of a bionic robot or within a miniature bionic muscle body. Its overall contour matches the tiny volume of the implanted cavity, and its thickness does not exceed the elastic deformation limit of the flexible epidermis of that area. For example, its dimensions can be specifically configured as <5mm × 5mm × 1mm. In one embodiment, this miniature embedded boost circuit employs a fully integrated charge pump topology, utilizing an on-chip MOSFET switch array and flying capacitors to achieve multi-stage voltage multiplication without the need for external inductors. Its power amplification is achieved as follows: a low-voltage pulse signal (e.g., 3.3V / 10mA) output from a pulse neural network processing unit is shaped and buffered before driving a non-overlapping clock generator inside the charge pump. Through three to six cascaded cross-coupled charge pump units, each stage boosts the voltage by approximately 1.5 to 2 times, ultimately outputting a drive signal in the 200-400V / mA range (power approximately 0.2-2W).

[0079] In some applications, this miniature embedded boost circuit can be directly embedded in the silicone inner wall of the bionic finger or integrated into the surface of the micro-bionic muscle within the finger. Its extreme miniaturization allows it to be implanted in the most distal bionic part. The all-solid-state, inductive-free structure is resistant to bending and impact, and exhibits high reliability during repeated finger flexion and extension. Simultaneously, the adjustable switching frequency allows for precise control of the contraction force of the fingertip muscles. For example, in this bionic dexterous hand, each fingertip embeds a 3mm × 3mm charge pump chip, boosting the 3.7V battery voltage to 300V to drive a 2mm diameter dielectric elastomer muscle unit at the fingertip, enabling the bionic finger to perform delicate actions such as gently pinching objects like strawberries.

[0080] A small, flexible, integrated boost circuit is configured to fit within a flexible cavity or muscle bundle gap in a medium-sized part of a bionic robot (such as the palm or wrist). Its length and width are adapted to the usable area on the dorsal side of the joint, and it possesses overall flexibility, without substantially restricting the normal flexion and extension freedom of the joint. For example, its dimensions are specifically configured as 10mm × 10mm × 2mm to 20mm × 15mm × 3mm. In one embodiment, the small, flexible, integrated boost circuit employs a hybrid integration process, mounting the boost controller, a miniature planar transformer or coupled inductor, capacitor, and rectifier diodes onto a flexible printed circuit board (FPC). Its power amplification is achieved as follows: a low-voltage pulse signal drives the controller to generate a high-frequency PWM wave, which is boosted by a miniature transformer (turns ratio 1:20 to 1:50), and then smoothed into high-voltage DC or a high-voltage pulse sequence by a voltage doubler rectifier network, resulting in an output voltage of 400–800V and an output power of 1–5W.

[0081] In some applications, this small, flexible, integrated boost circuit is bendable and can be encased in a flexible cavity on the back of a bionic hand or wrist, or encapsulated in a silicone jacket over a bionic muscle bundle. Compared to a charge pump, the transformer topology offers a higher boost ratio and power density, with lower output voltage ripple, making it suitable for driving medium loads requiring a stable electric field. The flexible substrate allows the circuit to adapt to dynamic deformation during joint bending, while reducing the number of cables. For example, in the flexion and extension movements of a bionic wrist, a 15mm × 15mm flexible boost circuit is embedded under the skin on the back of the wrist, providing 600V / 2W drive to three sets of parallel bionic muscles, enabling smooth up-and-down wrist movements.

[0082] A compact modular boost circuit is configured to be embedded within the internal cavity of a rigid or semi-rigid shell of a large proximal part of a bionic robot (such as the upper arm or thigh). Its compact shape avoids space constraints related to the arrangement of the main drive muscle bundles. It connects to the distal load via a flexible cable, without affecting the overall mobility and appearance of the robot. For example, its dimensions can be specifically configured as 30mm × 25mm × 5mm to 50mm × 40mm × 10mm. In one embodiment, the compact modular boost circuit employs a multi-phase interleaved Boost or flyback topology, using a small planar transformer, solid-state capacitors, and shielded inductors, all encapsulated in a metal / plastic shielded shell. It has an output power of 10–50W and an output voltage exceeding 1000V. Its power amplification is achieved by a low-voltage pulse signal generating a multi-phase interleaved control sequence via a driver chip, which drives multiple power MOSFETs. Energy is collected and output through a planar transformer and a voltage doubler rectifier network. The interleaving technique reduces input current ripple and output capacitor stress, resulting in a conversion efficiency typically >90%.

[0083] In some applications, this compact modular boost circuit can be embedded in the internal cavity of a bionic upper arm or thigh, connecting to multiple distributed bionic muscle groups via flexible cables. A silicone buffer layer can be added to the bottom of the module to reduce vibration transmission. High power output meets the explosive power requirements of large muscle groups such as the upper arm and thigh, such as rapid arm raising or leg kicking movements. The modular design facilitates maintenance and replacement, and the shielded outer shell effectively suppresses high-voltage switching noise. For example, in a bionic upper limb robot, a 40mm × 30mm × 8mm modular boost circuit is embedded inside the upper arm shell, boosting the 12V battery voltage to 1200V, simultaneously driving six bionic muscle bundles, enabling the robot to lift a 5kg weight.

[0084] In some embodiments, during the process of converting multi-channel pulse control signals into drive signals, the output parameters of each drive signal corresponding to a channel are dynamically adjusted; wherein the output parameters include at least one of the following: amplitude, pulse width, or duration; the output parameters of the channel are associated with the deformation or deformation rate required by the corresponding flexible drive unit.

[0085] In a specific implementation, each drive channel integrates a programmable drive circuit, which includes: a digital control interface (such as an SPI or I2C slave), a configuration register set, a programmable gain amplifier (PGA), a pulse width modulator (PWM), and a high-precision timer / counter. The spiking neural network processing unit calculates the required deformation and deformation rate of each flexible drive unit in the current motion cycle based on the target motion command and real-time feedback signal through a cooperative control algorithm. These requirements are then mapped to configuration values ​​for the output parameters: the deformation determines the total energy of the drive signal, achieved by adjusting the amplitude or pulse train duration; the deformation rate determines the instantaneous power, achieved by adjusting the pulse width or amplitude. The mapping coefficients can be pre-calibrated and stored in the configuration register. Subsequently, the pulse neural network processing unit writes parameter values ​​to the configuration registers of each channel via the control bus: for example, writing an 8-bit binary number to the PGA gain control word to make the output voltage amplitude adjustable in approximately 4V steps within the range of 0~1000V; writing a count value to the PWM comparison register to make the output pulse width adjustable in 1μs resolution within the range of 10μs~10ms; and writing a count value to the timer reload value register to make the duration of the drive signal output programmable within the range of 100μs~1s. Each channel performs adjustment independently without interference. Through closed-loop dynamic adjustment, the drive energy received by each flexible drive unit is precisely matched to its instantaneous deformation requirements, avoiding deformation saturation, overshoot, or energy waste in traditional constant voltage drives, while allowing the same controller to drive hybrid flexible drive unit arrays of different sizes and materials.

