Bionic artificial muscle bundling structure and control method thereof

Abstract

The invention relates to the technical field of bionic artificial muscles, in particular to a bionic artificial muscle bundling structure and a control method thereof, and the bionic artificial muscle bundling structure comprises a silica gel supporting structure and a plurality of electric drive bionic fibers; the bionic muscle function unit is composed of a silica gel supporting structure and electric drive bionic fibers, the silica gel supporting structure provides a six-degree-of-freedom basic working space, and the electric drive bionic fibers drive the unit to move through electric control contraction. Due to complexity (strong nonlinearity and difficulty in measuring physical parameters of materials) of an execution system, a control algorithm adopts sequence generation type reinforcement learning to carry out track control and matching.

Description

Technical Field

[0001] This invention relates to the field of bionic artificial muscle technology, and in particular to a bionic artificial muscle bundle structure and its control method. Background Technology

[0002] With the deepening development of flexible robots and the emergence of electro-actuated polymer materials and smart materials, artificial muscles, as a key component of flexible robots and intelligent actuation, have become a research focus in the field of bionics. Most existing muscle components in flexible robots are rigid devices, lagging significantly behind biological muscles in terms of compliance, power density, energy efficiency, and structural-functional integration. However, electro-actuated polymer materials have become a hot research topic in recent years, capable of generating reversible bending deflection and output force under electrical excitation. Compared to traditional polymer materials, electro-actuated polymers offer advantages such as large deformation, low driving voltage, light weight, and good flexibility, demonstrating enormous potential in many engineering fields such as bionic artificial muscles, biomedicine, underwater robots, and bionic robots. However, as a driving component, bionic artificial muscles also suffer from drawbacks such as small strain, weak output force, and a single deformation direction. Furthermore, increasing functionality leads to more complex control strategies, preventing the widespread application of bionic artificial muscles in the current field of intelligent actuation materials. Therefore, developing a biomimetic artificial muscle structure with stable configuration, large response output force, and high degree of deformation freedom, along with a relatively simple electric drive control strategy, is of great value and significance for promoting the widespread application of biomimetic artificial muscle devices.

[0003] Currently, Chinese patent CN115960396B discloses a method for preparing a biomimetic artificial muscle bundle structure and an electro-drive control method. The method involves preparing a biomimetic artificial muscle electro-actuation membrane solution, an electrode membrane solution, high-twist multi-walled carbon nanotube yarn, biomimetic artificial muscle fibers, and bundling the biomimetic artificial muscle fibers to obtain hexagonal prism-shaped artificial muscle fibers. Finally, the fiber cross-section is divided into eight independently electrified regions. This invention provides a biomimetic artificial muscle bundle structure with the above-mentioned structure. Its hexagonal prism configuration is stable, has high strength, is easy to assemble, has a simple manufacturing process, rapid response, stable structure, and high degree of freedom of movement.

[0004] The aforementioned existing technical solutions have the following drawbacks: they solve the structural problems, but do not solve the problem of accuracy compensation for the control method of muscle bundles, resulting in insufficient accuracy during end-effector control. Summary of the Invention

[0005] The purpose of this invention is to provide a biomimetic artificial muscle bundle structure and its control method to solve the problems of low control accuracy and poor trajectory tracking performance of biomimetic muscles caused by strong nonlinearity of the execution system, difficulty in measuring material physical parameters, and measurement errors in the prior art.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: A biomimetic artificial muscle bundle structure includes a silicone support structure and multiple electrically driven biomimetic fibers. The silicone support structure is a cylindrical structure with a vertically aligned axis. Multiple annular grooves are formed at the top of the silicone support structure from top to bottom. A through-hole is located at the center of the top of the silicone support structure, and an annular groove is formed on the outer side of the bottom end. The silicone support structure serves to form a basic working space for providing six degrees of freedom of motion. The multiple electrically driven biomimetic fibers are attached to the silicone support structure and drive the ends of the silicone support structure to generate multi-directional motion through electrically controlled contraction. The top end of each electrically driven biomimetic fiber is fixedly connected to the top end of the silicone support structure, and a fixing block is fixedly installed at the bottom end of each electrically driven biomimetic fiber. The silicone support structure is made of silicone material and achieves complex trajectory motion through the coordinated contraction of the electrically driven biomimetic fibers. The electrically driven biomimetic fibers are made of electroactive materials and can generate axial contraction under current stimulation, thereby pulling the top end of the silicone support structure.

