A method and system for controlling a biomimetic robot based on a muscle fiber model

Abstract

The application belongs to the technical field of robot control, and discloses a kind of bionic robot control method and system based on muscle fiber model, the method includes: based on scale division mechanism, construct multiscale muscle model, and utilize physiological state variable and fatigue accumulation variable, optimize multiscale muscle model, obtain muscle fiber model;According to muscle group cooperation relationship and the kinematic model of robot limb, combined with deformation compensation strategy, the geometry structure of model is adapted;Based on the fusion of multi-sensor data, the muscle fiber model and the model after geometry structure adaptation are used to construct feedback regulation mechanism;Based on adaptive neural network algorithm, muscle fiber model, geometry structure after adaptation and feedback regulation mechanism are optimized respectively, and combined with multi-objective optimization algorithm to solve, obtain the command information of bionic robot limb and execute.The application realizes accurate control of bionic robot limb joint position, torque and impedance.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and in particular to a biomimetic robot control method and system based on a muscle fiber model. Background Technology

[0002] In the continuous development of robotics technology, biomimetic robots, which need to simulate the complex limb movements of living organisms, have always focused on and challenged the precise, flexible, and adaptive control of their limbs. Traditional biomimetic robot control methods are often based on rigid mechanical structures and simple control algorithms, which show obvious limitations when facing complex and ever-changing tasks and environments.

[0003] For example, in human interaction scenarios, bionic robots need to adjust their limb movements in real time according to human actions, language, and environmental changes to achieve natural and harmonious interaction; when walking on complex terrain, they need to precisely control the movement of limb joints to maintain body balance and stability. However, existing control methods are insufficient to enable bionic robot limbs to achieve the same level of flexibility and intelligent response as biological muscles, and cannot fully meet the needs of these application scenarios.

[0004] Therefore, how to provide a biomimetic robot control method and system based on muscle fiber models is an urgent problem to be solved. Summary of the Invention

[0005] This invention provides a bionic robot control method and system based on a muscle fiber model to address the shortcomings of existing bionic robot limb control methods in terms of compliance, adaptability, and intelligence.

[0006] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or to describe the scope of protection of these embodiments. Its sole purpose is to present some concepts in a simple form as a prelude to the detailed description that follows.

[0007] According to a first aspect of the present invention, a biomimetic robot control method based on a muscle fiber model is provided.

[0008] In one embodiment, a biomimetic robot control method based on a muscle fiber model includes the following steps:

[0009] Based on the scale division mechanism, a multi-scale muscle model is constructed, and the parameters of the multi-scale muscle model are optimized using physiological state variables and fatigue accumulation variables to obtain a muscle fiber model.

[0010] Based on the muscle group coordination relationship and the kinematic model of the robot limb, and combined with the deformation compensation strategy, the muscle fiber model is geometrically adapted.

[0011] Based on fused multi-sensor data, a feedback regulation mechanism is constructed using muscle fiber models and geometrically adapted muscle fiber models.

[0012] Based on the adaptive neural network algorithm, the parameters of the muscle fiber model, the muscle fiber model after geometric structure adaptation, and the feedback regulation mechanism are optimized respectively. The solution is then obtained by combining a multi-objective optimization algorithm to obtain and execute the command information of the bionic robot limb.

[0013] According to a second aspect of the present invention, a biomimetic robot control system based on a muscle fiber model is provided.

[0014] In one embodiment, a biomimetic robot control system based on a muscle fiber model includes:

[0015] The muscle fiber model construction module is used to construct a multi-scale muscle model based on the scale division mechanism, and to optimize the parameters of the multi-scale muscle model using physiological state variables and fatigue accumulation variables to obtain a muscle fiber model.

[0016] The geometric structure adaptation module is used to adapt the geometric structure of the muscle fiber model based on the muscle group coordination relationship and the kinematic model of the robot limb, combined with the deformation compensation strategy.

[0017] The feedback regulation mechanism construction module is used to construct a feedback regulation mechanism based on fused multi-sensor data, utilizing muscle fiber models and geometrically adapted muscle fiber models.

[0018] The command generation and execution module is used to optimize the parameters of the muscle fiber model, the geometrically adapted muscle fiber model, and the feedback regulation mechanism based on the adaptive neural network algorithm, and to solve the problem by combining a multi-objective optimization algorithm to obtain and execute the command information of the bionic robot limb.

[0019] According to a third aspect of the present invention, a computer device is provided.

[0020] In some embodiments, the computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0021] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.

[0022] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the above method.

[0023] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0024] 1. This invention, through precise muscle fiber model simulation and combined with introduced formulas for force generation, force-length, and force-velocity, enables precise control of the position, torque, and impedance of the limb joints of bionic robots. Position control accuracy can reach ±0.1°, and torque control accuracy can reach ±0.5 N·m, making the robot's movement smoother and more natural, closely resembling the effect of biological muscle movement. In human interaction scenarios, it allows bionic robots to better cooperate with humans, avoiding harm; in delicate operation tasks, it improves the accuracy and stability of operations.

[0025] 2. This invention utilizes a real-time parameter adjustment mechanism based on internal and external sensor data, along with precise control of individual muscle fibers, enabling the robot to quickly adapt to different environmental changes and complex task requirements, with an environmental adaptation response time of less than 100ms. Whether navigating rugged terrain or manipulating objects of varying shapes and weights, the robot can efficiently complete tasks by automatically adjusting its control strategy. In dynamic environments, it can avoid obstacles in real time, maintaining the continuity and stability of its movement; during multi-task switching, it can quickly adjust control parameters to ensure smooth task execution.

[0026] 3. By combining external sensors (such as brain signal sensors and electromyography (EMG) sensors to perceive the human user's movement intentions, the robot can interact with humans more naturally and intelligently, with an intention recognition accuracy rate greater than 90%. Simultaneously, the adaptive learning capability of the muscle fiber model and the decision-making mechanism based on perceived data enable the bionic robot to possess a certain degree of autonomous decision-making ability, allowing it to make reasonable decisions in complex situations and improve the success rate of task execution. In human-robot collaborative work scenarios, the robot can autonomously adjust its working methods according to the human user's intentions and the on-site conditions, achieving efficient collaboration with humans; in autonomous navigation and obstacle avoidance tasks, it can autonomously plan its movement path based on environmental perception information to complete the predetermined task.

[0027] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0028] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0029] Figure 1 This is a flowchart illustrating a biomimetic robot control method based on a muscle fiber model according to an exemplary embodiment;

[0030] Figure 2 This is a schematic diagram illustrating the principle of a biomimetic robot control system based on a muscle fiber model, according to an exemplary embodiment.

[0031] Figure 3 This is a schematic diagram of a muscle fiber model structure for a biomimetic robot control method based on a muscle fiber model, according to an exemplary embodiment.

[0032] Figure 4 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment. Detailed Implementation

[0033] Figure 1 This invention illustrates a biomimetic robot control method based on a muscle fiber model.

[0034] In this optional embodiment, the biomimetic robot control method based on muscle fiber models includes:

[0035] S101. Based on the scale division mechanism, a multi-scale muscle model is constructed, and the parameters of the multi-scale muscle model are optimized using physiological state variables and fatigue accumulation variables to obtain a muscle fiber model.

[0036] It should be explained that the muscle fiber model construction provides a standardized library of muscle mechanics parameters (such as maximum tension, contraction velocity, and stiffness coefficient) as a benchmark for geometric adaptation and feedback control. For example, the constructed biceps model outputs the following characteristic parameter: maximum isometric tension. F iso =600N, muscle belly length L m =25cm, tendon stiffness K t =100 N / mm.

[0037] In this optional embodiment, a multi-scale muscle model is constructed based on a scale partitioning mechanism, and the parameters of the multi-scale muscle model are optimized using physiological state variables and fatigue accumulation variables to obtain a muscle fiber model, including the following steps:

[0038] A multi-scale muscle model is constructed based on a three-level division of molecular, fiber, and tissue scales.

[0039] By using physiological state variables, the nonlinear parameters of the multi-scale muscle model are dynamically adjusted to obtain the dynamically adjusted multi-scale muscle model.

[0040] Specifically, a multi-scale modeling approach is employed, employing a hierarchical modeling method from the microscopic sarcomere to the macroscopic muscle bundle. At the microscopic level, a contraction model of the sarcomere is established, considering the interaction between actin and myosin, and the force generation of the sarcomere is described by the following formula:

[0041] ;

[0042] In the formula, F cross Cross-bridge force, which is the instantaneous contractile force generated when myosin cross-bridges in the sarcomere bind to actin, is the microscopic force source of muscle contraction. n This refers to the number of cross bridges; p The force generated by a single cross bridge; i This refers to the angle at which cross-bridges bind to actin. At a macroscopic level, numerous sarcomeres are linked in series to form muscle fibers, and in parallel to form muscle bundles. Macroscopic muscle force is obtained through the force output of an integral microscopic model.

[0043] By using fatigue cumulative variables, muscle performance decay was simulated on a dynamically adjusted multi-scale muscle model to obtain a muscle fiber model.

[0044] It's important to explain that multi-scale modeling provides the structural foundation for the entire muscle fiber model. Through a three-tiered scale division (molecular scale: actin-myosin binding dynamics; fiber scale: tandem / parallel relationships of sarcomere contractile units; tissue scale: spatial distribution of muscle bundles), a mapping relationship is established between microscopic molecular interactions and macroscopic force output. For example, the maximum contractile force of a single muscle fiber can be calculated at the molecular scale. F fib =0.5 N The basic force of the muscle bundle is then obtained by the number of parallel fibers at the fiber level (e.g., 1000 fibers). F bundle =500 N .

[0045] Static parameters based on multi-scale models (such as initial sarcomere length) L 0), introduce physiological state variables (such as activation level) a ∈[0,1], contraction speed v The force output is corrected in real time through nonlinear equations:

[0046] ;

[0047] In the formula, F Indicates the current contractile force of the muscle; F bundle This indicates the force output of the muscle bundle; a Indicates activation level; v Indicates the rate of contraction; vmax The maximum contraction velocity is determined by the type of muscle fiber in the multi-scale model; fast-twitch fibers... v max =3 L 0 / s slow muscle fibers v max =1.5 L 0 / s .