[0086] For example, in a scenario where a bionic hand grasps an irregular object, the spiking neural network analyzes the contact force distribution in real time: the thumb requires large deformation (60° bend) but a low rate (30° / s), so the amplitude of the thumb channel is set to 800V, pulse width to 5ms, and duration to 200ms; the index finger requires moderate deformation (30° bend) and a high rate (60° / s), so the amplitude of the index finger channel is set to 500V, pulse width to 2ms, and duration to 50ms; the middle finger requires almost no deformation but needs a fast response, so the pulse width is set to 0.5ms and the amplitude to 200V. After each channel is adjusted independently, the bionic hand wraps around the object with a stable and coordinated grasping posture, and the grasping success rate and energy efficiency are better than the control group with uniform output parameters.

[0087] Additionally, it's important to note that for both optical and thermal driving methods, the "pulse width" and "duration" in the dynamically adjusted output parameters have specific physical meanings. In optical driving, the "pulse width" corresponds to the duration of a single light pulse (i.e., the conduction time of an LED or laser diode). A wider pulse width results in higher energy per irradiation, potentially causing greater local deformation in photoresponsive materials (such as photosensitive hydrogels). The "duration" corresponds to the total length of the light pulse train (i.e., the total irradiation time of multiple consecutive light pulses). A longer duration results in higher accumulated light energy, allowing the optical driving unit to continuously bend or contract. In thermal driving, the "pulse width" corresponds to the width of a single heating current pulse (i.e., the conduction time of a power MOSFET). A wider pulse width results in more energy injected by a single Joule heating pulse, leading to a higher instantaneous temperature rise in shape memory alloys or thermally expanding polymers. The "duration" corresponds to the total duration of the heating current pulse train (i.e., the total heating time of multiple pulses). A longer duration results in more accumulated heat, causing the driving unit's temperature to gradually rise to the phase transition point and maintain deformation. By adjusting the pulse width and duration separately, instantaneous power and total energy can be independently controlled, achieving precise decoupling of the deformation speed and amplitude of the optical / thermal actuation unit. For example, in an optically driven micro-gripper, when rapid bending with a small amplitude is required, a narrow pulse width (1ms) but high-frequency optical pulse train with a short duration (10ms) is used; when large-amplitude, slow bending is required, a wide pulse width (5ms) but low-frequency pulse train with a long duration (50ms) is used. In a thermally driven bionic finger, when rapid grasping is required, a large pulse width (20ms) high-current pulse is used to rapidly heat and contract the alloy wire with a short duration; when maintaining the grasping force is required, a narrow pulse width (5ms) low-current pulse train with a long duration (500ms) is used to maintain the temperature. This parameter decoupling allows this control method to flexibly adapt to the different dynamic characteristics of the optical / thermal actuation unit.

[0088] In some embodiments, the plurality of flexible drive units include a first drive unit and a second drive unit; when a drive signal is applied to the corresponding flexible drive unit, the first drive unit generates a first deformation and the second drive unit generates a second deformation; wherein the first deformation includes axial extension and contraction, and the second deformation includes bending.

[0089] In a specific implementation, the spiking neural network outputs pulse control signals with specific timing and amplitude relationships to the first and second channels respectively, based on the target motion command (such as "elbow bending"). The pulse control signal from the first channel, after power amplification, drives the first drive unit (using a dielectric elastomer column with high axial stiffness) to produce rapid axial contraction, simulating the force exertion of the agonist muscle. The pulse control signal from the second channel, after delay and amplitude adjustment, drives the second drive unit (using a dielectric elastomer film with lower axial stiffness) to produce bending deformation, simulating the damping and guiding effect of the antagonist muscle. The two signals are precisely coordinated in timing (e.g., the first drive unit contracts first, and the second drive unit bends 10ms later), resulting in a smooth bending motion of the overall actuator. By combining drive units with different mechanical properties and utilizing the spiking neural network's ability to coordinate multi-channel signals, a agonist-antagonist muscle collaborative working mechanism similar to that of biological muscles is achieved. This allows the flexible actuator to autonomously generate complex bending movements under a single control command, avoiding the redundant structure of multiple independent actuators required in traditional solutions, thus improving motion efficiency and compliance.

[0090] For example, in the bionic elbow joint flexion control, the spiking neural network outputs a pulse control signal sequence with a frequency of 5kHz and a duration of 100ms to the first channel according to the "flex 60°" command, and outputs a pulse sequence with the same frequency but starting 10ms later to the second channel. The two pulse signals enter their respective local high-voltage drive circuits, are boosted to 800V and 500V by charge pumps, respectively, to generate high-voltage electric field drive signals, which are applied to the high-stiffness dielectric elastomer pillar on the outer side of the elbow joint and the low-stiffness dielectric elastomer film on the inner side. Under the action of the drive signal, the outer pillar contracts axially (from 50mm to 35mm), and the inner film undergoes bending deformation (the radius of curvature changes from ∞ to 25mm). Together, they smoothly flex the elbow joint to 60°.

[0091] In the coordinated control of the first and second drive units, the spiking neural network outputs high-frequency pulses to the first channel (generating axial contraction) and low-frequency delayed pulses to the second channel (generating bending) according to the target bending direction. The bending angle can be changed by adjusting the amplitude ratio of the two pulses: for example, when the target bending angle increases, the amplitude of the first channel pulse is increased (increasing the axial contraction force) and the start time of the second channel pulse is appropriately delayed, resulting in a larger bending amplitude; when the bending angle needs to be reduced, the amplitude of the first channel is decreased and the delay of the second channel is reduced. The precise timing coordination between the two (e.g., the first drive unit contracts first, and the second drive unit bends 10ms later) enables the overall actuator to produce smooth bending motion. Its beneficial effects are: by combining drive units with different mechanical properties and utilizing the coordinated control capability of the spiking neural network for multi-channel signals, a synergistic working mechanism similar to that of biological muscles (agonist-antagonist muscles) is achieved. This allows the flexible actuator to autonomously generate complex bending motions under a single control command, avoiding the redundant structure of multiple independent actuators stacked in traditional solutions, and improving motion efficiency and compliance. For example, in bionic elbow joint flexion control, the spiking neural network, based on the "flex 60°" command, outputs a pulse control signal sequence with a frequency of 5kHz, an amplitude of 800V, and a duration of 100ms to the first channel, and a pulse sequence with a frequency of 3kHz, an amplitude of 500V, and a lag of 10ms to the second channel. The two pulse signals enter their respective local high-voltage drive circuits, are boosted by a charge pump to generate a high-voltage electric field drive signal, which is applied to the high-stiffness dielectric elastomer pillar on the outer side of the elbow joint and the low-stiffness dielectric elastomer film on the inner side. Under the action of the drive signal, the outer pillar contracts axially (from 50mm to 35mm), and the inner film undergoes bending deformation (the radius of curvature changes from ∞ to 25mm). Together, they smoothly flex the elbow joint to 60°, achieving the complete control process.

