A dexterous hand joint actuator control method, system, terminal and storage medium

Abstract

The application discloses a dexterous hand joint actuator control method, system, terminal and storage medium, and belongs to the technical field of intelligent control. The method is: acquiring a target instruction position and extracting a historical trajectory observation sequence, solving a second-order dynamic model to obtain an ideal output position sequence; splicing the historical trajectory observation sequence and the ideal output position sequence to obtain a physical enhanced feature vector and input a residual neural network to obtain an equivalent position deviation; determining a target position deviation physical threshold based on the dexterous hand joint actuator, performing amplitude limiting processing on the equivalent position deviation to obtain a limited deviation, and then superimposing the target instruction position to generate an equivalent target position instruction, so as to generate a joint driving torque to control the dexterous hand joint actuator. Therefore, by implementing the application, the virtual-real migration difference of the dexterous hand can be bridged from the bottom logic by fusing the physical mechanism and the residual network and introducing the hard boundary limiting mechanism of the physical limit, and the physical safety of the actuator operation is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a control method, system, terminal, and storage medium for a dexterous hand joint actuator. Background Technology

[0002] Deep reinforcement learning technology has shown significant advantages in the field of dexterous robotics, enabling complex arm-hand coordination operations by training agents in simulation environments. However, when directly deploying control strategies trained in simulations to real physical systems, a severe virtual-to-real migration discrepancy often arises. This discrepancy manifests as a deviation in dynamic response between the simulation and real physical environments, causing strategies that perform well in simulations to experience a significant performance drop or even complete failure in real-world deployments. Furthermore, the root cause of this virtual-to-real migration discrepancy lies in the fact that simulators typically apply oversimplified ideal assumptions to joint actuators, making it difficult to reproduce the complex nonlinear characteristics of real actuators. This severely restricts the efficient transfer of dexterous hands from simulation training to real-world applications. Therefore, ensuring the operational safety of dexterous hand joint actuators has become a crucial technical issue requiring further research.

[0003] Currently, existing technologies mainly employ simulation modeling methods based on purely analytical dynamic models and end-to-end neural network fitting methods based on purely data-driven approaches. The simulation modeling method based on purely analytical dynamic models typically approximates the actuator as a second-order linear system with time delay, reproducing its macroscopic dynamic response in the simulator. However, this method is limited by the continuity assumption of linear ordinary differential equations and cannot accurately characterize the non-continuous characteristics objectively present in micro-motors, such as micro-motion dead zones and nonlinear frictional creep, resulting in severe simulation distortion under low-speed or fine-tuning conditions. Secondly, the end-to-end neural network fitting method based on purely data-driven approaches directly learns the nonlinear mapping relationship between the actuator's input and output. Although neural networks have strong nonlinear fitting capabilities and can achieve high accuracy within the unloaded training data distribution, when the dexterous hand comes into rigid contact with the environment under unknown high loads, the input characteristics will deviate significantly from the unloaded training distribution, easily producing unpredictable output divergence, compromising the physical stability of the underlying control system, and posing serious safety hazards. Therefore, existing technologies cannot handle the differences between virtual and real-world transitions, resulting in poor operational safety for dexterous hand joint actuators. Summary of the Invention

[0004] This invention proposes a control method, system, terminal, and storage medium for dexterous hand joint actuators, which can solve the problem of poor operational safety of dexterous hand joint actuators caused by the inability of existing technologies to handle the differences between virtual and physical migration. This invention achieves this by integrating physical mechanisms and residual networks, and introducing a hard boundary limiting mechanism based on physical limits to overcome the differences between virtual and physical migration in the dexterous hand from the underlying logic, thereby improving the physical safety of actuator operation.

[0005] To achieve the above objectives, this invention provides a control method for a dexterous hand joint actuator, applied to a dexterous hand joint actuator. The control method includes: acquiring the target command position received in the current control step; extracting a historical trajectory observation sequence based on a preset historical time window; solving a pre-constructed second-order dynamic model based on the target command position to obtain an ideal output position sequence; concatenating the historical trajectory observation sequence and the ideal output position sequence to obtain a physical enhancement feature vector; inputting the physical enhancement feature vector into a pre-trained residual neural network to obtain an equivalent position deviation; determining a physical threshold for the target position deviation based on the dexterous hand joint actuator; and performing amplitude limiting processing on the equivalent position deviation according to the physical threshold to obtain an amplitude limiting deviation; superimposing the target command position and the amplitude limiting deviation to generate an equivalent target position command; inputting the equivalent target position command into a preset proportional-derivative controller to generate a joint driving torque, and controlling the dexterous hand joint actuator through the joint driving torque.

[0006] This invention proposes a control method for a dexterous hand joint actuator. The dynamic response of the dexterous hand joint actuator is decoupled into two parts: an ideal second-order dynamic response and residual network nonlinear compensation. An ideal physical model serves as the framework to ensure macroscopic stability. A residual neural network is used to predict equivalent position deviations to compensate for nonlinear disturbances such as dead zones and static friction within the actual dexterous hand joint actuator. Simultaneously, a hard boundary limiting mechanism based on actual physical limits is introduced at the output end. This allows the network output within the limiting range to be converted into feedforward compensation torque under no-load micro-motion conditions, accurately reproducing the low-speed stalling and dead zone phenomena of the actual actuator, achieving high-fidelity micro-motion. Furthermore, under loaded rigid contact conditions, the feedforward compensation is forcibly truncated, naturally degenerating into robust proportional-derivative control, reducing safety hazards caused by strategy divergence, and achieving high safety in rigid contact. Thus, by integrating physical mechanisms and residual networks and introducing a hard boundary limiting mechanism based on physical limits, the physical safety of actuator operation is improved from the underlying logic level, overcoming the differences between the virtual and real migration of the dexterous hand.

[0007] Furthermore, the step of obtaining the target command position received in the current control step, extracting the historical trajectory observation sequence based on a preset historical time window, and solving the pre-constructed second-order dynamic model based on the target command position to obtain the ideal output position sequence includes: extracting historical target commands of several historical control steps within the preset historical time window according to the target command position to form a historical command sequence, and extracting the joint actual positions of several historical control steps based on the historical command sequence to form a historical actual position sequence; performing backward difference on the historical actual position sequence to obtain a historical velocity sequence; using the historical command sequence, the historical actual position sequence, and the historical velocity sequence as the historical trajectory observation sequence; constructing a second-order dynamic model based on a preset steady-state gain, a preset damping ratio, a preset natural frequency, and the target command position, and sequentially calculating the corresponding ideal output position according to each historical target command in the historical command sequence to form the ideal output position sequence.