[0048] Dynamic adjustment of nonlinear parameters enables static models to have dynamic response capabilities, bridging structural modeling and functional simulation.

[0049] Based on the force output results of the first two layers, fatigue accumulation variables are introduced, such as ATP (adenosine triphosphate) consumption rate. C Dynamic decline in muscle performance;

[0050] ;

[0051] In the formula, F fatigue This represents the output force under fatigue conditions; F Indicates the current contractile force of the muscle; K Indicates the fatigue coefficient; t Indicates the duration of muscle contraction; C This indicates the consumption rate of adenosine triphosphate.

[0052] Its parameters (such as) C The initial value is determined by the myofiber metabolic properties in the multi-scale model, while the decay rate is adjusted by nonlinear parameters (such as during high-speed contraction). k (Increased by 2 times), enabling full lifecycle simulation of structure, dynamics, and decay.

[0053] In this optional embodiment, the multi-scale muscle model includes an actin-myosin binding dynamics model, an integrated model of the serial-parallel relationship of sarcomere contractile units, and a mapping model for establishing the spatial distribution of muscle bundles.

[0054] The actin-myosin binding kinetic model is used to simulate actin-myosin binding kinetics at the molecular scale and calculate the contractile force of a single muscle fiber.

[0055] The integrated model of series and parallel relationships of sarcomere contractile units is used to integrate the contractile force of a single muscle fiber at the fiber scale to obtain the force output of multiple muscle fibers.

[0056] The established mapping model for the spatial distribution of muscle bundles is used to integrate the force output of multiple muscle fibers at the tissue scale to obtain the force output of muscle bundles.

[0057] It needs to be explained that the microscopic cross-bridge force ( F cross The force originates from the transverse bridge. The hierarchical integration of sarcomeres, muscle fibers, and muscle bundles, combined with physiological factors such as probability correction, length correction, and activation correction, ultimately yields macroscopic muscle force.

[0058] Macroscopic muscle strength ( F muscle The calculation requires a three-stage integration process at the microscopic (molecular scale), mesoscopic (fiber scale), and macroscopic (tissue scale) levels to progressively accumulate the single cross-bridge force into the total output force of the muscle bundle. The specific steps are as follows:

[0059] 1. Integrating forces at the microscopic level for individual sarcomeres (from cross-bridge forces to sarcomere forces).

[0060] Total number of transverse bridges in a single sarcomere: a sarcomere contains M Group of myofilaments (e.g.) M =500), each group of myofilaments has N myosin (e.g., ... N =200), each myosin contains K A cross bridge (such as) K =6), then the total number of cross bridges in a single sarcomere is N total = M × N × K =500×200×6=6×10 5 (indivual).

[0061] Crossbridge binding probability correction: Not all crossbridges bind simultaneously. Introducing a binding probability P (related to Ca2+ concentration; P=0.1 at rest, P=0.8 upon activation), the actual number of crossbridges participating in contraction is... n active = N total × P .

[0062] Single sarcomere strength ( F sarcomere Calculation: Integral of forces acting on all active crossbridges (assuming crossbridge connection angles) i It follows a normal distribution with a mean of 45°, sin i The mean value is 0.707); the formula for calculating the force of a single sarcomere is: Substitute parameters p =10 -8 N / Crossbridge, get F sarcomere =6×10 5 ×0.8×10 -8×0.707≈0.0034N.

[0063] 2. Mesoscopic integration of muscle fiber forces (from sarcomere to muscle fiber).

[0064] The sarcomere sequence of muscle fibers: A muscle fiber consists of... L A series of sarcomeres (such as) L =1000), when connected in series, the force of each sarcomere is equal (force transmission), but the total contraction is the sum of the forces of each sarcomere. Therefore, the force of a single muscle fiber is equal to the force of a single sarcomere, that is... F fiber = F sarcomere ≈0.0034N.

[0065] Muscle fiber length-force correction: When the muscle fiber length deviates from the optimal length ( L When 0), by correction coefficient f ( L Adjusting force output (based on Hill's equations):

[0066] ;

[0067] If the current length L =1.1 L 0, then F fiber,修正 =0.0034×1=0.0034N.

[0068] 3. Integrate the forces of muscle bundles at the macroscopic level (from muscle fibers to muscle bundles).

[0069] Parallel connection of muscle fibers in a muscle bundle: A muscle bundle consists of... F Root muscle fibers are composed of parallel structures (such as...) F =5000), when connected in parallel, the total force is the sum of the forces of each muscle fiber:

[0070] ;

[0071] Substituting the parameters yields F bundle =5000×0.0034=170N.

[0072] Muscle bundle activation correction: taking into account overall muscle activation levels a (0≤) a ≤1 (controlled by nerve signals), the final macroscopic muscle force is: F muscle = F bundle × a ;when a When =0.6, Fmuscle =17×0.6=10.2N.

[0073] In this optional embodiment, physiological state variables are used to dynamically adjust the nonlinear parameters of the multi-scale muscle model, including force output correction and mechanical property correction.

[0074] The formula for force output correction is:

[0075] ;

[0076] In the formula, F Indicates the current contractile force of the muscle; F bundle This indicates the force output of the muscle bundle; a Indicates activation level; v Indicates the rate of contraction; v max Indicates the maximum contraction rate;

[0077] The formula for correcting mechanical properties is as follows:

[0078] ;

[0079] In the formula, k ( l ) represents the stiffness coefficient of a nonlinear spring. k With muscle length l The changing relationship; k 0、 k 1. k 2 represents the baseline coefficient (unit: N / m), obtained experimentally, such as for fast-twitch muscle fibers. k 0 = 500 k 1 = 200 k 2 = 100; α Indicates muscle type modifier, fast-twitch muscle fibers α =1.2, slow muscle fibers α =0.8; l This represents the ratio (dimensionless) between the current muscle fiber length and the optimal length. Optimal shrinkage length L 0 refers to the length of the muscle when it generates maximum force, determined by anatomical measurements, such as the human biceps brachii. L 0 ≈ 25cm.

[0080] It should be explained that the stiffness coefficient is not only related to the current length of the muscle (via...) l The terms (first and second terms) also differ depending on muscle type (fast-twitch / slow-twitch) (via... α ); normalization processing ( l This makes the formula applicable to different muscles, enhancing the model's versatility.

[0081] In this optional embodiment, the formula for simulating muscle performance decline in a dynamically adjusted multi-scale muscle model using fatigue accumulation variables is as follows:

[0082] ;

[0083] In the formula, F fatigue This represents the output force under fatigue conditions; F Indicates the current contractile force of the muscle; K Indicates the fatigue coefficient; t Indicates the duration of muscle contraction; C This indicates the consumption rate of adenosine triphosphate.

[0084] It should be explained that a fatigue accumulation and recovery model is introduced, and the maximum force is corrected in real time using the following formula. F 0:

[0085] ;

[0086] In the formula, t The upper limit of integration represents the cumulative contraction time from the initial time 0 to the current time. The integral term calculates the total cumulative fatigue during this time period. t The integral variable (intermediate time) is used to iterate from 0 to... t All moments of contraction; F 00 This is the initial maximum force; β The fatigue coefficient; c The coefficient of recovery; t rest This is a rest period.

[0087] Combining the nonlinear parameter formula mentioned above, This can be expressed as the energy expenditure rate during muscle contraction (unit: W / kg), and the complete formula is: ;

[0088] In the formula, c Using 0 as a baseline, the starting point for energy consumption is determined: when the contraction speed... v =0 (at rest), c ( v )= c 0×1×(1+0.01×( T c -25)), that is, basal metabolic rate varies with temperature. T c The correction value.

[0089] Example: Fast-twitch muscle fibers c 0 = 0.05, contraction speedv =1.2 L 0 / s ( (for optimal length) The energy consumption rate is: c (1.2) = 0.05 × e 0.8×1.2 ×(1+0.01×(30-25))≈0.05×2.6117×1.05 ≈0.137 W / kg.

[0090] S102. Based on the muscle group synergy relationship and the kinematic model of the robot limb, and combined with the deformation compensation strategy, the muscle fiber model is geometrically adapted.

[0091] It should be explained that geometric adaptation is based on the mechanical structure of the robot's limbs (such as the range of motion of the joints). i max =120°, bone length L b =30cm), adjust the attachment point position and muscle fiber direction of the muscle fiber model to establish the conversion relationship between muscle force and joint torque:

[0092] ;

[0093] In the formula, r For muscle lever arm; α The angle between the muscle fiber and the bone is calculated using geometric adaptation.

[0094] This step maps the abstract muscle model to a specific robot structure, serving as a crucial bridge for the model's practical application.

[0095] In this optional embodiment, the geometrical adaptation of the muscle fiber model based on the muscle group synergy and the kinematic model of the robot limb, combined with a deformation compensation strategy, includes the following steps:

[0096] Based on the synergistic relationship of muscle groups, a muscle group coupling matrix is ​​constructed;

[0097] Based on the muscle group coupling matrix and the kinematic model of the robot limb, combined with the joint angle, the initial dynamic lever arm of the muscle fiber is calculated.

[0098] By using a flexible connection model of muscle attachment points, the length of muscle fibers is corrected to obtain the effective length of muscle fibers;

[0099] Based on the effective length of the muscle fiber and the muscle fiber model, the initial dynamic lever arm is geometrically corrected to obtain the final dynamic lever arm.

[0100] It should be explained that, referencing the synergistic relationships of muscle groups in human anatomy, such as the flexor muscles (biceps brachii, brachioradialis, etc.) and extensor muscles (triceps brachii, anconeus, etc.) of the upper limb employing an antagonistic arrangement, with each simulated muscle composed of 10-50 independently controlled muscle fibers. This is achieved by establishing a muscle group coupling matrix. ( n For the number of muscle groups, m (Number of joints), describing the synergistic effect weight of muscle groups on joints.