[0092] In other embodiments, the multiple flexible actuation units include a first actuation unit, a second actuation unit, and a third actuation unit. When the multi-channel pulse control signal generated collaboratively by the spiking neural network is converted into a driving signal by the driving circuit and applied to the corresponding unit, the first actuation unit generates axial extension and contraction deformation, the second actuation unit generates bending deformation, and the third actuation unit generates torsional deformation. Specifically, the spiking neural network outputs pulse control signals with specific timing, amplitude, and phase relationships to the three channels according to the target motion command (such as "compound motion of a bionic wrist joint"): the pulse control signal of the first channel, after power amplification, drives the high-stiffness dielectric elastomer column to generate rapid axial extension and contraction, providing the main displacement; the pulse control signal of the second channel, after appropriate delay and amplitude adjustment, drives the low-stiffness dielectric elastomer film to generate bending deformation, changing the direction of motion; the pulse control signal of the third channel, after phase modulation, drives the helically wound dielectric elastomer fiber to generate torsional deformation, realizing rotational freedom. The three components work in precise timing coordination through a pre-defined cooperative relationship within the spiking neural network (e.g., the first unit contracts axially first, the second unit bends 5ms later, and the third unit twists 3ms later), enabling the flexible actuator to perform complex spatial movements. By introducing a third drive unit to increase the torsional degree of freedom, the flexible actuator can simulate the multi-axis movements of biological joints (such as flexion, extension, ulnar and radial deviation, and rotation of the wrist joint). Utilizing the multi-channel cooperative generation capability of the spiking neural network, axial, bending, and torsional deformations are fused into continuous, smooth composite movements, significantly improving the dexterity and task adaptability of the flexible robot.

[0093] For example, in bionic wrist joint control, the spiking neural network, based on the instruction of "palm-up rotation while bending 30°", outputs a pulse sequence with a frequency of 4kHz and a duration of 80ms to the first channel (driving axial contraction), a pulse sequence with the same frequency but starting 5ms later to the second channel (driving bending), and a pulse sequence with a frequency of 3kHz and a phase lag of 8ms to the third channel (driving torsion). These three pulse signals are boosted to 600V, 500V, and 700V respectively by a local high-voltage drive circuit, generating high-voltage electric field drive signals, which are applied to the axial extension unit on the outer side of the wrist joint, the bending unit on the inner side, and the spirally wound torsion unit. The contraction of the axial unit flexes the wrist, the bending unit changes the direction of flexion, and the torsion unit rotates the palm. These three elements work together to achieve a complex wrist movement, completely simulating the combined movements of dorsiflexion, palmar flexion, pronation, and supination of the human wrist joint.

[0094] It should be noted that the first, second, and third driving units mentioned above are merely examples. Multiple flexible driving units can also include four, five, or more driving units, each endowed with different mechanical response characteristics and deformation modes (such as axial extension, bending, torsion, shearing, and expansion). By leveraging the multi-channel collaborative generation capability of the spiking neural network, the timing, frequency, amplitude, and other parameters of the pulse control signal can be independently configured for each driving unit. This allows the units to be driven to produce combined deformations according to a preset coordination relationship, thereby achieving more complex and dexterous motion modes. For example, in the control of a biomimetic octopus tentacle, eight driving units can be arranged along the tentacle's axis, generating different degrees of bending and torsion. The spiking neural network sequentially triggers each unit to form a peristaltic wave, driving the tentacle to complete continuous actions of curling, grasping, and transferring objects. Other configurations of driving units can be flexibly designed according to specific application scenarios, and will not be elaborated upon here.

[0095] S4: Collect the deformation state information of the flexible actuator and convert it into a feedback pulse sequence.

[0096] In some embodiments, distributed deformation state information can be acquired in real time by multiple flexible sensors arranged on or inside the flexible actuator. This deformation state information is amplified, filtered, and level-converted by a signal conditioning circuit, and then converted into a feedback pulse sequence by a pulse coding circuit according to a preset coding strategy. This feedback pulse sequence is input to a spiking neural network processing unit as an adjustment signal for the network state evolution, enabling the network to correct the subsequent multi-channel pulse control signals in real time based on the error between the actual deformation and the target deformation, forming a closed-loop control. Converting distributed deformation information into a pulse sequence maintains the consistency of the entire control link in the pulse domain, avoiding the delay and accuracy loss caused by repeated conversion between analog and pulse signals in traditional schemes. Simultaneously, the multi-dimensional, multi-point deformation information allows the spiking neural network to more accurately perceive the overall shape of the flexible actuator, thereby achieving high-precision, highly robust collaborative deformation control.

[0097] For example, in bionic finger bending control, a flexible strain sensor attached to the back of the finger detects the resistance change corresponding to the bending angle (e.g., a 15% increase in resistance for 30°). The signal conditioning circuit converts this change into a 0–3.3V voltage, and the pulse coding circuit uses rate coding to convert it into a pulse frequency at a ratio of 10kHz / V (i.e., a 33kHz feedback pulse sequence for 30°). The pulse neural network processing unit compares this feedback frequency with the frequency corresponding to the target bending angle (e.g., 40kHz). If it finds that the frequency is too low, it automatically increases the density of the extensor channel output pulses, causing the finger to bend further until the feedback frequency reaches the target value, thereby achieving precise angle control.

[0098] In some embodiments, the deformation state information includes at least one of the following: local curvature information of the flexible actuator; surface strain distribution information of the flexible actuator; spatial pose information of the flexible actuator; deformation information of at least one of the plurality of flexible drive units; and contact force information between the flexible actuator and the external environment.

[0099] In a specific implementation, various sensors (such as flexible curvature sensors, strain sensor arrays, inertial measurement units, displacement sensors, and thin-film pressure sensors) deployed on the flexible actuator can collect local curvature information, surface strain distribution information, spatial pose information, deformation information of each driving unit, and contact force information with the external environment. These sensor signals, after being processed by a signal conditioning circuit (amplification, filtering, and level conversion), are converted into a multi-channel feedback pulse sequence by a pulse coding circuit according to a preset coding strategy (e.g., rate coding, which linearly maps curvature values, strain magnitudes, pose angles, deformation displacements, or contact force amplitudes to pulse frequencies; or time coding, which maps the rate of change to pulse intervals). This sequence is then input to the spiking neural network processing unit. By uniformly converting multi-dimensional, distributed deformation state information into pulse domain signals, the spiking neural network can comprehensively perceive the true shape of the flexible actuator and external interaction forces. Based on this rich feedback information, it can accurately correct the timing, amplitude, or pulse width of the multi-channel pulse control signals, achieving high-precision and highly robust closed-loop collaborative control.