[0008] In the above scheme, real physical parameters are obtained to construct a second-order dynamic model, so that the ideal response is highly consistent with the macroscopic motion characteristics of the real dexterous hand joint actuator, providing an accurate physical benchmark for subsequent residual compensation. At the same time, historical velocity sequences are extracted from the actual position sequence through backward difference, providing explicit differential damping information for subsequent feature vector construction, which helps to suppress transient overshoot.

[0009] Furthermore, concatenating the historical trajectory observation sequence and the ideal output position sequence to obtain a physical enhancement feature vector includes: obtaining a historical command sequence, a historical actual position sequence, and a historical velocity sequence from the historical trajectory observation sequence; subtracting the historical command sequence from the historical actual position sequence to obtain a historical tracking error sequence; standardizing the historical command sequence, the historical actual position sequence, the ideal output position sequence, the historical tracking error sequence, and the historical velocity sequence respectively, and concatenating them in a preset order to obtain a physical enhancement feature vector.

[0010] In the above scheme, the historical command sequence represents the system's motion intention, the historical actual position sequence represents the current physical state, the ideal output position sequence is injected as a priori physical law to guide the network to overcome dry friction, the historical tracking error sequence provides a physical priori anchor point similar to position closed-loop control, and the historical velocity sequence explicitly provides damping boundary constraints. The physical enhancement feature vector obtained by splicing these is then standardized to eliminate the dimensional differences between different features, so that the subsequent residual neural network does not need to implicitly fit complex physical laws from scratch, thus accelerating the network training convergence process and improving the response capability under multi-frequency domain signals.

[0011] Furthermore, in the process of inputting the physical enhancement feature vector into a pre-trained residual neural network to obtain the equivalent positional deviation, the pre-trained residual neural network includes: an input layer, at least two hidden layers, and an output layer; wherein: the input layer is used to receive the physical enhancement feature vector; the hidden layer is used to perform nonlinear mapping on the physical enhancement feature vector to obtain a physical enhancement feature vector mapping value; and the output layer is used to output the physical enhancement feature vector mapping value as the equivalent positional deviation.

[0012] In the above scheme, the residual neural network adopts a multilayer perceptron structure and is configured with multiple hidden layers. It can automatically extract abstract feature representations of nonlinear disturbances such as gear backlash and static friction from high-dimensional physical enhancement features. The equivalent position deviation output by the network quantitatively represents the angular lag caused by unmodeled nonlinear factors in a physical sense. It makes up for the deficiency of traditional linear dynamic models in representing discontinuous features in numerical form.

[0013] Furthermore, based on the dexterous hand joint actuator, a target position deviation physical threshold is determined, and the equivalent position deviation is limited according to the target position deviation physical threshold to obtain the limited deviation. This includes: obtaining the position hysteresis of the output end of the dexterous hand joint actuator from rest to motion when the actuator is in an unloaded or rotor-locked state; determining the position hysteresis that satisfies a preset hysteresis threshold as the target position deviation physical threshold; if the absolute value of the equivalent position deviation is greater than the target position deviation physical threshold, the equivalent position deviation is truncated to a symmetrical interval defined by the target position deviation physical threshold to obtain the limited deviation; if the absolute value of the equivalent position deviation is less than or equal to the target position deviation physical threshold, the equivalent position deviation is directly used as the limited deviation.

[0014] In the above scheme, a hard boundary limiting mechanism based on physical limits is introduced to ensure that the compensation amount output by the residual neural network is always constrained within the maximum deviation range that the real physical system can produce. Under no-load micro-motion conditions, the tracking error is small, and the limiting deviation of the network output is converted into feedforward compensation torque to accurately compensate for static friction and dead zone, achieving high fidelity micro-motion. When the dexterous hand comes into rigid contact with the environment, causing the tracking error to expand sharply, the network input features deviate significantly from the no-load training distribution, and the unconstrained network may output abnormal deviations. However, since the limiting mechanism forcibly cuts off the deviation within the physical threshold, the compensation torque generated is strictly limited to a very small range. The control is smoothly transferred to the basic impedance term of the proportional-derivative controller, reducing the safety hazards caused by strategy divergence from the underlying logic and ensuring the physical safety of the dexterous hand joint actuator operation.

[0015] Furthermore, the equivalent target position command is input into a preset proportional-differential controller to generate a joint driving torque, and the dexterous hand joint actuator is controlled by the joint driving torque, including: acquiring the proportional gain and differential gain of the preset proportional-differential controller; acquiring the target command position and the actual position command of the current control step; subtracting the equivalent target position command from the target command position to obtain the position deviation; producting the proportional gain with the position deviation to obtain the proportional torque component; producting the differential gain with the derivative of the actual position command to obtain the differential torque component; subtracting the proportional torque component from the differential torque component to obtain the joint driving torque; and applying the joint driving torque to the dexterous hand joint actuator to control the movement of the dexterous hand joint actuator.

[0016] In the above scheme, the equivalent target position command after the amplitude limiting deviation is superimposed is used as the input of the proportional-derivative controller to achieve decoupled distribution of the underlying torque. The product of the proportional gain and the position deviation constitutes the basic impedance torque, while the amplitude limiting deviation is converted into a feedforward compensation torque through the proportional gain. This torque indirectly intervenes in the physical engine through the stiffness coefficient by superimposing position dimensions. The product of the derivative gain and the actual position command provides the damping torque to suppress system oscillation. Thus, the introduced torque distribution mechanism accurately overcomes static friction and dead zone under no-load micro-motion conditions, mainly through the feedforward compensation torque. Under loaded rigid contact conditions, it ensures control stability mainly through the basic impedance torque and damping torque, achieving adaptive and smooth switching under different operating conditions, which helps to improve the physical safety of the dexterous hand joint actuator.

[0017] This invention also provides a dexterous hand joint actuator control system, applied to a dexterous hand joint actuator. The dexterous hand joint actuator control system includes a sequence acquisition module, a feature vector acquisition module, a position deviation calculation module, a limit deviation calculation module, a target position command generation module, and an actuator control module. Specifically: the sequence acquisition module acquires the target command position received in the current control step, extracts historical trajectory observation sequences based on a preset historical time window, and solves a pre-constructed second-order dynamic model based on the target command position to obtain an ideal output position sequence; the feature vector acquisition module concatenates the historical trajectory observation sequence and the ideal output position sequence to obtain a physical... The system includes several modules: an enhanced feature vector, a position deviation calculation module (which inputs the enhanced physical feature vector into a pre-trained residual neural network to obtain an equivalent position deviation), a limiting deviation calculation module (which determines a target position deviation physical threshold based on the dexterous hand joint actuator and limits the equivalent position deviation according to the target position deviation physical threshold to obtain a limiting deviation), a target position command generation module (which superimposes the target command position with the limiting deviation to generate an equivalent target position command), and an actuator control module (which inputs the equivalent target position command into a preset proportional-derivative controller to generate a joint driving torque and controls the dexterous hand joint actuator through the joint driving torque).