[0101] Based on the kinematic model of the robot's limbs, the lever arm of muscle fibers at the joints is calculated in real time. r Considering joint angles i Impact:

[0102] ;

[0103] In the formula, r 0、 r 1. r 2. Pre-calibration through finite element analysis; r 0 represents the basic lever arm (unit: m), which is the reference value of the muscle fiber's lever arm relative to the joint when the joint is in a neutral position (such as the elbow joint at 0° extension or the knee joint at 0° extension). It is determined by the anatomical location of the muscle attachment point. The basic lever arm of the biceps brachii at 0° of the elbow joint is... r 0 = 0.03m (3cm); r 1 represents the sinusoidal lever arm coefficient (unit: m), which is used to quantify the effect of sinusoidal changes in joint angle on the lever arm and reflects the longitudinal offset characteristics of the muscle attachment point in the joint rotation plane. r 2 represents the cosine lever arm coefficient (unit: m), used to quantify the effect of cosine changes in joint angle on the lever arm, reflecting the lateral offset characteristics of the muscle attachment point in the joint rotation plane. Both are pre-calibrated through finite element analysis (such as CT scan-based muscle-skeleton model force simulation), for example: the biceps brachii... r 1 = 0.01m, r 2 = 0.005m.

[0104] The contribution of a single muscle fiber to the torque of the joint is , t i Indicates the torque contribution of a single muscle fiber to the joint (unit: The rotational torque generated when the force produced by a single muscle fiber acts on the joint through a lever arm is the basic unit that constitutes the total torque of the joint. F i This represents the contractile force of a single muscle fiber (unit: N), i.e., the force of contraction of the first muscle fiber. i The force output by the root muscle fiber in its current activated state (calculated from a muscle fiber model, such as...) Fi =0.034N). This represents the angle (in radians or degrees) between the muscle fiber and the plane of joint rotation, reflecting the degree of deviation between the spatial orientation of the muscle fiber and the plane of joint movement. For example, muscle fibers oriented along the plane of joint rotation (such as the sagittal plane of the elbow joint). Muscle fibers that obliquely cross the plane of rotation .

[0105] Taking the calculation of the lever arm and torque of the biceps brachii during elbow flexion as an example:

[0106] Dynamic calculation of lever arm: joint angle i =90° (bending 90°), given r 0 = 0.03m r 1 = 0.01m, r 2 = 0.005m, then: .

[0107] Torque contribution of a single muscle fiber: contractile force of a single muscle fiber F i =0.034N, angle between muscle fiber and plane of rotation ,but: t i =0.034×0.04×cos(5°)≈0.034×0.04×0.996≈0.00135N·m.

[0108] lever arm r ( i The mechanism of dynamic changes in joint angle (through) r 0、 r 1、 r 2. Quantitative), rather than fixed values; the torque contribution of a single muscle fiber is not only related to force and lever arm, but also to the spatial orientation of the muscle fiber ( This approach better reflects biomechanical realities. A flexible connection model is set up at the muscle attachment point; when the force exceeds a threshold... F def Elastic deformation occurs at this time d And by correcting muscle fiber length l = l 0+ d Achieve deformation feedback.

[0109] ;

[0110] In the formula, d ( F This represents the elastic deformation at the muscle attachment point (unit: meter, m), i.e., when the muscle contraction force is... FAt that time, the deformation displacement generated by the flexible connection at the attachment point (such as the connection structure between the bionic tendon and the bone) is used to quantify the degree of elastic deformation under the action of external force. F This represents the current contractile force of the muscle (unit: Newtons, N), which is the total force output by the muscle fiber in the current active state (calculated from the muscle fiber model mentioned above, and is the input variable that triggers deformation). F def The deformation threshold force (unit: Newtons, N) is the critical force value at which a muscle attachment point transitions from a rigid state to an elastic deformation state. When the contractile force does not exceed this threshold, the attachment point is considered a rigid connection (no deformation); when it exceeds this threshold, elastic deformation occurs. For example, the attachment point threshold force of biomimetic silicone materials. F def =30N; d 0 represents the elastic deformation coefficient (unit: meters per Newton, m / N), reflecting the elastic properties of the flexible material at the attachment point, i.e., the deformation produced by a unit over-threshold force. The larger the coefficient, the easier the material is to deform (e.g., ...). d 0 = 0.0001m / N means that for every 1N exceeding the threshold, the deformation increases by 0.1mm. l 0 represents the initial muscle fiber length (unit: meter, m), that is, the original length of the muscle fiber when the muscle attachment point has not deformed, and the ratio of the current muscle fiber length to the optimal length mentioned earlier. l The reference length is consistent and is the initial value for deformation correction; D That is, the one mentioned above d ( F (elastic deformation), used here for simplified notation, and... d ( F Completely equivalent, representing the actual deformation displacement generated at the attachment point.

[0111] The formula clarifies the relationship between deformation and force, when the contraction force... F Less than or equal to the threshold F def When the attachment point is rigidly connected, the deformation is 0; when F When the threshold is exceeded, the deformation increases linearly with the force exceeding the threshold, demonstrating the elastic characteristic that the greater the force, the more significant the deformation.

[0112] The formula for correcting muscle fiber length is: l = l 0+ D ; Corrected l This feedback will be used in the force-length relationship calculation of the muscle fiber model. For example, when the muscle fiber is stretched, the length correction factor mentioned earlier will be applied. f ( L Reduce output force to simulate the force attenuation caused by deformation of the attachment point in real muscles.

[0113] Specific examples are as follows:

[0114] Known F def =30N, d 0 = 0.0001 m / N l 0 = 0.2m (initial length); when muscle contraction force F When the threshold is 50N (exceeding 20N), δ(50) = 0.0001 × (50 - 30) = 0.002m (2mm); the corrected muscle fiber length l =0.2+0.002=0.202m.

[0115] This length change will cause the subsequent force output calculation to be automatically adjusted through the force-length curve (such as the Hill equation), for example, reducing the original output force from 50N to 48N, thus realizing a biological simulation closed loop from deformation to force decay.

[0116] S103. Based on fused multi-sensor data, a feedback regulation mechanism is constructed using muscle fiber models and muscle fiber models adapted to geometric structures.

[0117] It needs to be explained that the force output based on the muscle fiber model... F Joint states after geometric adaptation (such as angles) i angular velocity i (This refers to a simulation of a biological reflex arc, such as a stretch reflex, where a joint angular velocity exceeding a threshold is detected.) i ′ th At 50° / s, the feedback signal is triggered to inhibit muscle activation. a (Reduced by 30%), output force is adjusted in reverse through a muscle fiber model to achieve rapid response control. Its feedback threshold and adjustment range are determined by the physiological characteristics of the muscle model (such as the sensitivity of stretch receptors), while the adjustment effect is reflected through the joint movement after geometric adaptation, forming a complete control closed loop.

[0118] In this optional embodiment, based on fused multi-sensor data, a feedback regulation mechanism is constructed using a muscle fiber model and a geometrically adapted muscle fiber model, comprising the following steps:

[0119] Multimodal sensor signals are acquired, and Kalman filters are used to fuse the multimodal sensor signals to obtain muscle force estimates.

[0120] Specifically, Kalman filtering is used to fuse data from different types of sensors, such as fusing strain gauge force sensor data with electromyography (EMG) sensor data to obtain a more reliable force estimate. ;

[0121] In the formula, The fused force estimate (unit: N) is the final force sensing result obtained through the fusion of multi-sensor data, which is more reliable than that of a single sensor. F sensor The measured value (unit: N) of the strain gauge force sensor directly reflects the magnitude of the contact force; F emg This represents the force estimate (in N) based on electromyography (EMG) signals, calculated through the mapping relationship between EMG signals and force. w F The force sensor weight (0-1) reflects the confidence level of the strain gauge force sensor data. A higher weight indicates more reliable sensor data (e.g., when the sensor is noise-free). w F ≈1).

[0122] Dynamic weight adjustment logic: When the strain gauge sensor noise increases (e.g., standard deviation) s When >0.5N), w F Reduce (e.g., from 0.8 to 0.3); when the EMG signal signal-to-noise ratio decreases (e.g., SNR < 20dB), w F Increase (e.g., from 0.5 to 0.9), prioritize the force sensor.

[0123] Example: If w F =0.7, F sensor =30N, F emg If the value is 28N, the fusion result is closer to the high-confidence sensor data.

[0124] Based on the comparison between the muscle force estimate and the preset safety threshold, the fast reflex regulation mechanism is triggered, and the activation decay is performed through hardware interrupt to obtain the new muscle fiber activation.

[0125] Using a geometrically adapted muscle fiber model, the rate of change of muscle fiber length is extracted, and combined with joint motion state, a mid-term regulatory mechanism is used to dynamically compensate for muscle force-velocity characteristic factors.

[0126] A long-term regulation mechanism is constructed based on a reinforcement learning framework. The weighted sum of energy consumption and motion accuracy is used as the reward function to iteratively optimize the control parameters of the fast reflex regulation mechanism and the medium-term regulation mechanism.

[0127] It needs to be explained that a three-level feedback adjustment mechanism is constructed.

[0128] 1. Rapid reflex regulation mechanism. When muscle force... F Exceeding the safety threshold Fsafe At that time, the activation level is reduced via a hardware interrupt within 10ms. a , .

[0129] In the formula, F safe The safety threshold force (unit: N) is the critical value that prevents the robot from exerting excessive force on the human body. It is set by clinical safety standards (such as upper limb rehabilitation robots). F safe =50N); a old This indicates the current activation level (0-1), which is the current contractile activation level of the muscle fiber; a new This indicates the adjusted activation level, which is the activation level that decreases after the force exceeds the safety threshold. It is used to quickly reduce muscle contraction force.

[0130] when F Exceed F safe At that time, the maximum decrease in activation is limited to 50% by the min function (to avoid force mutation).

[0131] Example: If F safe =50N, current F =60N, a old If the value is 0.8, the activation level decreases by 20%, and the force output decreases accordingly.

[0132] 2. Mid-term regulatory mechanism. Based on the rate of change in muscle length. and joint acceleration The force-velocity factor is adjusted by a PID controller. .