[0100] For example, when a bionic finger grasps a fragile object, a flexible curvature sensor detects excessive knuckle curvature (exceeding the target value), a strain array shows uneven strain distribution on the fingertip surface, and a contact force sensor senses that the pressure is about to exceed a threshold. The feedback interface converts the curvature deviation into a high-frequency pulse sequence, the uneven strain into multi-channel frequency difference pulses, and the contact force overload into extremely high-frequency pulses; all three are simultaneously input into a pulse neural network. After comprehensive judgment, the network immediately reduces the driving pulse density of the extensor channel and increases the pulse width of the flexor channel, allowing the finger to relax appropriately and adjust its posture, thus safely grasping the object without damaging it.

[0101] S5: Input the feedback pulse sequence into the pulse neural network to adjust the multi-channel pulse control signal in a closed loop, so as to realize the motion output indicated by the target motion command.

[0102] In some embodiments, after the feedback pulse sequence is input into the spiking neural network, the spiking neural network compares it with the expected pulse sequence corresponding to the target motion command. Through internal synaptic plasticity mechanisms (such as pulse timing-dependent plasticity) or dedicated error detection circuits, it calculates the deviation between the current output and the actual state, and then dynamically adjusts the generation parameters of the multi-channel pulse control signal (such as the pulse firing frequency, relative timing, pulse density, etc. of each channel). This allows the adjusted drive signal to correct the deformation error of the flexible actuator and gradually approach the motion output indicated by the target motion command. This closed-loop adjustment method utilizes the event-driven characteristics of the spiking neural network to directly complete error detection and control correction in the pulse domain without the need for digital-to-analog conversion or external processor intervention, greatly shortening the feedback delay and improving the real-time performance and accuracy of the control. At the same time, the network can adaptively adjust the cooperative relationship of the multi-channel signals according to the real-time changes of the feedback pulse sequence, enabling the flexible actuator to accurately and smoothly complete the target motion even when subjected to external disturbances or its own characteristic drift.

[0103] For example, during the continuous lifting of a weight by the bionic arm, the actual bending angle is less than the target value due to load changes, and the frequency of the feedback pulse sequence is lower than the expected value. After the spiking neural network detects this deviation, it automatically increases the pulse frequency of the extensor drive channel and appropriately delays the pulse start time of the flexor channel, so that the arm slowly rises to the target angle and then returns to steady-state output after reaching it, thus achieving precise motion control under disturbance resistance.

[0104] In some embodiments, closed-loop adjustment of the multi-channel pulse control signal includes: dynamically adjusting the gain, bias, or timing relationship of the multi-channel pulse control signal based on the error between the deformation state information and the target deformation state, so as to reduce the error.

[0105] In a specific implementation, the pulse neural network processing unit internally includes an error detection circuit and an adjustable parameter register. The error detection circuit compares the feedback pulse sequence with the target pulse sequence to calculate the amplitude error (corresponding to deformation amplitude deviation) and timing error (corresponding to phase or velocity deviation). Then, based on the magnitude and direction of the error, it adjusts the gain (e.g., changing the multiplier of the programmable gain amplifier to adjust the drive signal amplitude), bias (e.g., superimposing a DC bias voltage to change the initial position), or timing relationship (e.g., adjusting the relative phase of the pulses in each channel via a programmable delay line) of each drive channel via the control bus. By independently adjusting the gain, bias, and timing, errors in deformation amplitude, zero-point drift, and motion coordination can be compensated respectively, enabling the flexible actuator to quickly and smoothly return to the target motion trajectory even under load changes, material fatigue, or external disturbances, significantly improving the system's control accuracy and adaptability.

[0106] For example, in an experiment where the bionic finger repeatedly grasps objects of different weights, when grasping a heavy object results in insufficient bending amplitude, the spiking neural network detects the amplitude error and automatically increases the gain of the extensor drive channel (increasing the drive voltage from 400V to 600V); when the finger exhibits initial bending deviation due to long-term use, the bias is adjusted (increasing the static bias voltage of the flexor channel) to return the finger to a neutral position; when the response speeds of the two drive units are inconsistent, causing bending jitter, the timing relationship between the two channels is finely adjusted (advancing the pulse on one side by 2ms), so that the finger can always grasp objects smoothly and accurately.

[0107] It should also be noted that this application provides two preferred methods for the specific implementation of error detection and parameter correction in closed-loop adjustment.

[0108] The first approach employs error detection neurons: A special type of error detection neuron is placed within the spiking neural network. Its input includes a feedback pulse sequence and a target pulse sequence (derived from a target motion command). The error detection neuron uses pulse timing-dependent plasticity (STDP) to calculate the difference between the feedback pulse frequency and the target pulse frequency in real time, converting this difference into adjustments to synaptic weights. When the frequency is too low, the weight of excitatory synapses is increased; when the frequency is too high, the weight of inhibitory synapses is increased. The adjusted synaptic weights directly alter the output pulse frequency of downstream motor neurons (such as extensor and flexor neurons), thereby correcting the deformation amplitude deviation.

[0109] The second method employs a time difference detection circuit: this circuit consists of a pair of delay chains and a phase detector, receiving a feedback pulse sequence and a reference pulse sequence (obtained by frequency division of the target pulse sequence), respectively, and measuring the time difference between their rising edges. When the feedback pulse leads the reference pulse, the phase detector outputs a positive pulse, and the corresponding programmable delay line is added through control logic, causing the drive signal to be output with a lag; when the feedback pulse lags, the delay is reduced or the trigger is advanced. The output of this time difference detection circuit can also be used to adjust the phase of the output pulse, thereby correcting deviations in motion speed or relative timing. These two methods can be used independently or in combination, enabling the pulse neural network to simultaneously compensate for deformation amplitude and motion coordination errors, achieving high-precision closed-loop control. For example, in the experiment of repeated bending of the bionic wrist, when the load increases and the bending amplitude is insufficient, the error detection neuron detects that the feedback pulse frequency is lower than the target frequency, automatically increases the synaptic weight of the extensor channel, and raises the driving voltage from 400V to 600V, and the amplitude is restored; when the response of the two driving units is inconsistent and bending and twisting occurs, the time difference detection circuit finds that the feedback pulse on one side is 2ms ahead of the other side, automatically delays the driving signal of the ahead side by 2ms, and the two sides are restored to synchronization, and the wrist resumes pure bending motion.

[0110] The control method of the flexible actuator of this application will be described in detail below through several specific embodiments.