[0018] This invention proposes a control system for a dexterous hand joint actuator. The dynamic response of the dexterous hand joint actuator is decoupled into an ideal second-order dynamic response and a residual network nonlinear compensation. An ideal physical model serves as the framework to ensure macroscopic stability. A residual neural network is used to predict equivalent position deviations to compensate for nonlinear disturbances such as dead zones and static friction within the actual dexterous hand joint actuator. Simultaneously, a hard boundary limiting mechanism based on real physical limits is introduced at the output. This allows the network output within the limiting range to be converted into feedforward compensation torque under no-load micro-motion conditions, accurately reproducing the low-speed stalling and dead zone phenomena of the actual actuator, achieving high-fidelity micro-motion. Furthermore, under loaded rigid contact conditions, the feedforward compensation is forcibly truncated, naturally degenerating into robust proportional-derivative control, reducing safety hazards caused by strategy divergence and achieving high safety in rigid contact. Thus, by integrating physical mechanisms and residual networks and introducing a hard boundary limiting mechanism based on physical limits, the physical safety of actuator operation is improved from the underlying logic level, overcoming the differences between the virtual and real migration of the dexterous hand.

[0019] Furthermore, the amplitude limiting deviation calculation module includes a deviation physical threshold acquisition unit, a first threshold judgment unit, and a second threshold judgment unit. The deviation physical threshold acquisition unit is used to acquire the position lag amount when the output end of the dexterous hand joint actuator moves from rest to motion under no-load or rotor-locked conditions, and determines the position lag amount that satisfies a preset lag threshold as the target position deviation physical threshold. The first threshold judgment unit is used to truncate the equivalent position deviation to a symmetrical interval defined by the target position deviation physical threshold if the absolute value of the equivalent position deviation is greater than the target position deviation physical threshold, thereby obtaining the amplitude limiting deviation. The second threshold judgment unit is used to directly use the equivalent position deviation as the amplitude limiting deviation if the absolute value of the equivalent position deviation is less than or equal to the target position deviation physical threshold.

[0020] This invention also provides a dexterous hand joint actuator control terminal, the terminal including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the dexterous hand joint actuator control method when executing the computer program.

[0021] This invention also provides a dexterous hand joint actuator control storage medium, the storage medium including a stored computer program, wherein the computer program, when running, controls the device where the storage medium is located to execute the dexterous hand joint actuator control method. Attached Figure Description

[0022] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating the steps of a dexterous hand joint actuator control method according to a certain embodiment of the present invention; Figure 2 This is a schematic diagram of the module structure of a dexterous hand joint actuator control system provided in a certain embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0026] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0028] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0029] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0030] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0031] To address the issue of poor operational safety of dexterous hand joint actuators caused by the inability of existing technologies to handle differences between virtual and physical transfer, see [reference needed]. Figure 1 , Figure 1 This is a flowchart illustrating the steps of a dexterous hand joint actuator control method according to a certain embodiment of the present invention. Figure 1 As shown in the figure, this invention proposes a control method for a dexterous hand joint actuator, which is applied to a dexterous hand joint actuator. The control method for the dexterous hand joint actuator includes steps 101 to 106, and the specific steps are as follows: Step 101: Obtain the target command position received in the current control step, extract the historical trajectory observation sequence based on the preset historical time window, and solve the pre-constructed second-order dynamic model based on the target command position to obtain the ideal output position sequence; Step 102: Concatenate the historical trajectory observation sequence and the ideal output position sequence to obtain the physical enhancement feature vector; Step 103: Input the physical enhancement feature vector into the pre-trained residual neural network to obtain the equivalent positional deviation; Step 104: Determine the physical threshold of the target position deviation based on the dexterous hand joint actuator, and perform amplitude limiting processing on the equivalent position deviation according to the physical threshold of the target position deviation to obtain the amplitude limiting deviation; Step 105: Superimpose the target command position with the amplitude limiting deviation to generate an equivalent target position command; Step 106: Input the equivalent target position command into a preset proportional differential controller to generate joint driving torque, and control the dexterous hand joint actuator through the joint driving torque.

[0032] One possible implementation method is to first obtain the target command position received in the current control step, denoted as... The target command location is usually issued by the upper-level operation strategy or remote operation device. Then, using a preset historical time window as the span, with a historical time window length of H, historical control commands and historical actual position commands from multiple consecutive control steps in the past are extracted from the memory buffer to form a historical trajectory observation sequence. Specifically, the historical trajectory observation sequence includes a historical target command sequence. Historical actual position sequence obtained by joint sensors and analog-to-digital conversion And the historical velocity sequence obtained by backward difference approximation based on the actual position sequence. At the same time, based on the target command position received in the current control step. The system calls a pre-constructed second-order dynamic model to calculate the corresponding ideal response. This second-order dynamic model is a continuous-time dynamic system characterizing the linear macroscopic dynamics of a dexterous hand joint actuator. Its model parameters include steady-state gain, damping ratio, and natural frequency. These parameters are obtained by inputting step and frequency sweep excitation signals to the real physical actuator beforehand and fitting the collected input-output data using a system identification algorithm. This process is achievable with existing technology. Finally, the system analyzes historical command sequences. Each historical target instruction is used to calculate the corresponding ideal output position based on the identified dynamic parameters, ultimately forming a complete sequence of ideal output positions. .

[0033] Then, the historical trajectory observation sequence is compared with the ideal output position sequence. Feature fusion and splicing are performed. Specifically, historical command sequences are retrieved from historical trajectory observation sequences. and historical actual location sequence Perform element-wise difference operations to generate a historical tracking error sequence representing the historical following deviation. Subsequently, the historical instruction sequence will be... Historical actual location sequence Ideal output position sequence Historical tracking error sequence and historical velocity sequences These five sets of time-series data with clear physical meanings are Z-score standardized to eliminate numerical biases caused by different physical dimensions such as current, angle, and angular velocity during network training. After standardization, all five sets of sequences are concatenated in a predetermined fixed order to form a high-dimensional physical enhancement feature vector, denoted as [vector name missing]. This serves as the input for subsequent neural networks.