[0133] In the formula, It represents the rate of change of muscle fiber length (unit: m / s), that is, the speed at which the length of muscle fiber changes over time, reflecting the dynamic trend of muscle contraction or stretching (positive value is lengthening, negative value is shortening). It represents the joint angular acceleration (unit: rad / s²), which is the second derivative of the joint angle with time, reflecting the acceleration characteristics of joint movement (positive value for acceleration, negative value for deceleration). f V The base force-velocity factor (0-1) is a correction factor for the change in muscle force output with contraction velocity (based on the Hill equation). This indicates that the force-velocity factor after adjustment is dynamically adjusted by a PID controller to adapt the force output to the joint motion state. K p This represents the proportionality coefficient (unit: s / m), used to calculate the rate of change of length. The deviation can be directly adjusted; K d The differential coefficient (unit: s² / rad) is used to determine the joint angular acceleration. Suppress overshoot and enhance stability.

[0134] Example: If f V =0.9, K p =0.5s / m, , K d =0.01s² / rad, Then the force output increases as the factor increases.

[0135] 3. Long-term adjustment mechanism. The reflection parameter weights are updated through reinforcement learning, and the objective function is a weighted sum of energy consumption and motion accuracy.

[0136] Reflection parameter weights refer to key parameters in rapid reflection and intermediate conditioning (such as...) F safe , K p , K d The dynamic weighting coefficients are optimized through reinforcement learning to make the feedback mechanism more adaptable to the user's exercise habits. The objective function represents the comprehensive evaluation index of energy consumption and accuracy, and the formula is: J = l 1× E + l 2× e In the formula, E For muscle energy consumption; e For motion trajectory error; l 1. l The weight is 2, and the goal is to improve motion accuracy while reducing energy consumption.

[0137] The three-level feedback mechanism, with its hierarchical division of labor—rapid reflection (hardware-level security protection), mid-term adjustment (dynamic performance optimization), and long-term adaptation (personalized adaptation)—forms a complete human-computer interaction security control chain. Sensor fusion, through dynamic allocation of confidence weights, solves the problem of perception failure of a single sensor when there is noise or signal loss.

[0138] The precise control of a single muscle fiber is as follows:

[0139] 1. Force Generation Model: For each muscle fiber, the contractile force F it generates is determined by the activation level. a (0≤) a ≤1) Force-length factor f L Force-velocity factor fV The decision is made jointly, and the formula is as follows:

[0140] ;

[0141] In the formula, F 0 represents the maximum force of the muscle fiber under optimal length and isometric contraction conditions; l This is the ratio of the current muscle fiber length to the optimal length. v It is the ratio of the contraction rate to the maximum shortening rate.

[0142] 2. Force-length relationship: A modified quadratic function model is used to describe the force-length factor:

[0143] ;

[0144] The model is in l When the value is 1, the maximum value of 1 is reached, which is consistent with the characteristic that biological muscles generate the maximum force at their optimal length.

[0145] 3. Force-velocity relationship: Based on an improved form of Hill's equation:

[0146] ;

[0147] In the formula, b It is a constant (usually taken as 0.25), when v =0 (isometric contraction) f V =1, when v =1 (maximum shortening speed) f V =0.6.

[0148] 4. Activation control: via neural drive signals u (0-5V) Regulation of activation level a A first-order dynamic model is adopted:

[0149] ;

[0150] In the formula, t a To achieve the delayed activation characteristics of biological muscle (50-100ms), reflecting the delay from receiving nerve signals to reaching the target activation level, biological experiments have measured a range of 50-100ms (for fast-twitch muscle fibers). t a ≈50ms, slow muscle fibers t a (≈100ms) uThe neural drive signal (unit: V) simulates the control command of the central nervous system on the muscle, with a value range of 0-5V (0V corresponds to no drive, and 5V corresponds to maximum drive). The activation rate (unit: 1 / ms) is the rate at which the activation level changes over time, reflecting the dynamic trend of the activation process.

[0151] This formula describes activation. a With neural drive signals u The dynamic response process reflects the delayed activation characteristics of biological muscles, when nerve signals... u After input, activation level a It won't be reached instantly. u Instead of corresponding levels, it gradually approaches in an exponential manner (time constant). t a Determines the approach velocity); in steady state ( ), a = u / 5 (because) u The maximum is 5V, therefore a The maximum value is 1 (consistent with the definition of activation). The formula can be transformed into: (Will u Normalize to the range of 0-1, i.e. u / 5), the solution is: In the formula, a 0 represents the initial activation level.

[0152] Example: Fast-twitch muscle fibers t a =50ms, initial activation level a 0=0, input neural signal u =3V (after normalization, u / 5 = 0.6):

[0153] At t=0ms, a =0;

[0154] At t=50ms (one time constant), (Reaching 63% of steady-state value);

[0155] At t=150ms (3 time constants), a ≈0.6×(1-1 / e 3 The value is approximately 0.56 (close to the steady-state value of 0.6), reflecting the gradual activation characteristics of biological muscles.

[0156] First-order dynamic model through t a Simulating the delayed activation characteristics of biological muscles (such as the approximately 80ms delay between the force command and the maximum force output of the human biceps), it is closer to physiological reality than the instantaneous activation model, thus improving the realism of the bionic controller.

[0157] S104. Based on the adaptive neural network algorithm, the parameters of the muscle fiber model, the muscle fiber model after geometric structure adaptation, and the feedback regulation mechanism are optimized respectively, and the solution is obtained by combining the multi-objective optimization algorithm to obtain the command information of the bionic robot limb and execute it.

[0158] It should be explained that, based on information from muscle fiber models, muscle geometry, and reflex feedback mechanisms, a multi-objective optimization algorithm is used to accurately determine and generate torque or impedance commands to be sent to the bionic robot's limbs. During the bionic robot's task execution, based on the current motion target, environmental conditions, and the robot's own state, an objective function is established, comprehensively considering factors such as motion accuracy, energy consumption, and motion smoothness, to solve for the required torque or impedance value for each joint.

[0159] In this optional embodiment, based on an adaptive neural network algorithm, the parameters of the muscle fiber model, the geometrically adapted muscle fiber model, and the feedback adjustment mechanism are optimized respectively, and a multi-objective optimization algorithm is used to solve the problem to obtain the command information of the bionic robot limb and execute the following steps:

[0160] The parameters of the muscle fiber model, the geometrically adapted muscle fiber model, and the feedback regulation mechanism are optimized using an adaptive neural network algorithm.

[0161] Based on the muscle force information output by the optimized muscle fiber model, the dynamic lever arm information output by the optimized geometrically adapted muscle fiber model, and the multimodal sensor data processed by the optimized feedback adjustment mechanism, a multi-objective optimization function is constructed.

[0162] Based on the robot's physical constraints, the multi-objective optimization function is solved using a quadratic programming algorithm to obtain the optimal torque sequence;

[0163] The optimal torque sequence is converted into command information for the bionic robot limbs and executed.

[0164] It should be explained that, for the optimization of joint torque / resistance in biomimetic robots, the following multi-objective optimization function is established:

[0165] ;

[0166] In the formula, J The overall objective function (dimensionless) is a weighted comprehensive evaluation index that comprehensively measures motion accuracy, energy consumption, and motion smoothness. The goal is to minimize... J . l 1. l 2. l 3 indicates the weighting coefficient (dimensionless). l 1+ l 2+ l 3=1), used to adjust the priority of each sub-objective, example: high-precision tasks (such as assembly): l 1 = 0.6, l 2=0.2, l 3 = 0.2; Long-endurance missions (such as patrols): l 1 = 0.2, l 2 = 0.6, l 3 = 0.2; Human-computer interaction tasks (such as rehabilitation): This represents the motion accuracy sub-target (unit: rad² or m²), quantifying the deviation between the actual trajectory and the target trajectory. This represents the energy consumption sub-target (unit: J), quantifying the energy required to drive the joint. Represents the motion smoothness sub-objective (unit: or ), which quantifies the impact of the motion process.

[0167] Motion accuracy sub-target:

[0168] ;

[0169] In the formula, i d ( t The target joint angle (in rad) is represented by the desired trajectory given by the task planning module, such as... i d ( t )= i 0+ ωt Rotating at a constant speed; i ( t This represents the actual joint angle (unit: rad), which is measured in real time by the encoder; T This indicates the total task time (in seconds), such as the fetch task. T =2s.

[0170] Energy consumption sub-objective:

[0171] ;

[0172] In the formula, t ( t This indicates the joint output torque (unit: ...). ), are the optimization variables to be solved; This represents the joint angular velocity (unit: rad / s), measured by the angle differential or a velocity sensor.

[0173] Motion smoothness sub-objective:

[0174] ;

[0175] In the formula, It represents the joint angular acceleration (unit: rad / s²), which is measured by the differential of angular velocity or an acceleration sensor. The smaller the integral of the square, the smoother the motion (the smaller the impact).

[0176] The solution process, from the objective function to the joint torque / resistance, is as follows:

[0177] 1. The solution must satisfy the robot's physical constraints:

[0178] Maximum torque: | t ( t )|≤ t max (e.g., maximum torque of the servo motor) );

[0179] Speed ​​limit: (such as the maximum speed of the joint) ).

[0180] 2. Minimize the overall objective function J using the quadratic programming (QP) algorithm. The steps are as follows:

[0181] Will T Divided into N At any moment (e.g.) N =100, step size 0.02s), torque at each time step. t k ( k =1... N ) is the optimization variable. J Transform into a quadratic form In the formula, H It is a Hessian matrix; c The coefficient vector is used to solve for the optimal torque sequence under constraints. .

[0182] 3. If the impedance parameters (stiffness) need to be solved... K Damping B Then, the torque can be expressed in impedance form:

[0183] ;

[0184] After substituting into the objective function, with K , B To optimize the variables, the above quadratic programming solution process is repeated to obtain the optimal impedance parameters.

[0185] Example: Solving a single-joint grasping task.

[0186] Task parameters: Target trajectory id ( t )= πt / 2 (0→π / 2rad, completed within 2 seconds), Weight l 1 = 0.5, l 2 = 0.3, l 3 = 0.2.

[0187] Sub-objective calculation:

[0188] J 精度 : If the actual angle i ( t If the root mean square deviation from the target is 0.05 rad, then the integral result is approximately 0.005 rad. 2 ·s;

[0189] J 能耗 The average torque is 2 N·m, the average angular velocity is 0.785 rad / s, and the integral result is approximately 2 × 0.785 × 2 ≈ 3.14.