[0007] By adopting the above technical solution, the silicone support structure constitutes a basic workspace with six degrees of freedom, possessing flexibility and elasticity, capable of multi-directional deformation under external force. Multiple electro-driven bionic fibers are evenly aligned with the top of the silicone support structure, and their electro-controlled contraction drives the ends of the silicone support structure to generate multi-directional movement. The silicone support structure, with its flexibility and elasticity, can deform under external force and achieve complex trajectory movements through the coordinated contraction of the electro-driven bionic fibers. The integrated design of the silicone support structure and the electro-driven bionic fibers gives the entire structure the flexibility and high load-bearing capacity of bionic muscles, making it suitable for precision control scenarios. The electro-driven bionic fibers, made of electroactive materials, are attached to the silicone support structure and undergo axial contraction under electrical stimulation, thereby driving the end effector points of the silicone support structure to generate spatial movement. The silicone support structure provides the necessary flexibility and workspace, while the electro-driven bionic fibers, as actuators, achieve a contraction-driven function similar to biological muscles. This bundled structure combines the adaptability of flexible materials and the rapid response characteristics of electro-driven fibers, providing a hardware foundation for high-precision bionic motion control.

[0008] In a further embodiment, the number of the multiple electro-driven bionic fibers is three, and the three electro-driven bionic fibers are symmetrically arranged at a 120-degree angle to each other on the silicone support structure; one end of the electro-driven bionic fiber is fixed to the base of the silicone support structure 1, and the other end is connected to the end actuation point of the silicone support structure; by the independent contraction of a single electro-driven bionic fiber, the end actuation point can be driven to move in the direction of the corresponding fiber; by the simultaneous contraction of two electro-driven bionic fibers, the end actuation point can be driven to move in the area of ​​the angle between the two fibers; by the coordinated contraction of the three electro-driven bionic fibers, the end actuation point can be driven to move in the vertical direction or in the spatial range; the arrangement of the electro-driven bionic fibers optimizes the workspace coverage and improves the motion accuracy and stability.

[0009] By adopting the above technical solution, the symmetrical layout of the root fibers can realize the complex and multi-directional movement of the end execution point in space through different contraction combinations, which optimizes the coverage and movement flexibility of the workspace. The contraction combination methods include single-fiber contraction, two-fiber contraction or three-fiber coordinated contraction, etc.

[0010] This invention also discloses a method for controlling a biomimetic artificial muscle bundle structure, comprising the following steps: Step 1: Collect dynamic response data of the bionic artificial muscle bundle system through multi-spectral input signals to obtain a dataset corresponding to the current input signal and the motion position; the multi-spectral input signals include DC current signals of different frequencies and amplitudes, which are applied to one or more electrically driven bionic fibers 2 in a step-by-step manner, and three-dimensional motion position data are recorded. Step 2: Based on the dataset, a sequence-generative reinforcement learning algorithm is used for trajectory control and matching to generate a control sequence for the target pose; the reinforcement learning algorithm optimizes the control strategy through a reward function to ensure that the trajectory of the end point is consistent with the desired trajectory. Step 3: Error compensation is performed on the measurement data through repeated experiments and filtering to improve the reliability of trajectory data and control accuracy; the error compensation includes calculating the average trajectory, estimating the process noise variance, and applying Kalman filtering to enhance the data.

[0011] By adopting the above technical solution, this method combines system identification, intelligent control and error compensation, effectively overcoming the challenges brought by system nonlinearity and measurement noise, and realizing high-precision trajectory tracking in a high-uncertainty environment.

[0012] In a further embodiment, the multi-spectral input signal in step one specifically includes a DC current signal that increments at a fixed frequency. The input current of a single electrically driven bionic fiber increases in increments according to a preset step size, and the current value and the corresponding three-dimensional motion position vector are recorded simultaneously. For the joint input of multiple electrically driven bionic fibers, sinusoidal current signals of different frequencies and amplitudes are applied to the three fibers simultaneously, and the three-dimensional current vector and the three-dimensional motion position vector are recorded. The data collection process is carried out in a test scenario, capturing the motion of the end effector point through a binocular vision system and calibrating the pose using an image recognition algorithm.

[0013] By adopting the above technical solutions, multi-spectral excitation can fully stimulate the dynamic characteristics of the system, collect a more comprehensive dataset, and provide a rich information foundation for subsequent modeling and control.