[0034] Then, the concatenated physical enhancement feature vector Input is fed into a pre-trained residual neural network. Forward inference is performed. In this embodiment, the residual network can be implemented using a multilayer perceptron structure. Its input layer receives the number of neurons matching the dimension of the feature vector. In this embodiment, the dimension is 5. It contains at least two fully connected hidden layers, and the hidden layers use the hyperbolic tangent function as the activation function. Then, in each nonlinear mapping layer, the network extracts abstract features of nonlinear factors that are difficult to model explicitly, such as gear clearance and static friction. Finally, the output layer, containing a single neuron, outputs a continuous scalar value, denoted as . In this embodiment, the scalar value output by the above process In terms of physical concepts, it quantitatively describes the equivalent position deviation caused by unmodeled factors at the current moment.

[0035] Furthermore, a target position deviation physical threshold is determined based on the actual physical characteristics of the dexterous hand joint actuator. In this embodiment, the target position deviation physical threshold is obtained as follows: with the actual actuator in a completely unloaded state or with the rotor locked, small rotation commands are gradually sent to the dexterous hand joint actuator starting from zero position. Simultaneously, the actual angular displacement of the output end is monitored through a high-precision encoder. When the output end just begins to move macroscopically from a stationary state, the hysteresis angle difference between the command position and the sensor feedback position is recorded, and the maximum hysteresis is determined as the position deviation physical threshold. The position deviation physical threshold essentially corresponds to the maximum dead zone range generated by the superposition of gear backlash and static friction inside the dexterous hand joint actuator. After obtaining the equivalent position deviation and the target physical threshold, the equivalent position deviation output by the network is hard-limited. If the absolute value of the equivalent position deviation exceeds the threshold boundary, it is forcibly truncated to within the symmetrical threshold range to obtain the limited deviation. If it does not exceed the threshold, the original output value is retained and directly transmitted.

[0036] Then, the target instruction position received in the current control cycle is... The position command is calculated by adding the position measurement to the amplitude limiting deviation, and then calculating an equivalent target position command after nonlinear compensation correction. Therefore, the dead zone crossing information learned by the network is incorporated into the position command channel in a feedforward manner. When mechanical backlash causes the actuator to stall, it provides an additional displacement increment for the underlying control of that cycle to quickly cross the backlash zone.

[0037] Finally, the equivalent target position command The data is fed into a preset proportional-derivative controller, and the actual position and velocity of the joint are obtained. Specifically, the preset proportional gain and derivative gain of the proportional-derivative controller are first obtained. The proportional gain is determined based on the product of the joint's equivalent moment of inertia and the square of its natural frequency, while the derivative gain is determined based on the product of the joint's equivalent moment of inertia, twice the damping ratio, and the natural frequency. Then, the controller sends the equivalent target position command... The actual joint position of the current control step The position deviation is obtained by subtraction. The position deviation is multiplied by the proportional gain to obtain the proportional torque component. At the same time, the actual joint speed is multiplied by the differential gain to obtain the velocity damping torque component. Finally, the proportional torque component is subtracted from the velocity damping torque component to obtain the joint driving torque for this control cycle. This torque command is then sent to the motor drive unit to drive the dexterous hand joint actuator to complete one control cycle.

[0038] This invention proposes a control method for a dexterous hand joint actuator. The dynamic response of the dexterous hand joint actuator is decoupled into two parts: an ideal second-order dynamic response and residual network nonlinear compensation. An ideal physical model serves as the framework to ensure macroscopic stability. A residual neural network is used to predict equivalent position deviations to compensate for nonlinear disturbances such as dead zones and static friction within the actual dexterous hand joint actuator. Simultaneously, a hard boundary limiting mechanism based on actual physical limits is introduced at the output end. This allows the network output within the limiting range to be converted into feedforward compensation torque under no-load micro-motion conditions, accurately reproducing the low-speed stalling and dead zone phenomena of the actual actuator, achieving high-fidelity micro-motion. Furthermore, under loaded rigid contact conditions, the feedforward compensation is forcibly truncated, naturally degenerating into robust proportional-derivative control, reducing safety hazards caused by strategy divergence, and achieving high safety in rigid contact. Thus, by integrating physical mechanisms and residual networks and introducing a hard boundary limiting mechanism based on physical limits, the physical safety of actuator operation is improved from the underlying logic level, overcoming the differences between the virtual and real migration of the dexterous hand.

[0039] In a preferred embodiment, the step of obtaining the target command position received in the current control step, extracting a historical trajectory observation sequence based on a preset historical time window, and solving a pre-constructed second-order dynamic model based on the target command position to obtain an ideal output position sequence includes: extracting historical target commands of several historical control steps within a preset historical time window based on the target command position to form a historical command sequence, and extracting the joint actual positions of several historical control steps based on the historical command sequence to form a historical actual position sequence; performing backward differencing on the historical actual position sequence to obtain a historical velocity sequence; using the historical command sequence, the historical actual position sequence, and the historical velocity sequence as the historical trajectory observation sequence; constructing a second-order dynamic model based on a preset steady-state gain, a preset damping ratio, a preset natural frequency, and the target command position, and sequentially calculating the corresponding ideal output position for each historical target command in the historical command sequence to form the ideal output position sequence.

[0040] For example, at the beginning of each discrete control step, the current target command position is received from the upper controller. Then, a historical data extraction process is initiated. Specifically, a sliding time window of fixed length H is maintained, which continuously records the target commands and actual joint feedback positions of all control steps over a period of time. The historical target command arrays and actual joint feedback position arrays corresponding to several historical control steps within the window are extracted sequentially. The extracted historical target command arrays are then arranged according to the acquisition time sequence to form a historical command sequence. The extracted joint actual feedback position arrays are arranged according to the same time base to form a historical actual position sequence. Next, the obtained historical actual location sequence was analyzed. Perform backward differencing, which involves calculating the difference between the actual position values ​​of every two adjacent time points in the sequence and dividing it by the control period duration to generate a point-by-point historical velocity sequence. Historical speed sequence The instantaneous motion trend and differential damping information are explicitly carried. Subsequently, the above three sets of sequences are packaged together to form a historical trajectory observation sequence.