[0190] J 平滑性 Angular acceleration 0.785 rad / s² (uniform acceleration), integral result approximately 0.785 2 ×2≈1.23rad 2 / s²·s.

[0191] Overall goal and solution: J =0.5×0.005+0.3×3.14+0.2×1.23≈1.19, the optimal torque sequence is obtained by solving QP (e.g. (Smooth transition).

[0192] In this optional embodiment, the optimization of the parameters of the muscle fiber model, the geometrically adapted muscle fiber model, and the feedback regulation mechanism using an adaptive neural network algorithm includes the following steps:

[0193] Obtain standardized multi-source sensor data and combine it with real-time observation information from internal and external sensors to obtain the input vector;

[0194] By utilizing the hidden layers of an adaptive neural network model, a nonlinear feature mapping is performed on the input vector to obtain a higher-order feature vector;

[0195] Based on the high-order feature vector, the adjustment vector is obtained by linear calculation using the output layer of the adaptive neural network model.

[0196] Based on the adjustment vector, the parameters of the muscle fiber model, the geometrically adapted muscle fiber model, and the feedback regulation mechanism are corrected.

[0197] Based on the deviation between the actual motion performance and the expected target, a loss function is constructed, and the gradient is calculated through the backpropagation algorithm to iteratively update the weights and biases of the adaptive neural network model.

[0198] It should be explained that when the bionic robot needs to perform precise grasping actions, it will use an optimization algorithm to calculate the torque required for each joint of the arm based on the weight and shape of the object, as well as the position and method of grasping, with a control accuracy of ±0.5 N·m, to ensure stable grasping of the object. When the bionic robot walks on uneven ground, it will adjust the impedance of the leg joints according to the undulation of the terrain and the robot's posture through an impedance control algorithm, so that the stiffness and damping of the joints can adapt to changes in the ground in real time to maintain the body's balance and stability.

[0199] Based on sensory data from intrinsic sensors (such as high-precision absolute digital encoders with up to 16-bit resolution; Hall effect sensors with linearity error less than 0.5%; high-precision torque sensors with a measurement range of 0-50 N·m and an accuracy of ±0.2%; and six-axis inertial measurement units with an acceleration measurement range of ±16g and an angular velocity measurement range of ±2000° / s) and extrinsic sensors (such as high-definition stereo vision sensors with a resolution of 1920×1080 and a frame rate of 30fps; EEG signal sensors with a sampling rate of 250Hz and 16 channels; and surface electromyography (EMG) signal sensors with a sampling rate of 1000Hz and a common-mode inhibition ratio greater than 100dB), an adaptive neural network algorithm is used to dynamically adjust one or more parameters that determine the relationship between feedback data and muscle fiber model activation. The core function of the adaptive neural network is to dynamically adjust the mapping parameters of feedback data and muscle activation through sensor data, thereby achieving real-time adaptation of the model to the environment and task. Its structure includes an input layer (sensor data), a hidden layer (feature extraction), an output layer (parameters to be adjusted), and a parameter update mechanism. The specific process is as follows:

[0200] 1. Input layer: Multi-sensor data fusion, integrating standardized data from internal / external sensors to form an input vector x.

[0201] ;

[0202] In the formula, e θ This indicates the joint angle error (unit: rad). e θ = i d - i , id From the perspective of the target, i (From a practical perspective) e τ Indicates torque error (unit: ), e τ = t d - t , t d For the target torque, t This refers to the actual torque. This indicates the joint angular velocity (unit: rad / s). F ext This represents the external contact force (unit: N), measured by a six-axis force sensor; s emg The value represents the intensity of the electromyographic signal (dimensionless, 0-1), which is the normalized value of the surface electromyographic sensor. s eeg This indicates the intensity of the EEG signal (dimensionless, 0-1), representing the motor intention after decoding by the EEG sensor.

[0203] 2. Hidden layer: Extracts high-order features from sensor data through neural networks to reflect environmental changes (such as external force disturbances) and task requirements (such as accuracy / speed priority).

[0204] Hidden layer output (ReLU activation function): h = ReLU(W1x + b1);

[0205] In the formula, W1∈R 10×6 That is, the weight matrix from the input layer to the hidden layer (10 neurons); b1∈R 10 That is, the hidden layer bias vector.

[0206] 3. Output layer: Outputs the correction values ​​of parameters related to activation in the muscle fiber model;

[0207] Output layer output (parameter adjustment): ;

[0208] In the formula, That is, the parameter correction vector; W2∈R 3×10 That is, the weight matrix from the hidden layer to the output layer; b2∈R 3 That is, the output layer bias vector.

[0209] 4. Feedback Update: Error-Driven Learning.

[0210] The network weights are adjusted in reverse based on the actual motion error (such as trajectory deviation) to optimize the parameter mapping relationship.

[0211] The feedback data and muscle activation parameters that need to be dynamically adjusted include: k p , t a and α adj .

[0212] k p The force feedback proportionality coefficient (unit: 1 / N) determines the magnitude of the effect of muscle force changes on activation. t a It represents the activation time constant (unit: ms), which determines the delay characteristics from nerve signal to muscle activation; α adj This represents a dynamic adjustment factor for muscle type (dimensionless), which dynamically adjusts the contribution ratio of fast / slow muscle fibers based on the task (as opposed to the baseline adjustment factor). α ).

[0213] The corrected parameters are: ;

[0214] In the formula, p old This represents the parameter value at the previous moment; or This represents the learning rate (dimensionless, 0.01-0.1), controlling the adjustment range of the parameter.

[0215] Example: If .

[0216] Define the loss function (motion performance error): ;

[0217] In the formula, l e , l s Indicates the error weight (dimensionless); It represents the joint angular acceleration (unit: rad / s²), reflecting the smoothness of motion.

[0218] Update network weights using gradient descent: ;

[0219] In the formula, c This represents the network learning rate (dimensionless, 0.001-0.01).

[0220] The specific process of dynamic adjustment (taking a crawling task as an example):

[0221] Initial state: parameter k p =0.02 1 / N, t a =80ms α adj =1.0;

[0222] Sensor data input: angle error e θ =0.1rad (trajectory deviation), electromyographic signal s emg =0.8 (user intent enhancement); input vector x=[0.1,0.5,0.3,5,0.8,0.6] (other parameters omitted);

[0223] Network computation: Hidden layer features h = ReLU(W1x + b1) output feature vector;

[0224] The output layer calculates Δp=[0.005,-10,0.1] (increasing force feedback sensitivity, accelerating activation speed, and enhancing the contribution of fast-twitch muscle fibers).

[0225] Parameter update: k p,new =0.02+0.5×0.005=0.0225 1 / N; τ a,new =80 + 0.5 × (-10) = 75 ms;

[0226] Feedback: The angle error drops to 0.03 rad in the next moment, and the network updates the weights in reverse through the loss function L, continuously optimizing.

[0227] The adaptive adjustment mathematical mechanism (forward propagation calculates the correction amount, and backpropagation optimizes the mapping relationship) solves the problem of dynamic correlation between perceived data and model parameters.

[0228] Internal sensors are primarily used to monitor the motion state and mechanical parameters of the bionic robot's limbs, while external sensors are used to perceive information about the external environment and the intentions of the human user. For example, when the visual sensor detects an obstacle in front of the robot, the system adjusts the stiffness parameters of the muscle fiber model based on the obstacle's position and size, thereby changing the trajectory and speed of the bionic robot's limbs to avoid collisions. When the brain signal sensor or electromyography (EMG) sensor detects the human user's movement intentions, the system adjusts the activation parameters of the muscle fiber model accordingly, enabling the robot to follow the user's intentions with a response time of less than 100ms. Through this dynamic parameter adjustment mechanism, the bionic robot can quickly adapt to environmental changes and task requirements, achieving more intelligent and flexible control.

[0229] Figure 2 An embodiment of the biomimetic robot control system based on a muscle fiber model of the present invention is shown.

[0230] In this optional embodiment, the bionic robot control system based on a muscle fiber model includes:

[0231] The muscle fiber model construction module 201 is used to construct a multi-scale muscle model based on the scale division mechanism, and to optimize the parameters of the multi-scale muscle model using physiological state variables and fatigue accumulation variables to obtain a muscle fiber model.

[0232] The geometric structure adaptation module 202 is used to adapt the geometric structure of the muscle fiber model based on the muscle group coordination relationship and the kinematic model of the robot limb, combined with the deformation compensation strategy.

[0233] The feedback regulation mechanism construction module 203 is used to construct a feedback regulation mechanism based on fused multi-sensor data, using a muscle fiber model and a muscle fiber model adapted to geometric structure.

[0234] The command generation and execution module 204 is used to optimize the parameters of the muscle fiber model, the muscle fiber model adapted to the geometric structure, and the feedback adjustment mechanism based on the adaptive neural network algorithm, and to solve the problem by combining a multi-objective optimization algorithm to obtain and execute the command information of the bionic robot limb.

[0235] It is necessary to explain this further by taking the control of the elbow joint of a bionic robot as an example.

[0236] 1. Implementation scenarios.

[0237] Simulating the flexion / extension of the human elbow joint, the robot arm structure has a humerus length of 28cm, a radius length of 25cm, and an elbow joint range of motion of 0° (extension) to 145° (flexion). It is driven by two sets of bionic muscles (flexors: simulating the biceps brachii; extensors: simulating the triceps brachii).

[0238] 2. The process and data of constructing the muscle fiber model.

[0239] 1) Multi-scale modeling:

[0240] Molecular scale: Actin-myosin binding rate k on =0.5 / ms, dissociation rate k off =0.1 / ms, calculate the force of a single sarcomere. F sar =0.002N;

[0241] Fiber scale: Flexors contain 5000 muscle fibers (80% fast-twitch and 20% slow-twitch), and their basic strength after parallel combination is... F 屈基础 =5000×0.002×0.8=8N (fast-twitch muscle) + 5000×0.002×0.2=2N (slow-twitch muscle) =10N;

[0242] Tissue dimensions: muscle belly length 20cm, tendon length 5cm. Establish force-length curve. When the muscle belly length is 18-22cm, the force output is maintained above 90%.