[0111] Example 1

[0112] This embodiment specifically provides a control method for a bionic finger. The bionic finger is configured with three segments, each containing a pair of antagonistic dielectric elastomer driving units: the first driving unit is a high-stiffness columnar structure responsible for axial contraction, and the second driving unit is a low-stiffness thin-film structure responsible for bending. Each driving unit corresponds to a local micro-embedded boost circuit (3mm × 3mm in size, integrating a charge pump and a programmable gain amplifier). All boost circuits are embedded in the silicone inner wall of the finger. A flexible curvature sensor, a fingertip-integrated thin-film pressure sensor, and a fingertip-distributed strain sensor array are attached to the back of the finger. In scenarios involving grasping fragile objects (such as strawberries), this bionic finger achieves precise grasping with a bending angle of 30° and a contact force ≤0.5N through a spiking neural network that collaboratively generates multi-channel pulse control signals, a distributed driving architecture, dynamic energy distribution, and multi-dimensional feedback closed-loop control. The specific control process is as follows: After receiving the instruction "bend 30°, force ≤ 0.5N", the pulse neural network, in coordination with the decision-making layer, generates three pairs (one pair per finger joint) of pulse control signals, totaling six channels. The extensor channel of each finger joint outputs a pulse sequence with a frequency of 5kHz, a reference amplitude of 600V, and a pulse width of 5ms. The flexor channel outputs a pulse sequence with the same frequency but a starting amplitude of 300V, delayed by 8ms. The six signals are transmitted to the local boost circuit of the corresponding finger joint through flexible wires. The boost circuit dynamically adjusts the output parameters according to the configuration word issued by the pulse neural network in real time: in the initial stage (0-50ms), the extensor amplitude is set to 800V and the pulse width to 8ms, and the flexor amplitude is set to 200V and the pulse width to 2ms to quickly start bending. When approaching the target angle, feedback is used to... The closed-loop mechanism reduces the extensor amplitude to 500V and increases the flexor pulse width to 5ms to provide damping. Simultaneously, a flexible curvature sensor detects the actual bending angle of each phalanx at a 1kHz sampling rate, a strain sensor detects surface strain distribution, and a thin-film pressure sensor detects contact force. After conditioning, the sensor signals are converted into a feedback pulse sequence via rate encoding (a 30° bending angle corresponds to a 30kHz pulse frequency). Error detection neurons within the spiking neural network compare the feedback frequency with the target frequency. If the deviation exceeds ±2kHz, the corresponding channel synaptic weight is adjusted via the STDP mechanism: if the angle is insufficient, the extensor channel gain is increased (amplitude step +20V); if the contact force exceeds the limit, the extensor pulse width is immediately reduced and the flexor pulse width is increased. The closed-loop adjustment period is 5ms. This embodiment achieves synchronous and coordinated bending of multiple phalanges through collaborative generation and dynamic energy distribution. The multi-dimensional feedback closed loop ensures precise control of position and force, thus avoiding damage to objects. The distributed micro-boost circuit eliminates high-voltage cables at the fingertips, improving flexibility and integration.

[0113] Example 2

[0114] This embodiment provides a control method for a bionic arm. The bionic arm is configured such that the elbow joint is driven by three sets of parallel dielectric elastomer muscle bundles. Each muscle bundle includes four first drive units (axially extendable) and two second drive units (flexible). Each muscle bundle is equipped with a small, flexible, integrated boost circuit, which measures 15mm × 15mm × 2mm, uses a transformer boost topology, and has an output power of 15W. The three boost circuits are integrated into the inner side of the upper arm via a flexible circuit board. An inertial measurement unit is installed at the elbow joint to detect posture, and a strain array is attached to the muscle surface, along with an integrated force sensor to detect load. In scenarios involving lifting heavy objects (such as a 5kg dumbbell), this bionic arm utilizes a spiking neural network to collaboratively generate multi-channel pulse control signals, a distributed drive architecture, dynamic energy distribution, and multi-dimensional feedback closed-loop control to achieve rapid lifting from a horizontal position to 60° and maintain this position for 5 seconds against external disturbances. The specific control process is as follows: After receiving the "rapidly lift 60°" command, the pulse neural network collaboratively generates 12 channels of pulse control signals (3 groups × 4 extension + 2 bending). In the initial stage (0-100ms), all extension channels output a pulse sequence with a frequency of 8kHz, an amplitude of 900V, and a pulse width of 10ms, while the bending channel outputs a pulse sequence with a frequency of 2kHz, an amplitude of 400V, a pulse width of 5ms, and a lag of 15ms. The 12 signals are sent to three local boost circuits for independent amplification through equal-length flexible wires, avoiding the high current transmission loss of the centralized high-voltage bus. The pulse neural network dynamically adjusts the output parameters of each channel according to the actual angle and angular velocity feedback from the inertial measurement unit: 0-100ms During the rapid lifting phase, all extension channels maintain full-amplitude output. As the target approaches within 100-200ms, the central muscle bundle extension amplitude decreases to 600V, and the lateral bundles decrease to 500V to adjust balance. During the holding phase (5 seconds), the extension channel amplitude decreases to 400V with a pulse width of 2ms, and the bending channel output stops, maintaining only isometric contraction to resist gravity. Simultaneously, the inertial measurement unit provides real-time angle feedback, the strain array detects muscle elongation, and the force sensor detects load changes. When an external disturbance causes an angle decrease, the feedback pulse frequency decreases, and the error detection neuron immediately increases the extension channel amplitude to 1000V and increases the pulse width to 15ms, generating an instantaneous compensating torque, restoring the original angle within milliseconds. This embodiment shortens the high-voltage path through a distributed drive architecture, improving safety and integration; dynamic energy distribution adjusts according to load and position, achieving rapid response and energy saving; and the multi-dimensional feedback closed loop effectively resists disturbances, ensuring the stability of lifting heavy objects.