[0041] In the process of constructing the ideal output position sequence, a set of physical parameters, namely the steady-state gain, which were previously identified through an offline system, are first invoked. Equivalent damping ratio and Undamped natural frequency. The specific method for obtaining these parameters is as follows: Excitation command signals containing step changes, sinusoidal sweeps, and triangular wave waveforms are sequentially input into the actual dexterous hand joint actuator. The actual angular response at the actuator output is synchronously recorded at a preset sampling frequency. The input-output dataset is fitted and solved using system identification algorithms such as least squares to obtain the equivalent second-order dynamic parameters of this actuator model. Then, based on the steady-state gain... Equivalent damping ratio and A second-order dynamic model describing the macroscopic linear dynamics of the actuator is established using parameters such as the undamped natural frequency, specifically expressed as follows: ; In the formula, Output the position of the ideal physical model for the current control step; Then, within each control cycle, the historical instruction sequence is traversed. Each historical target instruction stored in the system is used as the instantaneous input to the second-order dynamic model, combined with steady-state gain. Equivalent damping ratio and The time response relationship of the undamped natural frequency is determined by sequentially solving the ideal response angle value corresponding to each historical control step, thus forming a sequence of historical commands point by point. One-to-one ideal output position sequence In this embodiment, the ideal output position sequence This reflects the theoretical motion trajectory that the actuator should exhibit under the assumptions of zero backlash and zero friction.

[0042] In the above scheme, real physical parameters are obtained to construct a second-order dynamic model, so that the ideal response is highly consistent with the macroscopic motion characteristics of the real dexterous hand joint actuator, providing an accurate physical benchmark for subsequent residual compensation. At the same time, historical velocity sequences are extracted from the actual position sequence through backward difference, providing explicit differential damping information for subsequent feature vector construction, which helps to suppress transient overshoot.

[0043] A preferred embodiment involves concatenating the historical trajectory observation sequence and the ideal output position sequence to obtain a physical enhancement feature vector. This includes: obtaining a historical command sequence, a historical actual position sequence, and a historical velocity sequence from the historical trajectory observation sequence; subtracting the historical command sequence from the historical actual position sequence to obtain a historical tracking error sequence; standardizing the historical command sequence, the historical actual position sequence, the ideal output position sequence, the historical tracking error sequence, and the historical velocity sequence respectively, and concatenating them in a preset order to obtain the physical enhancement feature vector.

[0044] For example, the historical trajectory observation sequence is compared with the ideal output position sequence. Feature fusion and splicing are performed. Specifically, historical command sequences are retrieved from historical trajectory observation sequences. and historical actual location sequence Perform element-wise difference operations to generate a historical tracking error sequence representing the historical following deviation. Subsequently, the historical instruction sequence will be... Historical actual location sequence Ideal output position sequence Historical tracking error sequence and historical velocity sequences These five sets of time-series data with clear physical meanings are Z-score standardized to eliminate numerical biases caused by different physical dimensions such as current, angle, and angular velocity during network training. After standardization, all five sets of sequences are concatenated in a predetermined fixed order to form a high-dimensional physical enhancement feature vector, denoted as [vector name missing]. This serves as the input to the subsequent neural network. The historical instruction sequence is one such input. Characterizing the macroscopic motion intention of a dexterous hand joint actuator; historical actual position sequence Characterizing the actual physical state of a dexterous hand joint actuator; ideal output position sequence Characterization incorporates physical laws as prior observations to guide the network in overcoming dry friction; historical tracking error sequences. Characterized by physical prior anchors similar to position closed-loop proportional control; historical velocity sequence The characterization explicitly provides differential damping information, suppressing transient overshoot.

[0045] In the above scheme, the historical command sequence represents the system's motion intention, the historical actual position sequence represents the current physical state, the ideal output position sequence is injected as a priori physical law to guide the network to overcome dry friction, the historical tracking error sequence provides a physical priori anchor point similar to position closed-loop control, and the historical velocity sequence explicitly provides damping boundary constraints. The physical enhancement feature vector obtained by splicing these is then standardized to eliminate the dimensional differences between different features, so that the subsequent residual neural network does not need to implicitly fit complex physical laws from scratch, thus accelerating the network training convergence process and improving the response capability under multi-frequency domain signals.

[0046] In a preferred embodiment, in the process of inputting the physical enhancement feature vector into a pre-trained residual neural network to obtain an equivalent positional deviation, the pre-trained residual neural network includes: an input layer, at least two hidden layers, and an output layer; wherein: the input layer is used to receive the physical enhancement feature vector; the hidden layers are used to perform nonlinear mapping on the physical enhancement feature vector to obtain a physical enhancement feature vector mapping value; and the output layer is used to output the physical enhancement feature vector mapping value as the equivalent positional deviation.

[0047] For example, in this embodiment, the residual network can be implemented using a multilayer perceptron structure. Its input layer receives the number of neurons matching the dimension of the feature vector; in this embodiment, the dimension is 5. It contains at least two fully connected hidden layers, with the hyperbolic tangent function used as the activation function. Then, in each nonlinear mapping layer, the network abstracts features from nonlinear factors that are difficult to model explicitly, such as gear backlash and static friction. Finally, the output layer, containing a single neuron, outputs a continuous scalar value, denoted as […]. .

[0048] Specifically, the input layer first receives the 5H-dimensional physical enhancement feature vector concatenated during the current control cycle. Subsequently, the physical enhancement feature vector The data is passed layer by layer to the hidden layers. In this embodiment, there are at least two hidden layers, each containing multiple neurons, for example, 64 neurons per layer. Neurons between adjacent layers are fully connected, and the weights and biases of each connection are determined during training. Within each hidden layer, the data from the previous layer is first multiplied by the current layer's weight matrix and then linearly combined with the bias vector. The resulting linear transformation is then fed into the hyperbolic tangent activation function for nonlinear transformation, generating the nonlinear mapping output of that layer, which serves as the input to the next layer. The hyperbolic tangent activation function, due to its symmetric and continuously differentiable output range about the origin, effectively characterizes the smooth and bounded physical properties of friction and dead zones within a real actuator. Finally, after the nonlinear mapping of a hidden layer, the data is passed to the output layer, which contains only a single neuron. The received feature representation is linearly weighted and summed, directly outputting a continuous scalar value. It is the physical enhancement feature vector mapping value. In this embodiment, In a physical sense, it represents the unmodeled equivalent position deviation caused by nonlinear and discontinuous factors such as gear backlash and static friction superposition within the current control cycle, and completes the dimensionality reduction mapping from high-dimensional physical features to low-dimensional deviation compensation quantities.

[0049] In the above scheme, the residual neural network adopts a multilayer perceptron structure and is configured with multiple hidden layers. It can automatically extract abstract feature representations of nonlinear disturbances such as gear backlash and static friction from high-dimensional physical enhancement features. The equivalent position deviation output by the network quantitatively represents the angular lag caused by unmodeled nonlinear factors in a physical sense. It makes up for the deficiency of traditional linear dynamic models in representing discontinuous features in numerical form.