[0243] 2) Dynamic adjustment of nonlinear parameters:

[0244] Activation a Signal input from nerves u (0-5V) control: a =1-e -u / 2 (like u =3V a =0.78);

[0245] contraction speed v Changes with joint movement: When the elbow joint flexion speed v j When the rate is 30° / s, calculate the muscle fiber contraction velocity. v =0.5 L 0 / s, substitute the force-velocity equation to get F 屈动态 =10×0.78×[1-(0.5 / 3) 2 =7.6N.

[0246] 3) Fatigue model integration:

[0247] After 5 seconds of sustained contraction, the ATP consumption rate C =0.02 / ms, fatigue coefficient k =0.001, calculate fatigue force. F 屈动态 =7.6×e -0.001×5000×0.02 =7.6×0.9=6.84N.

[0248] 3. Geometric structure adaptation process and data.

[0249] Determine the flexor muscle attachment points: medial condyle of the humerus (4 cm from the center of elbow joint rotation), radial tuberosity (3 cm from the center of rotation), and calculate the lever arm. r =3cm, angle between muscle fibers and bone α =15°;

[0250] Joint torque conversion: t 屈 = F 屈动态 × r ×sin( i +15°), when the elbow joint is flexed i When =90°, t 屈=6.84×0.03×sin(105°)=6.84×0.03×0.97≈0.2N·m.

[0251] 4. Implementation and data of the reflection feedback mechanism.

[0252] Stretch reflex threshold: elbow extension angular velocity (Triggered during rapid stretching);

[0253] Feedback adjustment: When detected (Exceeding the threshold by 25%), the output inhibitory signal reduces the activation of the flexor muscles. a As the force decreases from 0.78 to 0.55, the force output decreases to 10 × 0.55 × 0.98 ≈ 5.4 N, the torque decreases to 0.16 N·m, and the extension speed is reduced to 50° / s (within the threshold).

[0254] 5. Implementation results data.

[0255] Tracking accuracy: The command trajectory is 0°→90°→0° (cycle 2 seconds), and the maximum error between the actual trajectory and the command is ≤3°, which is better than traditional PID control (error ≤8°).

[0256] Fatigue simulation realism: After 30 minutes of continuous weight-bearing (5N) exercise, the output force decreased by 35%, which is consistent with the fatigue characteristics of human biceps (decreased by 30%-40%).

[0257] Reflex response speed: The time from detecting the abnormality to completing the force adjustment is ≤80ms, which is close to the latency of human muscle spindle reflex (50-100ms).

[0258] The hardware support architecture of the bionic robot controller algorithm of this invention mainly includes a sensor module, an algorithm processing module, and a control execution module. The sensor module is responsible for collecting various internal and external sensor data, employing a distributed layout and communicating with the processing module via a high-speed bus (such as EtherCAT, with a 1ms cycle). The algorithm processing module uses a multi-core heterogeneous processor (such as an ARM Cortex-A76 quad-core + FPGA) to run muscle fiber model-related algorithms, process and analyze sensor data, and generate torque or impedance commands, achieving a computational capability of up to 10 TOPS. The control execution module uses a high-precision servo driver, supporting torque control and impedance control modes, with an output current accuracy of 0.1A and a control bandwidth of 1kHz. Based on the received commands, it drives the joints and simulated muscles of the bionic robot limbs to perform corresponding actions. For example, the digital encoder in the sensor module is used to measure the rotation angle of the joint and transmit the angle information to the algorithm processing module; the processor in the algorithm processing module runs the muscle fiber model algorithm based on the received angle information and other sensor data, calculates the torque command required by the joint, and sends the command to the control execution module; the servo driver in the control execution module drives the brushless DC motor according to the torque command, which drives the joint movement of the bionic robot limb, and the motor speed control accuracy can reach ±1 rpm.

[0259] The software architecture adopts a layered design, including a bottom-level driver layer, a middleware layer, and an application layer. The bottom-level driver layer, based on a real-time operating system (such as RT-Linux, with a real-time performance of 100µs), is responsible for interacting with hardware devices, acquiring sensor data, outputting control signals, and supporting sensor calibration and fault diagnosis. The middleware layer provides a series of functional modules, such as muscle fiber model algorithm implementation (written in C++, supporting parallel computing), multi-sensor data fusion and processing (based on Kalman filtering and particle filtering algorithms), and communication management (supporting protocols such as TCP / IP and CANopen), providing a unified interface and services for the application layer. The application layer, according to different task requirements, calls the functional modules of the middleware layer to realize various application scenarios of the biomimetic robot, adopting a modular design to support functional expansion and algorithm upgrades.

[0260] The sensor module contains 20-30 sensing nodes, each node integrating 3-5 types of sensors and an independent 16-bit ADC converter. It communicates via EtherCAT bus, with the bus period dynamically configurable from 500μs to 10ms. It also has a self-calibration circuit with a calibration error of less than 0.1%.

[0261] Each sensing node adopts a modular design, with a lightweight aluminum alloy shell that provides IP65 dust and water resistance, making it suitable for complex environments. In addition to integrating conventional sensors, each node also includes a temperature compensation circuit. This circuit addresses different sensor characteristics (e.g., the temperature drift coefficient of a strain gauge is 0.01% / ℃) by real-time acquisition of ambient temperature (accuracy ±0.5℃) and inputting it into the compensation formula. ;in k T For temperature coefficient, T 0 is the calibration temperature), which controls the influence of temperature on the measurement results to within ±0.2%; F comp The measured force value (unit: N) after temperature compensation is the sensor output result after temperature error correction, reflecting the true force value and eliminating the influence of ambient temperature changes on measurement accuracy. F raw This represents the sensor's original measured force value (unit: N), which is the sensor's initial output value before temperature compensation, including errors such as temperature drift (e.g., spurious force signals generated by strain gauges due to temperature changes). k T This represents the temperature coefficient (unit: 1 / ℃), quantifying the sensitivity of sensor measurements to temperature changes, and is determined by sensor characteristics (such as the strain gauge in the text). k T =0.01% / ℃=0.0001 / ℃); T represents the current ambient temperature (unit: ℃), which is collected in real time by the temperature sensor built into the sensing node, with an accuracy of ±0.5℃; T 0 represents the calibration temperature (unit: °C), which is the reference temperature used when the sensor is manufactured or calibrated in the field (usually set to 25 °C, i.e., ambient temperature). At this temperature, the sensor's measurement value has no temperature error. When the ambient temperature... T = T At 0 o'clock, F comp = F raw No compensation is required; when T > T At 0, if the sensor has a positive temperature drift (such as a strain gauge). k T If the value is positive, the error of the original value being too high is corrected by multiplication factor; when... T < T At 0, the error of the original value being too low is corrected, and the temperature effect is ultimately controlled within ±0.2%.

[0262] Example: Strain gauge sensor k T =0.0001 / ℃, T 0 = 25℃, currently T=35℃, original measurement value F raw =100N.

[0263] Temperature deviation is T - T 0 = 10℃; the compensation factor is 1 + 0.0001 × 10 = 1.001; the force value after compensation is... F comp =100×1.001=100.1N.

[0264] Without compensation, the original value will be too large by 100 × 0.0001 × 10 = 0.1 N due to temperature drift (error 0.1%). Compensation will eliminate this error.

[0265] The EtherCAT bus employs a dual-redundancy design, with master and slave stations connected via fiber optic cables, achieving a transmission rate of 1Gbps and a maximum transmission distance of 500 meters. Bus communication utilizes distributed clock synchronization technology, achieving a synchronization accuracy of ±10ns to ensure the consistency of timestamps across sensor data. Simultaneously, it supports data compression transmission, employing differential coding compression algorithms for highly redundant sensor data (such as continuous position information), achieving compression ratios of 30%-50% and reducing bus bandwidth usage.

[0266] The sensor module uses a 24V DC power supply and is equipped with a high-efficiency DC-DC converter (conversion efficiency of over 92%), providing multiple voltage outputs such as 3.3V, 5V, and 12V for different sensors. The module has a built-in power monitoring circuit that automatically triggers a protection mechanism when the input voltage is below 20V or above 28V, cutting off the power supply to non-core sensors and sending an alarm signal to the main controller.

[0267] The algorithm processing module adopts a heterogeneous architecture, with a main processor of ARM Cortex-A76 quad-core, a coprocessor of Xilinx Zynq UltraScale+ series FPGA, an AI acceleration unit of 8TOPS computing power NPU, supporting synchronous simulation of 1000+ root muscle fibers, and built-in 2GB LPDDR4 memory and 64GB eMMC storage.

[0268] The main processor (ARM Cortex-A76) in the algorithm processing module runs a real-time operating system (RTOS), employing a scheduling strategy combining time-slice round-robin scheduling and priority preemption. Tasks are divided into core tasks (highest priority, such as control command generation), auxiliary tasks (medium priority, such as data recording), and background tasks (low priority, such as log analysis), ensuring that the response time of core tasks is less than 1ms. The FPGA's internal logic adopts a pipelined design. For sensor data preprocessing, it implements a 5-stage pipeline operation with a 20ns delay per stage, enabling filtering (using a 16th-order FIR filter) and noise reduction of a single set of data within 100ns. The NPU supports mixed-precision computing (FP16 / INT8). For matrix operations in muscle fiber models (such as 1000×1000-dimensional matrix multiplication), when using INT8 precision, the computing power can reach 16 TOPS, and the computation latency is reduced to 500μs.

[0269] In addition to built-in storage, the algorithm processing module also supports SD card expansion (maximum capacity 256GB) and adopts RAID1 mirrored storage to ensure the security of critical data (such as fault records and training data). The storage chip is equipped with a power-loss protection circuit, which can write cached data to the non-volatile storage area within 10ms when a power failure is detected, thus avoiding data loss.

[0270] A combination of a vapor chamber and heat sink fins is used for heat dissipation. The vapor chamber is 2mm thick and covers the core areas of the processor, FPGA, and NPU, with a thermal resistance of less than 0.5℃ / W. The heat sink fins have a surface area of ​​100cm², and together with a 40mm diameter silent fan (3000rpm, 10CFM airflow), the operating temperature of the core components can be kept below 70℃, ensuring stable chip performance.