[0014] In a further embodiment, the sequence-generative reinforcement learning algorithm in step two learns a control strategy by constructing a state space, an action space, and a reward function. The state space includes the current end position, the target pose, and historical trajectory information; the action space is a current control signal; and the reward function is designed based on trajectory error and energy consumption. The algorithm generates control sequences through iterative training, enabling the system to achieve high-precision trajectory tracking in complex dynamic environments. The reinforcement learning algorithm also utilizes the collected dataset for offline pre-training and online fine-tuning to adapt to real-world application scenarios.

[0015] By adopting the above technical solutions, reinforcement learning can autonomously learn complex control strategies and adapt to the nonlinear dynamic characteristics of the system, achieving high-precision control without the need for an accurate physical model.

[0016] In a further embodiment, the repeated experiments in step three refer to conducting more than ten experiments under the same current input signal and recording the time-space coordinate data of the end execution point; by calculating the average position trajectory and variance at each time point, the statistical characteristics of the system's dynamic characteristics are extracted; the error compensation also involves correcting the calibration error and material response uncertainty of the vision measurement system, including camera parameter deviation, image distortion and temperature drift compensation.

[0017] By adopting the above technical solution and extracting the common features of the system response through statistical methods, the influence of random errors in a single experiment is effectively reduced, and the representativeness of the data is improved.

[0018] In a further embodiment, the filtering process in step three includes a Kalman filter algorithm for smoothing and suppressing outliers in the average trajectory data; the Kalman filter is calculated based on the system linear model and measurement noise covariance, and reduces random errors through prediction and update steps; the filtered data serves as the true label for reinforcement learning training, improving the accuracy and robustness of the control sequence.

[0019] By adopting the above technical solutions, Kalman filtering can effectively fuse predicted and observed values, reduce random noise in the data, and further improve the quality of trajectory data used to train control models.

[0020] In a further embodiment, the control method also includes a pose construction and recognition stage in a real-world scenario; in a test scenario, markers at the end effector point are identified using a binocular camera and the YOLO algorithm, and their three-dimensional coordinates are calibrated; in an application scenario, a control sequence is directly generated based on pose requirements, without the need for real-time recognition; the pose construction and recognition process is combined with error compensation to ensure the system's stable performance in dynamic environments.

[0021] By adopting the above technical solution, a smooth transition from experimental calibration to practical application is achieved, ensuring the practicality and effectiveness of the control method at different stages.

[0022] In summary, the present invention has the following beneficial effects: 1. Through the biomimetic design of silicone support structure and electro-driven biomimetic fiber bundles, it can simulate biological muscle movement patterns, provide compliant drive and flexible movement with multiple degrees of freedom; 2. By using multi-spectral signals for system data collection and combining sequence generative reinforcement learning for control, it is possible to effectively handle system nonlinearity and achieve high-precision tracking of complex trajectories. 3. The error compensation method, which combines repeated experiments with Kalman filtering, can significantly suppress measurement noise and random errors, and improve the system control accuracy and robustness. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of a biomimetic artificial muscle bundle structure according to the present invention; In the diagram: 1. Silicone support structure; 2. Electro-driven bionic fiber. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to the accompanying drawings.

[0025] Identical parts are indicated by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to the attached figures. Figure 1In this specification, the terms "bottom surface" and "top surface," "inner" and "outer" refer to the direction toward or away from the geometry of a specific component. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this specification, "a plurality of" means two or more, unless otherwise explicitly and specifically defined by the direction of the center.

[0026] Example 1: like Figure 1 As shown, a biomimetic artificial muscle bundle structure includes a silicone support structure 1 and multiple electrically driven biomimetic fibers 2. The silicone support structure 1 forms the basic workspace for providing six degrees of freedom of motion. The multiple electrically driven biomimetic fibers 2 are attached to the silicone support structure 1 and drive the ends of the silicone support structure 1 to generate multi-directional motion through electrically controlled contraction. The silicone support structure 1 is flexible and elastic, capable of deformation under external force, and achieves complex trajectory motion through the coordinated contraction of the electrically driven biomimetic fibers 2. The electrically driven biomimetic fibers 2 are made of electroactive materials and can generate axial contraction under current stimulation, thereby pulling the silicone support structure 1. The integrated design of the silicone support structure 1 and the electrically driven biomimetic fibers 2 gives the entire structure the flexibility and high load-bearing capacity of a biomimetic muscle, suitable for… For precision control applications; the system comprises three electrically driven bionic fibers 2, symmetrically arranged at a 120-degree angle to each other on a silicone support structure 1; one end of each electrically driven bionic fiber 2 is fixed to the base of the silicone support structure 1, and the other end is connected to the end effector of the silicone support structure 1; the independent contraction of a single electrically driven bionic fiber 2 can drive the end effector to move in the direction of the corresponding fiber; the simultaneous contraction of two electrically driven bionic fibers 2 can drive the end effector to move in the area between the two fibers; the coordinated contraction of the three electrically driven bionic fibers 2 can drive the end effector to move in the vertical direction or within a spatial range; the arrangement of the electrically driven bionic fibers 2 optimizes the workspace coverage and improves motion accuracy and stability.