[0050] A preferred embodiment involves determining a target position deviation physical threshold based on a dexterous hand joint actuator, and then limiting the equivalent position deviation according to the target position deviation physical threshold to obtain a limited deviation. This includes: acquiring the position hysteresis at the output end of the dexterous hand joint actuator when it transitions from rest to motion under no-load or rotor-locked conditions; determining the position hysteresis that satisfies a preset hysteresis threshold as the target position deviation physical threshold; if the absolute value of the equivalent position deviation is greater than the target position deviation physical threshold, then truncating the equivalent position deviation to a symmetrical interval defined by the target position deviation physical threshold to obtain the limited deviation; if the absolute value of the equivalent position deviation is less than or equal to the target position deviation physical threshold, then directly using the equivalent position deviation as the limited deviation.

[0051] For example, during the offline calibration phase, the dexterous hand joint actuator is first measured in one of two specific physical states. The first state is a completely unloaded state, where no load is connected to the joint output end and the fingers are in a free, suspended position. The second state is a rotor-locked state, where the motor rotor is fixed mechanically or electrically, leaving only the clearance space of the gear transmission chain. In both states, the controller sends incrementally small amplitude rotation commands to the dexterous hand joint actuator from zero at an extremely low rate of change. The step size of the commands is much smaller than the expected dead zone range to ensure measurement accuracy. Simultaneously, an angle sensor mounted on the actuator output shaft synchronously and continuously records the actual angular displacement response at the output end. In the initial stage of gradually increasing command, due to the combined resistance of internal gear backlash and static friction, the output end will remain stationary, and the actual angular displacement is essentially zero. At this time, there is a gradually increasing difference between the command position and the sensor feedback position. As the command amplitude continues to increase until the gear teeth inside the gearbox just overcome the backlash and break through the static friction torque, the output end will begin to produce observable macroscopic motion, and the angle value fed back by the sensor will jump. Recording the hysteresis difference between the command position and the feedback position at this moment is the maximum position deviation physical threshold determined by static friction and backlash of this dexterous hand joint actuator. In this embodiment, the preset hysteresis threshold is the maximum position hysteresis, and the maximum position deviation physical threshold is used as the target position deviation physical threshold, denoted as . .

[0052] During the online limiting phase, the residual neural network outputs the equivalent position deviation in each control cycle. Then, the physical threshold of the target position deviation obtained from offline calibration is read from the parameter table. Then the equivalent position deviation Physical threshold of deviation from target position The symmetrically defined amplitude-limiting intervals are compared and judged, as follows: The Clamp operation is a limiting operation. If the equivalent position deviation... The absolute value is greater than the physical threshold of the target position deviation. This indicates that the compensation output by the network under the current operating conditions exceeds the maximum dead zone deviation that the actual actuator may physically produce, and is therefore an unreliable output. A forced truncation operation is then performed to reduce the equivalent position deviation. The value is limited to the physical threshold of the target position deviation. Within the defined interval boundaries, a positive threshold is used when the deviation is positive and a negative threshold is used when the deviation is negative. The truncated value is then output as the amplitude limiting deviation. If the equivalent position deviation... The absolute value is less than or equal to the physical threshold of the target position deviation. This indicates that the network output is within a reasonable range that can be interpreted by the real physical system, and the equivalent position deviation can be directly calculated. The original value is passed to the subsequent overlay stage as the limiting deviation and is not modified in any way.

[0053] In the above scheme, a hard boundary limiting mechanism based on physical limits is introduced to ensure that the compensation amount output by the residual neural network is always constrained within the maximum deviation range that the real physical system can produce. Under no-load micro-motion conditions, the tracking error is small, and the limiting deviation of the network output is converted into feedforward compensation torque to accurately compensate for static friction and dead zone, achieving high fidelity micro-motion. When the dexterous hand comes into rigid contact with the environment, causing the tracking error to expand sharply, the network input features deviate significantly from the no-load training distribution, and the unconstrained network may output abnormal deviations. However, since the limiting mechanism forcibly cuts off the deviation within the physical threshold, the compensation torque generated is strictly limited to a very small range. The control is smoothly transferred to the basic impedance term of the proportional-derivative controller, reducing the safety hazards caused by strategy divergence from the underlying logic and ensuring the physical safety of the dexterous hand joint actuator operation.

[0054] In a preferred embodiment, the equivalent target position command is input into a preset proportional-differential controller to generate a joint driving torque, and the joint driving torque is used to control a dexterous hand joint actuator. This includes: acquiring the proportional gain and differential gain of the preset proportional-differential controller; acquiring the target command position and the actual position command of the current control step; subtracting the equivalent target position command from the target command position to obtain a position deviation; producting the proportional gain with the position deviation to obtain a proportional torque component; producting the differential gain with the derivative of the actual position command to obtain a differential torque component; subtracting the proportional torque component from the differential torque component to obtain the joint driving torque; and applying the joint driving torque to the dexterous hand joint actuator to control its movement.

[0055] For example, in the controller parameter preparation stage, the proportional gain parameter and derivative gain parameter stored in the preset proportional-derivative controller are first read. In this embodiment, the preset proportional-derivative controller is implemented using a standard proportional-derivative (PD) position controller. Specifically, the proportional gain... It is obtained by multiplying the equivalent moment of inertia of the joint by the square of the natural frequency, and is expressed as: In the formula, For the equivalent rotational inertia of the joint, the proportional gain The differential gain determines the response stiffness and steady-state accuracy of a dexterous hand joint actuator to positional deviations. It is obtained by multiplying the equivalent moment of inertia of the joint, twice the damping ratio, and the natural frequency, and is expressed as: Differential gain The damping ability of a dexterous hand joint actuator to suppress changes in velocity is determined by the equivalent moment of inertia. Inertial parameters can be extracted from the three-dimensional model of the dexterous hand joint actuator and the load link, or obtained through time-domain response experiments.

[0056] Within each control cycle, obtain the target command position for the current control step. and the actual joint position of the current control step. That is, the actual position command, and then the equivalent target position command. Subtract the actual joint position of the current control step The position deviation for this control cycle is calculated. Position deviation essentially represents the difference between the expected angle after residual compensation correction and the current true angle. Subsequently, parallel calculations of the torque components are performed. In the proportional channel, the proportional gain is... Multiplying the position deviation yields the proportional torque component. In this embodiment, the proportional torque component includes contributions from both the base impedance restoring force and the feedforward compensation force derived from the limiting deviation. In the differential channel, the differential gain is... Multiply by the actual joint velocity of the current control step, where the actual joint velocity of the current control step can be determined by the actual joint position of the current control step. Differentiating over the time scale yields, which is represented as Finally, the differential torque component is obtained. This component provides a damping force proportional to the motion velocity, used to suppress oscillations and overshoot during the system's approach to the target position. After calculating the proportional and differential torque components, a subtraction operation is performed, subtracting the differential torque component from the proportional torque component to generate the joint drive torque for the current control cycle. The specific calculation formula is as follows: ; in, For positional deviation; due to , These are proportional torque components; Differential torque components; joint driving torque A positive value indicates that the joint is driven to move in the direction of increasing angle. A negative value indicates that the drive joint moves in the direction of decreasing angle.