[0271] The servo drive unit of the control execution module supports torque closed-loop control (bandwidth 1kHz) and position closed-loop control (bandwidth 500Hz), uses CANopen protocol for communication with a cycle of 1ms, has built-in motor parameter self-identification function with identification accuracy of ±5%, output current range of 0-10A, and efficiency ≥95%.

[0272] The drive unit of the control execution module adopts a vector control algorithm, with a current loop control cycle of 100μs, a speed loop control cycle of 500μs, and a position loop control cycle of 1ms. Through space vector pulse width modulation (SVPWM) technology, the output voltage harmonic distortion (THD) is less than 3%, reducing motor operating noise (noise level below 55dB@1m). The drive unit has a built-in motor parameter self-identification function, which automatically identifies parameters such as motor resistance, inductance, and back electromotive force constant by injecting a high-frequency signal (1kHz), with an identification accuracy of ±5%, simplifying the debugging process.

[0273] The CANopen protocol employs a layered error handling mechanism. When a communication error (such as frame loss or checksum error) is detected, it first attempts to retransmit (up to 3 times). If retransmission fails, it switches to a backup communication channel (such as RS485) to ensure reliable transmission of control commands. The motor encoder is a photoelectric absolute encoder with multi-turn counting function (maximum 4096 turns) and supports wire breakage detection. When the encoder signal is interrupted, the drive unit automatically switches to a speed estimation mode based on back EMF to maintain basic motor operation.

[0274] The drive unit features a wide voltage input design (18-36V) and is equipped with a power factor correction circuit (power factor ≥ 0.95) to improve energy efficiency. The output current range is 0-10A, and synchronous rectification technology achieves an efficiency of over 95%, reducing heat generation.

[0275] In the software implementation architecture, the RT-Linux kernel uses the PREEMPT_RT patch, reducing interrupt response latency to within 5μs. The driver employs a memory locking (mlock) mechanism to prevent critical code and data from being swapped out of physical memory, ensuring stable access times. For sensor data acquisition, DMA (Direct Memory Access) is used to avoid CPU intervention, achieving data transfer rates of up to 10MB / s.

[0276] The underlying software driver layer is based on RT-Linux, with a real-time task scheduling accuracy of 10μs. It provides a hardware abstraction interface, integrates self-calibration and fault diagnosis, and uses DMA for sensor data acquisition with a transmission rate of 10MB / s.

[0277] The HAL layer provides a unified function interface, such as sensor_read() and motor_control(), with interface parameters using standardized data structures (e.g., SensorData includes fields such as timestamp, value, and status code). It supports dynamic hardware platform adaptation, describing hardware resources through a DeviceTree, allowing adaptation to different sensor and motor models without modifying driver code.

[0278] The hardware diagnostic module employs a monitoring method combining polling and interrupts, with a polling cycle of 10ms. It can monitor parameters such as voltage (accuracy ±1%), current (accuracy ±2%), and temperature (accuracy ±1℃). When a serious fault occurs (such as a motor overcurrent of 20A), a hardware interrupt is triggered, and a safety shutdown procedure (cutting off the motor output and recording fault information) is executed within 100μs.

[0279] The muscle fiber model engine employs thread pool technology (number of threads = number of CPU cores) to achieve parallel computation of the muscle fiber model. Each muscle fiber object contains state variables (such as length, velocity, and force) and methods (such as calculate_force()). It receives sensor data through a message queue and returns the calculation results. The engine supports online tuning of model parameters. Parameters such as stiffness coefficients and damping coefficients can be modified in real time through a RESTful API interface, and parameter configuration files (JSON format) are saved for loading on the next startup.

[0280] The multi-sensor data fusion module implements the federated Kalman filter algorithm, dividing the sensor network into multiple subsystems (such as force sensor subsystems and position sensor subsystems). Each subsystem performs independent local filtering, and then the fusion center performs global optimal estimation. The module has a built-in sensor error model library, which automatically selects matching filtering parameters for different sensor types (such as the error model of a laser displacement sensor, which is normally distributed with a mean of 0 and a standard deviation of 0.01 mm). The fusion accuracy is more than 30% higher than that of a single sensor.

[0281] The control algorithm module adopts a plug-in design, supporting dynamic loading of control algorithms. In addition to conventional control strategies, it also implements an adaptive robust control algorithm, which calculates the control quantity using the following formula to address model uncertainties and external disturbances:

[0282] ;

[0283] In the formula, e is the tracking error (unit: depending on the type of controlled variable, such as joint angle error in radians and end position error in meters). e = y d - y ,in, y d For the desired output, such as the target joint angle or the target position, y This refers to the actual output, such as the joint angle or end-effector position measured by the sensor. The value represents the disturbance estimate; `sat` is the saturation function to ensure system robustness. k p This is a proportional control factor (unit: 1 / rad or 1 / m, depending on the error type), used to proportionally amplify the tracking error. e Rapidly reduce static deviation; k d These are differential control coefficients (units: s / rad or s / m), used to determine the rate of change of error. Suppress system overshoot and improve stability; To track the rate of change of error (unit: rad / s or m / s), that is, to track the first derivative of the error with respect to time, reflecting the dynamic trend of error change (such as the rate at which the error increases). These are boundary parameters for the saturation function (units consistent with the tracking error e, such as rad or m), used to set the width of the linear interval of the saturation function. hour, (Linear segment), when hour, (Saturation section, output ±1) Its function is to maintain control linearity when the error is small, and limit the increase of control quantity when the error is too large, so as to avoid system oscillation.

[0284] The communication management module adopts a publish-subscribe communication model, achieving loosely coupled communication between modules through a message middleware (such as ZeroMQ). Data transmission uses a custom protocol format, including a frame header (4 bytes), data length (2 bytes), data field (variable length), and checksum (2 bytes), supporting data fragmentation and reassembly (maximum fragment size 1024 bytes).

[0285] A finite state machine includes five states: initialization, standby, running, paused, and fault. State transition conditions are defined by logical expressions (e.g., the condition for transitioning from fault to standby is that the fault is resolved and a reset command is received). The path planning algorithm introduces an energy consumption cost function. (in P For power, (Using joint angular velocity), while ensuring the shortest path length, energy consumption is optimized, achieving energy savings of up to 15%.

[0286] 1. Initialization state.

[0287] Core behaviors: Perform hardware initialization (sensor, driver, and communication module self-test), load software parameters (control algorithm parameters, joint limit values), and calibrate the system (joint zero-position calibration, sensor offset compensation).

[0288] Status indicator: Status code = 0x00, indicator light flashes red (frequency 2Hz).

[0289] Exit conditions: All self-test items return to normal (e.g., encoder communication is normal, torque sensor zero drift <0.1N·m) and calibration error <0.5°, otherwise a fault state is triggered.

[0290] 2. Standby mode.

[0291] Core behaviors: Maintain low-power mode (driver is powered off, only the CPU and communication module work), listen for external instructions (such as startup and configuration instructions), and cache task parameters (path points, speed thresholds).

[0292] Status indicator: Status code = 0x01, indicator light is solid green.

[0293] Exit conditions: A valid start command (including task ID and verification code) is received, and the task parameters are verified (number of path points ≥ 2).

[0294] 3. Operating status.

[0295] Core behavior: Perform motion control according to the planned path (real-time calculation of joint torque), update sensor data every 10ms and monitor for anomalies (such as torque over-limit, path deviation).

[0296] Status indicator: Status code = 0x02, indicator light flashes green (frequency 5Hz).

[0297] Exit conditions: from mission completion (reaching the destination and speed <0.01rad / s) to returning to standby; from receiving a pause command to entering pause; from detecting a fault to entering fault mode.

[0298] 4. Paused state.

[0299] Core behavior: Maintain the current joint position (driver enable lock), save task breakpoint data (current path index, running time, cumulative energy consumption).

[0300] Status indicator: Status code = 0x03, indicator light is solid yellow.

[0301] Exit conditions: from receiving a continue command to resuming operation (loading breakpoint data); from receiving a reset command to clearing breakpoints and returning to standby.

[0302] 5. Fault status.

[0303] Core actions: Immediately disconnect the driver output (joint free state), record fault information (fault code, occurrence time, sensor snapshot), and report the fault through the communication module.

[0304] Status indicators: Status code = 0x04, indicator light is solid red, buzzer alarm (1Hz).

[0305] Fault types: sensor fault (fault code = 0x01), torque over-limit (fault code = 0x02), communication timeout (fault code = 0x03), etc.

[0306] Table 1 shows the logical expressions and parameter definitions for state transitions.

[0307] Table 1: Logical expressions and parameter definitions for state transitions

[0308] ;

[0309] The interface, based on the Qt framework, adopts an MVC (Model-View-Controller) architecture. The view layer includes components such as real-time curves (sampling rate 100Hz), status indicators, and parameter input boxes; the model layer stores system status data; and the controller layer processes user input and updates the model. Remote control functionality is supported, communicating with clients via the WebSocket protocol (port 8080) with a transmission latency of less than 200ms.

[0310] The reinforcement learning algorithm employs an experience replay mechanism (with a replay buffer of 1 million data points) and uses an ε-greedy strategy (ε decays linearly from 1.0 to 0.1) to balance exploration and exploitation. Offline training utilizes a distributed training framework (such as Horovod), supporting 8-GPU parallel training, which improves model convergence speed by 6 times. During online learning, an incremental learning algorithm is used, with each update of model parameters requiring less than 100KB of computation, avoiding system performance fluctuations.

[0311] System integration layer startup process: The system startup is divided into 5 stages: BIOS initialization (500ms), operating system loading (2s), driver initialization (1s), middleware loading (500ms), and application startup (1s). Each stage includes a self-test step. If the self-test fails, an alarm will be triggered by a buzzer (different frequencies of buzzer sounds correspond to different faults).

[0312] The muscle fiber model engine in the middleware layer uses thread pool parallel computing and supports GPU acceleration, with a synchronous simulation latency of less than 1ms for 1000+ muscle fibers; multi-sensor fusion uses federated Kalman filtering, which improves fusion accuracy by more than 30% compared to single sensor fusion.