[0027] Specific implementation process: The silicone support structure 1 constitutes a basic workspace with six degrees of freedom, possessing flexibility and elasticity, capable of multi-directional deformation under external force. Multiple electro-driven bionic fibers 2 are evenly arranged along their axes at the connection points between their tips and the top of the silicone support structure 1. Through electrically controlled contraction, they drive the ends of the silicone support structure 1 to generate multi-directional movement. The silicone support structure 1, with its flexibility and elasticity, can deform under external force and achieve complex trajectory movements through the coordinated contraction of the electro-driven bionic fibers 2. The integrated design of the silicone support structure 1 and the electro-driven bionic fibers 2 gives the entire structure the flexibility and high load-bearing capacity of bionic muscles, making it suitable for precision control scenarios. The electro-driven bionic fibers 2 are made of electroactive materials, attached to the silicone support structure 1, and generate axial contraction under electrical stimulation, thereby driving the end actuator of the silicone support structure 1 to generate spatial movement. The silicone support structure 1 provides the necessary flexibility and workspace, while the electro-driven bionic fibers 2, as actuators, realize a contraction-driven function similar to biological muscles. This bundled structure combines the adaptability of flexible materials with the rapid response characteristics of electric drive fibers, providing a hardware foundation for high-precision bionic motion control. The ring groove is designed to enable the top of the driving silicone support structure 1 to have better deformation capability when it is pulled by the electric drive bionic fibers 2 at different positions.

[0028] Example 2: A method for controlling a biomimetic artificial muscle bundle structure includes the following steps: Step 1: Collect dynamic response data of the bionic artificial muscle bundle system using multi-spectral input signals to obtain a dataset corresponding to the current input signal and the motion position. The multi-spectral input signals include DC current signals of different frequencies and amplitudes, which are applied to one or more electrically driven bionic fibers 2 in a step-by-step manner, and the three-dimensional motion position data is recorded. Specifically, the multi-spectral input signals include DC current signals that step at a fixed frequency. The input current of a single electrically driven bionic fiber 2 increases according to a preset step size, and the current value and the corresponding three-dimensional motion position vector are recorded. For the joint input of multiple electrically driven bionic fibers 2, sinusoidal current signals of different frequencies and amplitudes are applied to three fibers simultaneously, and the three-dimensional current vector and the three-dimensional motion position vector are recorded. The data collection process is carried out in a test scenario, capturing the motion of the end effector point through a binocular vision system and calibrating the pose using an image recognition algorithm. Step 2: Based on the dataset, a sequence-generative reinforcement learning algorithm is used for trajectory control and matching to generate a control sequence for the target pose. The reinforcement learning algorithm optimizes the control strategy through a reward function to ensure that the trajectory of the end effector point is consistent with the desired trajectory. The sequence-generative reinforcement learning algorithm learns the control strategy by constructing a state space, an action space, and a reward function. The state space includes the current end effector position, the target pose, and historical trajectory information; the action space is the current control signal; and the reward function is designed based on trajectory error and energy consumption. The algorithm generates control sequences through iterative training, enabling the system to achieve high-precision trajectory tracking in complex dynamic environments. The reinforcement learning algorithm also utilizes the collected dataset for offline pre-training and online fine-tuning to adapt to real-world application scenarios. Step 3: Error compensation is performed on the measurement data through repeated experiments and filtering to improve the reliability of trajectory data and control accuracy. Error compensation includes calculating the average trajectory, estimating the process noise variance, and applying Kalman filtering to enhance the data. Repeated experiments refer to conducting more than ten experiments under the same current input signal, recording the time-space coordinate data of the end execution point. By calculating the average position trajectory and variance at each time point, statistical features of the system's dynamic characteristics are extracted. Error compensation also involves correcting the calibration error and material response uncertainty of the vision measurement system, including compensation for camera parameter deviation, image distortion, and temperature drift. Filtering includes the Kalman filtering algorithm, used to smooth the average trajectory data and suppress outliers. Kalman filtering is calculated based on the system's linear model and measurement noise covariance, reducing random errors through prediction and update steps. The filtered data serves as the true label for reinforcement learning training, improving the accuracy and robustness of the control sequence.