[0057] After the joint driving torque is generated, the joint driving torque will be... The motor drive interface register of the dexterous hand joint actuator is written, and the current or voltage of the motor winding is controlled by the pulse width modulation signal. After being amplified by the reduction gearbox, it acts on the joint output shaft, driving the dexterous hand joint actuator to move to the equivalent target command position.

[0058] In the above scheme, the equivalent target position command after the amplitude limiting deviation is superimposed is used as the input of the proportional-derivative controller to achieve decoupled distribution of the underlying torque. The product of the proportional gain and the position deviation constitutes the basic impedance torque, while the amplitude limiting deviation is converted into a feedforward compensation torque through the proportional gain. This torque indirectly intervenes in the physical engine through the stiffness coefficient by superimposing position dimensions. The product of the derivative gain and the actual position command provides the damping torque to suppress system oscillation. Thus, the introduced torque distribution mechanism accurately overcomes static friction and dead zone under no-load micro-motion conditions, mainly through the feedforward compensation torque. Under loaded rigid contact conditions, it ensures control stability mainly through the basic impedance torque and damping torque, achieving adaptive and smooth switching under different operating conditions, which helps to improve the physical safety of the dexterous hand joint actuator.

[0059] Based on the above method embodiments, corresponding apparatus embodiments are provided; see [link to apparatus embodiments]. Figure 2 , Figure 2 This is a schematic diagram of the module structure of a dexterous hand joint actuator control system provided in one embodiment of the present invention. Figure 2 As shown, this embodiment of the invention also provides a dexterous hand joint actuator control system, applied to a dexterous hand joint actuator. The dexterous hand joint actuator control system includes a sequence acquisition module 201, a feature vector acquisition module 202, a position deviation calculation module 203, a limiting deviation calculation module 204, a target position command generation module 205, and an actuator control module 206. Specifically: the sequence acquisition module 201 is used to acquire the target command position received in the current control step, extract historical trajectory observation sequences based on a preset historical time window, and solve a pre-constructed second-order dynamic model based on the target command position to obtain an ideal output position sequence; the feature vector acquisition module 202 is used to concatenate the historical trajectory observation sequence and the ideal output position sequence. The sequence is set to obtain a physical enhancement feature vector; the position deviation calculation module 203 is used to input the physical enhancement feature vector into a pre-trained residual neural network to obtain an equivalent position deviation; the amplitude limiting deviation calculation module 204 is used to determine the target position deviation physical threshold based on the dexterous hand joint actuator, and to perform amplitude limiting processing on the equivalent position deviation according to the target position deviation physical threshold to obtain an amplitude limiting deviation; the target position command generation module 205 is used to superimpose the target command position with the amplitude limiting deviation to generate an equivalent target position command; the actuator control module 206 is used to input the equivalent target position command into a preset proportional differential controller to generate a joint driving torque, and to control the dexterous hand joint actuator through the joint driving torque.

[0060] This invention proposes a control system for a dexterous hand joint actuator. The dynamic response of the dexterous hand joint actuator is decoupled into an ideal second-order dynamic response and a residual network nonlinear compensation. An ideal physical model serves as the framework to ensure macroscopic stability. A residual neural network is used to predict equivalent position deviations to compensate for nonlinear disturbances such as dead zones and static friction within the actual dexterous hand joint actuator. Simultaneously, a hard boundary limiting mechanism based on real physical limits is introduced at the output. This allows the network output within the limiting range to be converted into feedforward compensation torque under no-load micro-motion conditions, accurately reproducing the low-speed stalling and dead zone phenomena of the actual actuator, achieving high-fidelity micro-motion. Furthermore, under loaded rigid contact conditions, the feedforward compensation is forcibly truncated, naturally degenerating into robust proportional-derivative control, reducing safety hazards caused by strategy divergence and achieving high safety in rigid contact. Thus, by integrating physical mechanisms and residual networks and introducing a hard boundary limiting mechanism based on physical limits, the physical safety of actuator operation is improved from the underlying logic level, overcoming the differences between the virtual and real migration of the dexterous hand.

[0061] Furthermore, the limiting deviation calculation module 204 includes a deviation physical threshold acquisition unit 301, a first threshold judgment unit 302, and a second threshold judgment unit 303. The deviation physical threshold acquisition unit 301 is used to acquire the position lag amount when the output end of the dexterous hand joint actuator moves from rest to motion under no-load or rotor-locked conditions, and to determine the position lag amount that satisfies a preset lag threshold as the target position deviation physical threshold. The first threshold judgment unit 302 is used to truncate the equivalent position deviation to a symmetrical interval defined by the target position deviation physical threshold if the absolute value of the equivalent position deviation is greater than the target position deviation physical threshold, thereby obtaining the limiting deviation. The second threshold judgment unit 303 is used to directly use the equivalent position deviation as the limiting deviation if the absolute value of the equivalent position deviation is less than or equal to the target position deviation physical threshold.

[0062] It is understood that the above-described device embodiments correspond to the method embodiments of the present invention, and can implement the dexterous hand joint actuator control method provided by any of the above-described method embodiments of the present invention.

[0063] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can specifically be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0064] Based on the above-described embodiment of a dexterous hand joint actuator control method, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a dexterous hand joint actuator control method according to any embodiment of the present invention.

[0065] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0066] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0067] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0068] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the dexterous hand joint actuator control method described in any of the above-described method embodiments of the present invention.

[0069] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0070] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A control method for a dexterous hand joint actuator, characterized in that, A control method for a dexterous hand joint actuator, applicable to a dexterous hand joint actuator, includes: The target command position received in the current control step is obtained, the historical trajectory observation sequence is extracted based on the preset historical time window, and the pre-constructed second-order dynamic model is solved based on the target command position to obtain the ideal output position sequence. By concatenating the historical trajectory observation sequence and the ideal output position sequence, a physical enhancement feature vector is obtained; The physical enhancement feature vector is input into a pre-trained residual neural network to obtain the equivalent positional deviation. The target position deviation physical threshold is determined based on the dexterous hand joint actuator, and the equivalent position deviation is limited according to the target position deviation physical threshold to obtain the limited deviation. The target command position is superimposed with the amplitude limiting deviation to generate an equivalent target position command; The equivalent target position command is input into a preset proportional-derivative controller to generate a joint driving torque, and the joint driving torque is used to control the dexterous hand joint actuator.