[0115] Example 3

[0116] This embodiment provides a control method for a soft crawling robot. The robot is configured to consist of eight segments connected in series. Each segment includes a pair of photosensitive bending units made of photoresponsive hydrogel and a pair of thermosensitive contraction units made of shape memory alloy wires. The photosensitive units are driven by micro-LED driver chips attached to the surface of the hydrogel, and the thermosensitive units are driven by micro-heating driver chips attached to the alloy wires. All driver chips are connected to a central pulse neural network processing unit via flexible wires. Each segment is equipped with a flexible curvature sensor and a temperature sensor. In a scenario where the soft crawling robot moves forward in a creeping motion on rough ground, stable creeping wave propulsion from the tail to the head is achieved through the autonomous evolution of multi-channel pulse control signals by the central pattern generator inside the pulse neural network, distributed drive and energy form conversion, dynamic energy distribution, and multi-dimensional feedback closed-loop control. The specific control process is as follows: The spiking neural network is configured with a topology that simulates a spinal cord central pattern generator, consisting of eight groups of cross-inhibited neurons. It only requires a single "forward" initiation pulse to autonomously generate eight sets of multi-channel pulse control signals with fixed phase differences (each lagging by 45°). Each signal set includes an optically driven pulse (1kHz frequency, 10ms pulse width) and a thermally driven pulse (500Hz frequency, 20ms pulse width). The LED driver chip in each segment converts the electrical pulses into 450nm blue light pulses. The light intensity is controlled by adjusting the pulse width, and the heated driver chip amplifies the pulse current. Up to 2A, the driving alloy wire generates Joule heating, raising the temperature from 25°C to 70°C. A pulse neural network dynamically adjusts the energy distribution of each segment based on ground slope feedback. When going uphill, the thermal driving pulse width of the rear segments is increased to 30ms and the optical driving pulse width to 15ms to generate greater thrust; when going downhill, the energy is reduced. Simultaneously, a curvature sensor provides feedback on the bending angle, and a temperature sensor provides feedback on the alloy wire temperature. These feedback pulse sequences are input to a central pattern generator network. The network fine-tunes the pulse start time of each segment (via a programmable delay line) based on the phase difference between the actual bending and the target waveform, ensuring the creeping wave maintains a stable rhythm on rough ground. This embodiment greatly simplifies the control logic through autonomous collaborative generation by the central pattern generator, eliminating the need for real-time calculation of the segment timing. Distributed optical / thermal actuation achieves multi-physics compatibility and improves energy conversion efficiency. A multi-dimensional feedback closed loop enables the robot to adapt to different terrains, exhibiting biological-like adaptive capabilities.

[0117] Example 4

[0118] This embodiment provides a control method for the wrist of a bionic hand. The drive unit of the bionic hand's wrist is configured to consist of a first drive unit (axial extension, responsible for bending), a second drive unit (bending, responsible for assisting bending), and a third drive unit (spirally wound dielectric elastomer fiber, responsible for torsion). Each drive unit corresponds to a local micro-embedded boost circuit. The wrist is equipped with a six-axis inertial measurement unit, multiple curvature sensors, and torque sensors. In scenarios requiring simultaneous 30° bending and 20° torsion (such as wringing a towel), this bionic wrist achieves decoupled composite bending and torsion motion through the coordinated generation of multi-channel pulse control signals via a spiking neural network, distributed drive and dynamic energy distribution, multi-dimensional feedback closed loop, and coordinated deformation of different drive units. The specific control process is as follows: The pulse neural network collaboratively generates three-channel pulse control signals according to the "bend 30° + twist 20°" command. The bending main channel (first drive unit) outputs a frequency of 6kHz, an amplitude of 700V, and a pulse width of 8ms. The bending auxiliary channel (second drive unit) outputs with a 5ms lag. The twist channel (third drive unit) outputs a frequency of 4kHz, an amplitude of 500V, and a pulse width of 6ms, with its start time synchronized with the bending main channel but its phase modulation is in a spiral mode. Each channel signal is independently amplified by its corresponding local boost circuit. During the 0-80ms bending phase, the amplitude of the bending main channel increases linearly from 700V to 900V, the auxiliary channel increases from 300V to 500V, and the twist channel maintains 500V. During the 80-150ms twist phase, the amplitude of the bending channel drops to 400V to maintain the posture and twist. The channel amplitude is increased to 800V and the pulse width is increased to 10ms to achieve torsion. Simultaneously, the inertial measurement unit provides real-time feedback of the wrist Euler angles, the curvature sensor detects local bending curvature, and the torque sensor detects torsional torque. After the feedback pulse sequence is input into the spiking neural network, error detection neurons calculate the bending angle deviation and torsion angle deviation respectively. The synaptic weights of the bending and torsion channels are independently adjusted through a pulse timing-dependent plasticity mechanism. When torsion causes unexpected bending coupling, the time difference detection circuit fine-tunes the starting phase of the torsion channel (advancing or lagging by 2ms) to decouple the two movements. The first driving unit generates the main bending displacement, the second driving unit provides bending smoothness, and the third driving unit generates torsion. These three elements achieve composite motion based on a pre-set synergistic relationship in the spiking neural network (bending precedes torsion by 20ms). In this embodiment, the coordinated deformation of three different driving units achieves multi-degree-of-freedom composite motion. Dynamic energy distribution ensures a reasonable allocation of energy between bending and torsion, and the multi-dimensional feedback closed loop effectively decouples motion coupling, improving the accuracy and compliance of the composite motion.

[0119] Example 5

[0120] This embodiment provides a bionic facial expression driving method. The bionic face is configured with 32 miniature dielectric elastomer driving units, each 2mm × 2mm in size. These driving units control actions such as raising the corners of the mouth, contracting the orbicularis oculi muscle, and lifting the frontalis muscle. Each driving unit corresponds to a miniature embedded boost circuit, which can be 1.5mm × 1.5mm in size. All boost circuits are integrated onto a flexible film and attached to the inner side of the facial silicone skin. Miniature curvature sensors and strain sensors are arranged throughout the face. In scenarios where the bionic face simulates expressions such as smiling and frowning, a delicate and natural expression simulation is achieved through the collaborative generation of multi-channel pulse control signals by a spiking neural network, a distributed driving architecture, dynamic energy distribution, and multi-dimensional feedback closed-loop control. The specific control process is as follows: The spiking neural network collaboratively generates 32 channels of pulse control signals based on the "smile" command. Eight channels related to the corners of the mouth output pulses with a frequency of 7kHz and an amplitude of 900V; six channels related to the eye area output pulses with a frequency of 5kHz and an amplitude of 600V; other areas output pulses with a frequency of 2kHz and an amplitude of 200V, or no output at all. The timing of each channel is preset by the collaborative control layer based on facial anatomy (the corners of the mouth unit starts 5ms before the eye area unit). All boost circuits receive pulse signals through flexible microstrip lines, and each circuit independently drives its corresponding micro-drive unit. The high-voltage transmission distance is less than 5mm. The electromagnetic interference is extremely low. The pulse neural network dynamically adjusts the output parameters of each channel based on the feedback curvature information. When the upward movement of the left corner of the mouth is insufficient, the pulse width (from 5ms to 8ms) and amplitude (from 900V to 1000V) of the relevant channel on the left are increased. When the movement around the eyes is too large, the amplitude is reduced. The curvature sensor detects the curvature of each area of ​​the face in real time, and the strain sensor detects the skin stretching. After the feedback pulse sequence is input, the pulse neural network calculates the deformation error of each area. By independently adjusting the gain and bias of each channel, the expression is accurately matched to the target (curvature of the corner of the mouth ≥0.2mm⁻¹ when smiling, curvature of the eye area ≤0.05mm⁻¹). This embodiment achieves delicate facial expressions through ultra-high density distributed drive. Cooperative generation ensures the natural coordination of the expression. The multi-dimensional feedback closed loop ensures the accuracy and repeatability of the expression. The miniaturized boost circuit allows the entire system to be embedded in ultra-thin silicone skin, with excellent invisibility and flexibility.