[0029] The control method also includes a pose construction and recognition stage in real-world scenarios; in test scenarios, markers at the end effector point are identified and their three-dimensional coordinates are calibrated using a binocular camera and the YOLO algorithm; in application scenarios, control sequences are generated directly based on pose requirements without the need for real-time recognition; the pose construction and recognition process is combined with error compensation to ensure the system's stable performance in dynamic environments.

[0030] In addition, this invention also includes the extraction of dynamic statistical features of biomimetic muscle units: to suppress random errors and improve measurement accuracy, it is necessary to conduct multiple repeated experiments (more than 10 times) under the same current input signal and record the time-space coordinate dataset of the end. (End of current signal) Spatial coordinates under action ,in It's a timestamp. (This refers to the number of trials), where: ; in and These are the maximum experimental step size and the number of trials, respectively. and These are the three-dimensional end-position coordinates and the three-dimensional current input signal, respectively. Nonlinear transformation and Koopman linear expansion are performed on the coordinates at each time point to calculate the global linear mode, average trajectory, and its variance. This yields the global linear characteristics of the current signal, the covariance of the average end-position trajectory and process noise, and the main process is as follows: 1. Calculation of average state trajectory: ; 2. Global linear mode calculation: observation function Selection: Select a polynomial function for nonlinear coordinate transformation, that is:

[0031] Where L represents the dimension increase, and the observation function g will... The 3D coordinates are mapped to the 3L-dimensional state space.

[0032] 3. Data Matrix Transformation: Construct forward / backward data matrices based on the average state trajectory. The format is as follows: ; ; 4. Global Linear Mode Calculation: For forward / backward data matrices Mean mode The calculation is as follows: ; in Represents the Pencemore generalized inverse of a matrix, and the matrix ; 5. System Reconfiguration Form: For the system's average mode, the global linear expansion form of the system is as follows: ;where the matrix ; 6. Process noise Variance estimation: .

[0033] To further improve data accuracy, Kalman filtering was used to cross-validate the experimental data under the condition of calculating the global linear expansion. The random error band of the average trajectory was reduced by calculating the filter gain to avoid the influence of extreme outliers on the average trajectory. The process is as follows: Camera measurement variance calibration: Three-dimensional coordinates of preset target points on the calibration map Measure the coordinates of the points using a camera Coordinate points are constructed after multiple samplings. -Dataset Then the camera measurement covariance value is: ; Based on the above calculations Calculate the Kalman filter gain The process is as follows: Prediction equation: ; Update equation: ; ; Covariance recurrence relation: ; Based on the above calculation process, Kalman filtering is used sequentially to enhance the data.

[0034] Ultimately, the mean-compensated trajectory data will serve as the true labels for reinforcement learning training, used for training and validating the control model. This compensation method, with a limited number of experiments, simplifies the computational complexity of filtering through recursive filtering, significantly improving the accuracy of control sequence generation and the actual motion performance of the system, making it suitable for biomimetic control tasks requiring high precision.

[0035] In the embodiments disclosed in this invention, the terms "installation," "connection," "linking," and "fixing" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; "linking" can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the embodiments disclosed in this invention according to the specific circumstances.

[0036] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A biomimetic artificial muscle bundle structure, characterized in that: The device includes a silicone support structure (1) and multiple electrically driven biomimetic fibers (2); the silicone support structure (1) is a cylindrical structure with its axis vertically aligned. Multiple annular grooves are provided at the top of the silicone support structure (1) from top to bottom. A through-hole is provided at the center of the top of the silicone support structure (1), and an annular groove is provided on the outer side of the bottom end of the silicone support structure (1). The silicone support structure (1) forms the basic working space for providing six degrees of freedom of motion. The multiple electrically driven biomimetic fibers (2) are attached to the silicone support structure (1). The end of the silicone support structure (1) is driven by the electric control contraction to generate multi-directional movement. The top of each electric drive bionic fiber (2) is fixedly connected to the top of the silicone support structure (1), and a fixing block is fixedly installed at the bottom of each electric drive bionic fiber (2). The silicone support structure (1) is made of silicone material and achieves complex trajectory movement through the coordinated contraction of the electric drive bionic fiber (2). The electric drive bionic fiber (2) is made of electroactive material and can generate axial contraction under current stimulation, thereby pulling the top of the silicone support structure (1).