2. The dexterous hand joint actuator control method as described in claim 1, characterized in that, The steps of obtaining the target command position received in the current control step, extracting historical trajectory observation sequences based on a preset historical time window, and solving a pre-constructed second-order dynamic model based on the target command position to obtain an ideal output position sequence include: Based on the target instruction position, extract the historical target instructions of several historical control steps within a preset historical time window to form a historical instruction sequence, and extract the joint actual positions of several historical control steps based on the historical instruction sequence to form a historical actual position sequence. The historical actual location sequence is subjected to backward differencing to obtain the historical velocity sequence; The historical command sequence, the historical actual position sequence, and the historical velocity sequence are used as the historical trajectory observation sequence; Based on the preset steady-state gain, preset damping ratio, preset natural frequency and the target command position, a second-order dynamic model is constructed, and the corresponding ideal output position is calculated sequentially according to each historical target command in the historical command sequence to form the ideal output position sequence.

3. The dexterous hand joint actuator control method as described in claim 2, characterized in that, By concatenating the historical trajectory observation sequence and the ideal output position sequence, a physically enhanced feature vector is obtained, including: The historical command sequence, historical actual position sequence, and historical velocity sequence are obtained from the historical trajectory observation sequence. The historical tracking error sequence is obtained by subtracting the historical command sequence from the historical actual position sequence. The historical instruction sequence, the historical actual position sequence, the ideal output position sequence, the historical tracking error sequence, and the historical velocity sequence are standardized respectively, and then concatenated in a preset order to obtain a physical enhancement feature vector.

4. The dexterous hand joint actuator control method as described in claim 1, characterized in that, In obtaining the equivalent positional bias by inputting the physical enhancement feature vector into a pre-trained residual neural network, the pre-trained residual neural network comprises: an input layer, at least two hidden layers, and an output layer; wherein: The input layer is used to receive the physical enhancement feature vector; The hidden layer is used to perform a non-linear mapping on the physical enhancement feature vector to obtain the physical enhancement feature vector mapping value; The output layer is used to output the physical enhancement feature vector mapping value as the equivalent positional deviation.

5. The dexterous hand joint actuator control method as described in claim 1, characterized in that, Based on the dexterous hand joint actuator, a physical threshold for the target position deviation is determined, and the equivalent position deviation is limited according to the physical threshold to obtain the limited deviation, including: The position hysteresis of the output terminal of the dexterous hand joint actuator from rest to motion is obtained when the dexterous hand joint actuator is in an unloaded or rotor-locked state. The position hysteresis that meets the preset hysteresis threshold is determined as the target position deviation physical threshold. If the absolute value of the equivalent position deviation is greater than the target position deviation physical threshold, then the equivalent position deviation is truncated to the symmetrical interval defined by the target position deviation physical threshold to obtain the amplitude limiting deviation. If the absolute value of the equivalent position deviation is less than or equal to the physical threshold of the target position deviation, then the equivalent position deviation is directly used as the amplitude limiting deviation.

6. The dexterous hand joint actuator control method as described in claim 1, characterized in that, The equivalent target position command is input into a preset proportional-derivative controller to generate a joint driving torque, and the dexterous hand joint actuator is controlled by the joint driving torque, including: Obtain the proportional gain and derivative gain of the preset proportional-derivative controller; Obtain the target command position and the actual position command of the current control step; The position deviation is obtained by subtracting the equivalent target position command from the target command position. The proportional gain is multiplied by the position deviation to obtain the proportional torque component; The differential torque component is obtained by multiplying the differential gain with the derivative of the actual position command. The joint driving torque is obtained by subtracting the proportional torque component from the differential torque component. The joint driving torque is applied to the dexterous hand joint actuator to control the movement of the dexterous hand joint actuator.

7. A dexterous hand joint actuator control system, characterized in that, The system is applied to a dexterous hand joint actuator, and includes a sequence acquisition module, a feature vector acquisition module, a position deviation calculation module, a amplitude limiting deviation calculation module, a target position command generation module, and an actuator control module; wherein: The sequence acquisition module is used to acquire the target command position received in the current control step, extract the historical trajectory observation sequence based on the preset historical time window, and solve the pre-constructed second-order dynamic model based on the target command position to obtain the ideal output position sequence. The feature vector acquisition module is used to concatenate the historical trajectory observation sequence and the ideal output position sequence to obtain a physical enhancement feature vector; The position deviation calculation module is used to input the physical enhancement feature vector into a pre-trained residual neural network to obtain the equivalent position deviation; The amplitude limiting deviation calculation module is used to determine the physical threshold of the target position deviation based on the dexterous hand joint actuator, and to perform amplitude limiting processing on the equivalent position deviation according to the physical threshold of the target position deviation to obtain the amplitude limiting deviation; The target position command generation module is used to superimpose the target command position with the amplitude limiting deviation to generate an equivalent target position command. The actuator control module is used to input the equivalent target position command into a preset proportional-derivative controller to generate joint driving torque, and to control the dexterous hand joint actuator through the joint driving torque.

8. A dexterous hand joint actuator control system as described in claim 7, characterized in that, The amplitude limiting deviation calculation module includes a deviation physical threshold acquisition unit, a first threshold judgment unit, and a second threshold judgment unit; The deviation physical threshold acquisition unit is used to acquire the position lag amount when the output end of the dexterous hand joint actuator changes from stationary to motion under no-load or rotor-locked conditions, and to determine the position lag amount that meets the preset lag threshold as the target position deviation physical threshold. The first threshold determination unit is used to cut off the equivalent position deviation to a symmetrical interval defined by the target position deviation physical threshold if the absolute value of the equivalent position deviation is greater than the target position deviation physical threshold, thereby obtaining the amplitude limiting deviation. The second threshold determination unit is used to directly use the equivalent position deviation as the amplitude limiting deviation if the absolute value of the equivalent position deviation is less than or equal to the target position deviation physical threshold.

9. A dexterous hand joint actuator control terminal, characterized in that, The terminal includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements a dexterous hand joint actuator control method as described in any one of claims 1 to 6.

10. A dexterous hand joint actuator control storage medium, characterized in that, The storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the storage medium is located to perform a dexterous hand joint actuator control method as described in any one of claims 1 to 6.

Images