[0121] This application provides a control device for a flexible actuator, which may include multiple independently driveable flexible drive units.

[0122] Figure 2 This is a schematic diagram of the structure of a control device for a flexible actuator provided in an embodiment of this application.

[0123] Reference Figure 2 As shown, the control device for the flexible actuator provided in this application embodiment may include:

[0124] Command conversion module 201 is used to convert target motion commands into pulse sequences;

[0125] The first processing module 202 is used to input the pulse sequence into the spiking neural network, and the spiking neural network collaboratively generates multi-channel pulse control signals;

[0126] The second processing module 203 is used to convert the multi-channel pulse control signal into a drive signal and apply it to multiple flexible drive units so that the multiple flexible drive units can generate coordinated deformation.

[0127] The acquisition module 204 is used to acquire the deformation state information of the flexible actuator and convert it into a feedback pulse sequence;

[0128] The third processing module 205 is used to input the feedback pulse sequence into the pulse neural network to adjust the multi-channel pulse control signal in a closed loop, so as to realize the motion output indicated by the target motion command.

[0129] In one possible implementation, the instruction conversion module 201 converts the target motion instruction into a pulse sequence by generating a corresponding pulse sequence according to at least one of the motion direction, motion amplitude, motion speed, or motion trajectory of the target motion instruction, using at least one of rate encoding, time encoding, or group encoding.

[0130] In one possible implementation, the first processing module 202 generates multi-channel pulse control signals in a coordinated manner through a spiking neural network. The spiking neural network determines the generation strategy of each channel's pulse control signal according to the target motion command, so that the pulse control signals of each channel are correlated in terms of timing, frequency, or pulse density, so as to jointly realize the motion output indicated by the target motion command.

[0131] In one possible implementation, the pulse control signals of each channel are correlated with each other in terms of timing, frequency, or pulse density, including: there is a preset cooperative relationship between the multi-channel pulse control signals, and the cooperative relationship includes at least one of the following: there is a predetermined relative timing relationship between the pulse control signals of each channel; the pulse frequency ratio of the pulse control signals of each channel is a preset value; and the pulse density distribution of the pulse control signals of each channel conforms to a preset cooperative mode.

[0132] In one possible implementation, the spiking neural network also dynamically adjusts the compliance and stability of the coordinated deformation by forming dynamic antagonism among multiple flexible drive units based on the dynamic relative temporal relationship of the real-time feedback signal.

[0133] In one possible implementation, the spiking neural network includes a cooperative control layer, comprising: multiple input neurons for receiving pulse sequences; multiple interneurons forming excitatory and / or inhibitory connections with the multiple input neurons; and multiple output neurons forming connections with the multiple interneurons, wherein the spiking activities of the multiple output neurons correspond to multi-channel pulse control signals; wherein the interneurons are configured to cooperatively drive the multiple output neurons to generate pulses according to the spiking activities of the multiple input neurons, thereby forming a preset cooperative relationship between the multi-channel pulse control signals.

[0134] In one possible implementation, the spiking neural network collaboratively generates multi-channel pulse control signals, including: the spiking neural network dynamically adjusts the collaborative relationship between the multi-channel pulse control signals according to the feedback pulse sequence to adapt to the deformation state changes of the flexible actuator.

[0135] In one possible implementation, the second processing module 203 converts the multi-channel pulse control signal into a drive signal by amplifying the power of the multi-channel pulse control signal and / or converting its energy form to generate a drive signal having a voltage, field strength, or energy sufficient to drive the corresponding flexible drive unit to produce deformation. The drive signal is at least one of an electric field signal, an optical signal, a thermal signal, or a magnetic field signal.

[0136] In one possible implementation, the multi-channel pulse control signal is converted into a drive signal and executed by multiple drive circuits corresponding one-to-one with each channel. The multiple drive circuits are physically distributed on or near the flexible actuator body.

[0137] In one possible implementation, during the process of converting the multi-channel pulse control signal into a drive signal, the second processing module 203 dynamically adjusts the output parameters of the corresponding channel of each drive signal; wherein the output parameters include at least one of the following: amplitude, pulse width or duration; the output parameters of the channel are related to the deformation or deformation rate required by the corresponding flexible drive unit.

[0138] In one possible implementation, the plurality of flexible drive units include a first drive unit and a second drive unit; when a drive signal is applied to the corresponding flexible drive unit, the first drive unit generates a first deformation and the second drive unit generates a second deformation; wherein, the first deformation includes axial extension and contraction, and the second deformation includes bending.

[0139] In one possible implementation, the deformation state information includes at least one of the following: local curvature information of the flexible actuator; surface strain distribution information of the flexible actuator; spatial pose information of the flexible actuator; deformation information of at least one of the multiple flexible drive units; and contact force information between the flexible actuator and the external environment.

[0140] In one possible implementation, closed-loop adjustment of the multi-channel pulse control signal includes: comparing the feedback pulse sequence with the pulse sequence corresponding to the target motion command, and adjusting the network parameters of the spiking neural network according to the comparison result to update the generation strategy of the multi-channel pulse control signal.

[0141] In one possible implementation, the network parameters include at least one of the following: synaptic weights between neurons; impulse firing thresholds of neurons; membrane potential time constants of neurons; and connection topology between neurons.

[0142] In one possible implementation, closed-loop adjustment of the multi-channel pulse control signal includes: dynamically adjusting the gain, bias, or timing relationship of the multi-channel pulse control signal based on the error between the deformation state information and the target deformation state, so as to reduce the error.

[0143] Figure 3 This is a schematic diagram of the structure of a computing device provided in one embodiment of this application.

[0144] Reference Figure 3 As shown, this application embodiment also provides a computing device 300, including a processor 301, which is used to execute program code, causing the computing device 300 to execute the control method of the flexible actuator provided in any embodiment of this application.

[0145] Figure 4 This is a schematic diagram of the structure of a computing device cluster provided in one embodiment of this application.

[0146] Reference Figure 4 As shown, this application embodiment also provides a computing device cluster 400, including multiple computing devices 300, each computing device 300 including a processor 301, the processor 301 being used to execute program code, causing the computing device cluster to execute the control method of the flexible actuator provided in any embodiment of this application.

[0147] This application also provides a computer program product containing instructions that, when executed by a computing device cluster, cause the computing device cluster to perform the control method for the flexible actuator provided in any embodiment of this application.

[0148] This application also provides a computer-readable storage medium including computer program instructions. When the computer program instructions are executed by a computing device, the computing device executes the control method of the flexible actuator provided in any embodiment of this application.

[0149] This application also provides a robot, including the computing device 300 or computing device cluster 400 provided in any of the above embodiments.

[0150] In this application, the terms "first," "second," etc., are used to distinguish identical or similar items that have substantially the same function and purpose. It should be understood that there is no logical or temporal dependency between "first" and "second," nor does it limit the quantity or order of execution. It should also be understood that although the following description uses the terms "first," "second," etc., to describe various elements, these elements should not be limited by the terms. These terms are merely used to distinguish one element from another.