[0313] The application-layer task planner introduces an energy consumption cost function. The adaptive learning module employs an experience replay mechanism (capacity of 1 million data points) and an ε-greedy strategy. Offline training supports 8-GPU parallelism, and online update computation is less than 100KB.

[0314] A memory pool is used to pre-allocate a 10MB memory block for real-time tasks, avoiding the uncertainty of dynamic memory allocation. CPU resource scheduling adopts a proportional share algorithm, allocating 60% of CPU time to core tasks (such as control algorithms) to ensure that their operation is not affected by other tasks.

[0315] The system integration layer uses memory pool technology to pre-allocate 10MB of real-time memory, allocates 60% of CPU resource scheduling time to core tasks, and supports version control and AES-256 encrypted storage for configuration management. Only authorized users can modify key parameters.

[0316] The system parameter configuration supports version control. Each modification automatically generates a version number (in the format Vx.yz, where x is the major version number, y is the feature version number, and z is the revision number), and it can be rolled back to any historical version. The configuration file is stored encrypted (AES-256 algorithm), and only authorized users (authenticated via USB key) can modify key parameters (such as the maximum torque limit).

[0317] Figure 3 This is a schematic diagram of the muscle fiber model structure of the present invention, which clearly shows the connection relationship and data flow between the muscle fiber model, muscle geometry and reflection feedback mechanism.

[0318] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0319] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0320] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0321] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0322] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0323] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A biomimetic robot control method based on a muscle fiber model, characterized in that, The method includes the following steps: Based on the three-level division of molecular, fiber and tissue scales, a multi-scale muscle model is constructed, and the parameters of the multi-scale muscle model are optimized by using physiological state variables and fatigue accumulation variables to obtain a muscle fiber model. Based on the muscle group synergy relationship and the kinematic model of the robot limb, and combined with the deformation compensation strategy, the muscle fiber model is geometrically adapted. Specifically, this includes: constructing a muscle group coupling matrix based on the muscle group synergy relationship; calculating the initial dynamic lever arm of the muscle fiber based on the muscle group coupling matrix and the kinematic model of the robot limb, combined with the joint angles; correcting the muscle fiber length using a flexible connection model of the muscle attachment points to obtain the effective length of the muscle fiber; and geometrically correcting the initial dynamic lever arm based on the effective length of the muscle fiber and the muscle fiber model to obtain the final dynamic lever arm. Based on fused multi-sensor data, a feedback regulation mechanism is constructed using a muscle fiber model and a geometrically adapted muscle fiber model. Specifically, this includes: acquiring multimodal sensor signals; fusing these signals using a Kalman filter to obtain muscle force estimates; triggering a fast reflex regulation mechanism based on a comparison of the muscle force estimates and a preset safety threshold, performing activation decay via hardware interrupt to obtain new muscle fiber activation; extracting the rate of change of muscle fiber length using the geometrically adapted muscle fiber model, and combining this with joint motion status to dynamically compensate for muscle force-velocity characteristic factors using a mid-term regulation mechanism; and constructing a long-term regulation mechanism based on a reinforcement learning framework, using a weighted sum of energy consumption and motion accuracy as the reward function to iteratively optimize the control parameters of the fast reflex regulation mechanism and the mid-term regulation mechanism. Based on the adaptive neural network algorithm, the parameters of the muscle fiber model, the muscle fiber model after geometric structure adaptation, and the feedback regulation mechanism are optimized respectively. The solution is then obtained by combining a multi-objective optimization algorithm to obtain and execute the command information of the bionic robot limb.

2. The biomimetic robot control method based on a muscle fiber model according to claim 1, characterized in that, The process of optimizing the parameters of a multi-scale muscle model using physiological state variables and fatigue accumulation variables to obtain a muscle fiber model includes the following steps: By using physiological state variables, the nonlinear parameters of the multi-scale muscle model are dynamically adjusted to obtain the dynamically adjusted multi-scale muscle model. By using fatigue cumulative variables, muscle performance decay was simulated on a dynamically adjusted multi-scale muscle model to obtain a muscle fiber model.

3. The biomimetic robot control method based on a muscle fiber model according to claim 1, characterized in that, The multi-scale muscle model includes an actin-myosin binding dynamics model, an integrated model of the series-parallel relationship of sarcomere contractile units, and a mapping model for establishing the spatial distribution of muscle bundles. The actin-myosin binding kinetic model is used to simulate actin-myosin binding kinetics at the molecular scale and calculate the contractile force of a single muscle fiber. The integrated model of series and parallel relationships of sarcomere contractile units is used to integrate the contractile force of a single muscle fiber at the fiber scale to obtain the force output of multiple muscle fibers. The established mapping model for the spatial distribution of muscle bundles is used to integrate the force output of multiple muscle fibers at the tissue scale to obtain the force output of muscle bundles.

4. The biomimetic robot control method based on a muscle fiber model according to claim 2, characterized in that, The method of using physiological state variables to dynamically adjust the nonlinear parameters of a multi-scale muscle model includes force output correction and mechanical property correction. The formula for the force output correction is as follows: ； In the formula, F Indicates the current contractile force of the muscle; F bundle This indicates the force output of the muscle bundle; a Indicates activation level; v Indicates the rate of contraction; v max Indicates the maximum contraction rate; The formula for expressing the mechanical property correction is: ； In the formula, k ( l ) represents the stiffness coefficient of a nonlinear spring. k With muscle length l The changing relationship; k 0、 k 1. k 2 represents the base coefficient; α Indicates muscle type correction factor; l This represents the ratio of the current muscle fiber length to the optimal length.

5. The biomimetic robot control method based on a muscle fiber model according to claim 2, characterized in that, The formula for simulating muscle performance decline using the fatigue cumulative variable in a dynamically adjusted multi-scale muscle model is as follows: ； In the formula, F fatigue This represents the output force under fatigue conditions; F Indicates the current contractile force of the muscle; K Indicates the fatigue coefficient; t Indicates the duration of muscle contraction; C This indicates the consumption rate of adenosine triphosphate.

6. The biomimetic robot control method based on a muscle fiber model according to claim 1, characterized in that, The adaptive neural network algorithm optimizes the parameters of the muscle fiber model, the geometrically adapted muscle fiber model, and the feedback regulation mechanism, and solves the problem using a multi-objective optimization algorithm to obtain the command information for the bionic robot limb and execute the following steps: The parameters of the muscle fiber model, the geometrically adapted muscle fiber model, and the feedback regulation mechanism are optimized using an adaptive neural network algorithm. Based on the muscle force information output by the optimized muscle fiber model, the dynamic lever arm information output by the optimized geometrically adapted muscle fiber model, and the multimodal sensor data processed by the optimized feedback adjustment mechanism, a multi-objective optimization function is constructed. Based on the robot's physical constraints, the multi-objective optimization function is solved using a quadratic programming algorithm to obtain the optimal torque sequence; The optimal torque sequence is converted into command information for the bionic robot limbs and executed.

7. The biomimetic robot control method based on a muscle fiber model according to claim 6, characterized in that, The optimization of the parameters of the muscle fiber model, the geometrically adapted muscle fiber model, and the feedback regulation mechanism using an adaptive neural network algorithm includes the following steps: Obtain standardized multi-source sensor data and combine it with real-time observation information from internal and external sensors to obtain the input vector; By utilizing the hidden layers of an adaptive neural network model, a nonlinear feature mapping is performed on the input vector to obtain a higher-order feature vector; Based on the high-order feature vector, the adjustment vector is obtained by linear calculation using the output layer of the adaptive neural network model. Based on the adjustment vector, the parameters of the muscle fiber model, the muscle fiber model after geometric adaptation, and the feedback regulation mechanism are corrected. Based on the deviation between the actual motion performance and the expected target, a loss function is constructed, and the gradient is calculated through the backpropagation algorithm to iteratively update the weights and biases of the adaptive neural network model.

8. A biomimetic robot control system based on a muscle fiber model, characterized in that, The system includes: The muscle fiber model construction module is used to construct a multi-scale muscle model based on a three-level division of molecular, fiber, and tissue scales. It also uses physiological state variables and fatigue accumulation variables to optimize the parameters of the multi-scale muscle model and obtain the muscle fiber model. The geometric structure adaptation module is used to adapt the muscle fiber model geometrically based on the muscle group synergy relationship and the kinematic model of the robot limb, combined with a deformation compensation strategy. Specifically, it includes: constructing a muscle group coupling matrix based on the muscle group synergy relationship; calculating the initial dynamic lever arm of the muscle fiber based on the muscle group coupling matrix and the kinematic model of the robot limb, combined with joint angles; correcting the muscle fiber length using a flexible connection model of muscle attachment points to obtain the effective length of the muscle fiber; and geometrically correcting the initial dynamic lever arm based on the effective length of the muscle fiber and the muscle fiber model to obtain the final dynamic lever arm. The feedback regulation mechanism construction module is used to construct a feedback regulation mechanism based on fused multi-sensor data, utilizing a muscle fiber model and a geometrically adapted muscle fiber model. Specifically, it includes: acquiring multimodal sensor signals; fusing the multimodal sensor signals using a Kalman filter to obtain muscle force estimates; triggering a fast reflex regulation mechanism based on the comparison between the muscle force estimates and a preset safety threshold, executing activation decay through hardware interrupts to obtain new muscle fiber activation; extracting the rate of change of muscle fiber length using the geometrically adapted muscle fiber model, and combining it with joint motion state to dynamically compensate for muscle force-velocity characteristic factors using a mid-term regulation mechanism; and constructing a long-term regulation mechanism based on a reinforcement learning framework, using a weighted sum of energy consumption and motion accuracy as the reward function to iteratively optimize the control parameters of the fast reflex regulation mechanism and the mid-term regulation mechanism. The command generation and execution module is used to optimize the parameters of the muscle fiber model, the geometrically adapted muscle fiber model, and the feedback regulation mechanism based on the adaptive neural network algorithm, and to solve the problem by combining a multi-objective optimization algorithm to obtain and execute the command information of the bionic robot limb.

Images