2. The biomimetic artificial muscle bundle structure according to claim 1, characterized in that: The number of the electric drive bionic fibers (2) is three, and the three electric drive bionic fibers (2) are evenly arranged axially around the axis of the silicone support structure (1); one end of the electric drive bionic fiber (2) is fixed to the top of the silicone support structure (1), and the other end is connected to the end execution point of the silicone support structure (1), and there is a gap between the end execution point of the silicone support structure (1) and the body of the silicone support structure (1); by the independent contraction of a single electric drive bionic fiber (2), the end execution point can be driven to move in the direction of the corresponding fiber; by the simultaneous contraction of two electric drive bionic fibers (2), the end execution point can be driven to move in the angle region between the two fibers; by the coordinated contraction of the three electric drive bionic fibers (2), the end execution point can be driven to move in the vertical direction or in the spatial range.

3. A control method for the biomimetic artificial muscle bundle structure according to any one of claims 1-2, characterized in that, Includes the following steps: Step 1: Collect dynamic response data of the bionic artificial muscle bundle system through multi-spectral input signals to obtain a dataset corresponding to the current input signal and the motion position; the multi-spectral input signals include DC current signals of different frequencies and amplitudes, which are applied to one or more electrically driven bionic fibers (2) in a step-by-step manner, and three-dimensional motion position data are recorded. Step 2: Based on the dataset, a sequence-generative reinforcement learning algorithm is used for trajectory control and matching to generate a control sequence for the target pose; the reinforcement learning algorithm optimizes the control strategy through a reward function to ensure that the trajectory of the end point is consistent with the desired trajectory. Step 3: Error compensation is performed on the measurement data through repeated experiments and filtering to improve the reliability of trajectory data and control accuracy; the error compensation includes calculating the average trajectory, estimating the process noise variance, and applying Kalman filtering to enhance the data.

4. The control method according to claim 3, characterized in that: The multi-spectral input signal in step one specifically includes a DC current signal that steps at a fixed frequency. The input current of a single electrically driven bionic fiber (2) increases by a preset step size, and the current value and the corresponding three-dimensional motion position vector are recorded. For the joint input of multiple electrically driven bionic fibers (2), sinusoidal current signals with different frequencies and amplitudes are applied to the three fibers simultaneously, and the three-dimensional current vector and the three-dimensional motion position vector are recorded. The data collection process is carried out in the test scenario. The motion of the end effector is captured by a binocular vision system and the pose is marked by an image recognition algorithm.

5. The control method according to claim 3, characterized in that: The sequence generative reinforcement learning algorithm described in step two learns control strategies by constructing a state space, an action space, and a reward function. The state space includes the current end position, target pose, and historical trajectory information. The action space is the current control signal, and the reward function is designed based on trajectory error and energy consumption. The algorithm generates control sequences through iterative training, enabling the system to achieve high-precision trajectory tracking in complex dynamic environments. The reinforcement learning algorithm also utilizes the collected dataset for offline pre-training and online fine-tuning to adapt to real-world application scenarios.

6. The control method according to claim 3, characterized in that: The repeated experiments mentioned in step three refer to conducting more than ten experiments under the same current input signal and recording the time-space coordinate data of the end execution point; by calculating the average position trajectory and variance at each time point, the statistical characteristics of the system's dynamic characteristics are extracted; the error compensation also involves correcting the calibration error and material response uncertainty of the vision measurement system, including camera parameter deviation, image distortion and temperature drift compensation.

7. The control method according to claim 3, characterized in that: The filtering process in step three includes the Kalman filter algorithm, which is used to smooth the average trajectory data and suppress outliers. The Kalman filter is calculated based on the system linear model and the measurement noise covariance, and reduces random errors through prediction and update steps. The filtered data serves as the true label for reinforcement learning training, improving the accuracy and robustness of the control sequence.

8. The control method according to claim 3, characterized in that: The control method also includes a pose construction and recognition stage in real-world scenarios; in test scenarios, markers at the end effector point are identified using a binocular camera and the YOLO algorithm, and their three-dimensional coordinates are calibrated; in application scenarios, control sequences are generated directly based on pose requirements, without the need for real-time recognition; the pose construction and recognition process is combined with error compensation to ensure the system's stable performance in dynamic environments.

Images