[0151] In this application, the term "at least one" means one or more, and the term "multiple" means two or more.

[0152] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A control method for a flexible actuator, the flexible actuator comprising a plurality of independently driveable flexible drive units, characterized in that, include: Convert the target motion command into a pulse sequence; The pulse sequence is input into a pulsating neural network, which generates a multi-channel pulse control signal. The multi-channel pulse control signal is converted into a drive signal and applied to multiple flexible drive units to cause the multiple flexible drive units to produce coordinated deformation; The deformation state information of the flexible actuator is collected and converted into a feedback pulse sequence; The feedback pulse sequence is input into the pulse neural network to adjust the multi-channel pulse control signal in a closed loop, thereby realizing the motion output indicated by the target motion command. The spiking neural network generates multi-channel pulse control signals, including: The pulse neural network determines the generation strategy of pulse control signals for each channel in a unified manner according to the target motion command, so that the pulse control signals of each channel are correlated in terms of timing, frequency or pulse density, so as to jointly realize the motion output indicated by the target motion command.

2. The method according to claim 1, characterized in that, The pulse control signals of each channel are interconnected in terms of timing, frequency, or pulse density, including: The multi-channel pulse control signals have a preset coordination relationship, which includes at least one of the following: There is a predetermined relative timing relationship between the pulse control signals of each channel; The pulse frequency ratio of each channel's pulse control signal is a preset value; The pulse density distribution of the pulse control signals in each channel conforms to the preset cooperative mode.

3. The method according to claim 2, characterized in that, The spiking neural network also dynamically adjusts the relative timing relationship based on real-time feedback signals to form dynamic antagonism among the multiple flexible drive units, thereby regulating the compliance and stability of the cooperative deformation.

4. The method according to claim 1, characterized in that, The spiking neural network includes a cooperative control layer, which includes: Multiple input neurons are used to receive the pulse sequence; Multiple interneurons form excitatory and / or inhibitory connections with the multiple input neurons; Multiple output neurons are connected to multiple intermediate neurons, and the pulse activity of the multiple output neurons corresponds to the multi-channel pulse control signal respectively; The intermediate neuron is configured to collaboratively drive the multiple output neurons to generate pulses based on the pulse activity of the multiple input neurons, thereby forming a preset collaborative relationship among the multi-channel pulse control signals.

5. The method according to claim 1, characterized in that, The spiking neural network generates multi-channel pulse control signals, including: The spiking neural network dynamically adjusts the coordination relationship between the multi-channel pulse control signals according to the feedback pulse sequence to adapt to the deformation state changes of the flexible actuator.

6. The method according to claim 1, characterized in that, Converting the multi-channel pulse control signal into a drive signal includes: The multi-channel pulse control signal is amplified and / or its energy form is converted to generate a driving signal with sufficient voltage, field strength or energy to drive the corresponding flexible driving unit to produce deformation. The driving signal is at least one of an electric field signal, an optical signal, a thermal signal or a magnetic field signal.

7. The method according to claim 6, characterized in that, The conversion of multi-channel pulse control signals into drive signals is executed by multiple drive circuits corresponding one-to-one with each channel. These multiple drive circuits are physically distributed on or near the body of the flexible actuator.

8. The method according to claim 1 or 6, characterized in that, During the process of converting the multi-channel pulse control signal into the drive signal, the output parameters of each drive signal corresponding to the channel are dynamically adjusted. The output parameters include at least one of the following: amplitude, pulse width, or duration; The output parameters of the channel are associated with the deformation or deformation rate required by the corresponding flexible drive unit.

9. The method according to claim 1, characterized in that, The plurality of flexible drive units include a first drive unit and a second drive unit; when the drive signal is applied to the corresponding flexible drive unit, the first drive unit generates a first deformation and the second drive unit generates a second deformation; wherein, the first deformation includes axial extension and contraction, and the second deformation includes bending.

10. The method according to claim 1, characterized in that, The process of converting target motion commands into pulse sequences includes: Based on at least one of the motion direction, motion amplitude, motion speed, or motion trajectory of the target motion command, a corresponding pulse sequence is generated using at least one of rate encoding, time encoding, or group encoding.

11. The method according to claim 1, characterized in that, The deformation state information includes at least one of the following: local curvature information of the flexible actuator; surface strain distribution information of the flexible actuator; and spatial pose information of the flexible actuator. Deformation information of at least one of the plurality of flexible drive units; The contact force information between the flexible actuator and the external environment.

12. The method according to claim 1, characterized in that, The closed-loop adjustment of the multi-channel pulse control signal includes: The feedback pulse sequence is compared with the pulse sequence corresponding to the target motion command, and the network parameters of the spiking neural network are adjusted according to the comparison result to update the generation strategy of the multi-channel pulse control signal.

13. The method according to claim 12, characterized in that, The network parameters include at least one of the following: synaptic weights between neurons; impulse firing thresholds of neurons; membrane potential time constants of neurons; and connection topology between neurons.

14. The method according to claim 1, characterized in that, The closed-loop adjustment of the multi-channel pulse control signal includes: Based on the error between the deformation state information and the target deformation state, the gain, bias, or timing relationship of the multi-channel pulse control signal is dynamically adjusted to reduce the error.

15. A control device for a flexible actuator, the flexible actuator comprising a plurality of independently driveable flexible drive units, characterized in that, include: The instruction conversion module is used to convert target motion instructions into pulse sequences; The first processing module is used to input the pulse sequence into a pulse neural network, and the pulse neural network generates a multi-channel pulse control signal. The second processing module is used to convert the multi-channel pulse control signal into a drive signal and apply it to the multiple flexible drive units so that the multiple flexible drive units can generate coordinated deformation. The acquisition module is used to acquire the deformation state information of the flexible actuator and convert it into a feedback pulse sequence; The third processing module is used to input the feedback pulse sequence into the pulse neural network to adjust the multi-channel pulse control signal in a closed loop, so as to realize the motion output indicated by the target motion command. The spiking neural network generates multi-channel pulse control signals, including: The pulse neural network determines the generation strategy of pulse control signals for each channel in a unified manner according to the target motion command, so that the pulse control signals of each channel are correlated in terms of timing, frequency or pulse density, so as to jointly realize the motion output indicated by the target motion command.

16. A computing device, characterized in that, The computing device includes a processor for executing program code, causing the computing device to perform the control method for the flexible actuator as described in any one of claims 1-14.

17. A computing device cluster, characterized in that, It includes multiple computing devices, each computing device including a processor, the processor being used to execute program code, causing the cluster of computing devices to perform a control method for a flexible actuator as described in any one of claims 1-14.

18. A robot, characterized in that, This includes the computing device as described in claim 16 or the cluster of computing devices as described in claim 17.

Images