A control method and system for a robotic arm for tethered and collimated robots.

By constructing a static transmission mapping and dynamic angle residual model, and combining the main torque network of the rigid body of the whole arm and the grouped torque residual model, the accuracy problem of force control modeling for rope-driven and quasi-direct-drive robots was solved, and stable control of multi-degree-of-freedom robots was achieved.

CN123103707APending Publication Date: 2026-11-17BEIHANG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611618188.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to accurately model torque control in rope-driven and quasi-direct-drive robots, particularly in areas such as joint coupling, dynamic hysteresis, mixing of underlying controller behavior, and differences in error scales between large and small joints, resulting in inaccurate force control modeling.

Method used

By combining ridge regression processing and GRU network with multilayer perceptron, a static transmission mapping model, a dynamic angle residual model, a rigid body main torque network of the whole arm, and a grouped torque residual model are constructed. By separating and modeling the signal boundary and group structure, an impedance feedback model is formed to control the movement of the robotic arm.

Benefits of technology

It improves the generalization ability and stability of the model, can accurately describe the joint mapping relationship of multi-rope coupled joints, capture the historical effect and velocity-related error in flexible transmission, and is applicable to multi-degree-of-freedom robot systems such as humanoid robot upper limbs, dexterous hands, and tendon-driven robotic arms, providing a real-time available controller.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123103707A_ABST
    Figure CN123103707A_ABST
Patent Text Reader

Abstract

This invention relates to the field of humanoid robot upper limb control technology, specifically to a control method and system for a robotic arm of a rope-driven and quasi-direct-drive robot. The method includes: establishing a training dataset based on robot operation data; establishing a static transmission mapping model for estimating static joint angles from motor positions; training a dynamic angle residual model for obtaining angle residuals from motor states and static joint angle estimates; training a rigid body backbone torque network for estimating motor torques from corrected joint angles; training a grouped torque residual model for estimating torque residuals from corrected joint angle estimates and motor torque estimates; fitting an impedance feedback model for obtaining impedance feedback torques from torque feedforward estimation; acquiring real-time robot operation data, processing it sequentially through the above models, and controlling the robotic arm movement with the obtained impedance feedback torques. This invention can improve the accuracy of force control modeling for robotic arms.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of humanoid robot upper limb control technology, specifically to a control method and system for a robotic arm of a rope-driven and quasi-direct-drive robot. Background Technology

[0002] Rope-driven, tendon-driven, wire rope-driven, and collimated-driven robots have significant advantages in lightweight design, high power density, compliant interaction, and safe collaboration, making them particularly suitable for humanoid robot arms, dexterous hands, mobile manipulation platforms, and rehabilitation robots. However, compared to traditional rigid-drive robots, these robots present significant modeling and control challenges: First, the relationship between motor-side angles and joint-side angles is not a fixed ratio, but is affected by variations in rope length, winding radius, flexible elongation, gaps, friction, backlash, and multi-rope coupling, resulting in configuration-dependent nonlinear static mappings. Second, dynamic lags exist during motion, including velocity-dependent, acceleration-dependent, and history-dependent lags, making it difficult to recover the true joint state simply by relying on instantaneous motor positions. Third, there are coupled inertia, gravity terms, friction terms, and servo system closed-loop responses between robot joints, especially at the shoulder, wrist, or combinations of multiple rope-driven joints, which traditional rigid-body dynamics models struggle to fully cover. Fourth, real-world control typically requires outputting stable torque commands within limited computational resources and real-time cycles; therefore, the model must possess sufficient expressive power, interpretability, layered debugging capabilities, and degenerate operation characteristics.

[0003] Existing methods typically include analytical dynamics modeling, data-driven calibration or compensation, and feedback force control methods such as impedance / PID. For example, in the field of tethered surgical robots, existing solutions use linear regression or deep networks to correct inaccurate joint position estimates and compensate for tether hysteresis and historical errors using networks such as LSTM and DNN. In tendon-driven dexterous hands, tethered flexible arms, and tethered parallel mechanisms, there are publicly available solutions that utilize neural networks, RBF networks, or friction models to compensate for unmodeled force / torque terms.

[0004] However, the existing solutions still have several shortcomings: First, joint angle calibration models mostly focus on position estimation and do not further integrate with the whole-arm torque feedforward model to form a diagnosable two-stage link; second, torque learning models tend to simultaneously absorb target commands, tracking errors, and low-level controller behavior, resulting in insufficient generalization of the backbone model to new trajectories or control gains; third, unified networks or unified contexts can easily allow large-load joint errors in the shoulder to mask small-inertia joint errors in the wrist, end effector, etc., and lack a heterogeneous grouping mechanism that configures contexts, windows, and loss functions according to anatomical / transmission structure differences; fourth, existing solutions suffer from insufficient historical degradation, missing feature checks, and residual availability flag support when applied to align-drive, humanoid upper limb, and multi-level hybrid transmission robots in real-time. Therefore, a control method for a robotic arm that combines multiple transmission requirements into a trainable, verifiable, and deployable model chain is needed. Summary of the Invention

[0005] In view of the above problems, the present invention provides a control method and system for a robotic arm for rope-driven and quasi-direct-drive robots, which solves the technical problems of inaccurate force control modeling caused by transmission flexibility, joint coupling, dynamic lag, mixing of behavior of the underlying controller, differences in error scale of large and small joints and data noise in the prior art.

[0006] On one hand, the present invention provides a control method for a robotic arm for rope-driven and quasi-direct-drive robots, comprising the following steps: Step S1: Collect robot operation data and preprocess it to obtain a training dataset; the robot has a robotic arm with multiple joints; the data in the training dataset includes the motor state, joint state, target joint command, and motor torque of the robotic arm; the motor state includes motor position, motor speed, and motor acceleration, and the joint state includes joint angle, joint speed, and joint acceleration; Step S2: Perform ridge regression processing based on the training dataset to obtain the static transmission mapping model; the static transmission mapping model can obtain the static joint angle estimate from the input motor position; Step S3: A dynamic angle residual model is trained based on the training dataset and static joint angle estimation. The dynamic angle residual model adopts a GRU network, which can obtain the angle residual from the input motor state and static joint angle estimation. Step S4: Based on the training dataset and the corrected joint angle estimation of some joints, the rigid body main torque network of the whole arm is trained. The rigid body main torque network of the whole arm adopts a multilayer perceptron and can obtain the motor torque estimation from the input corrected joint angle. The corrected joint angle estimate of the partial joint is obtained by adding the static joint angle estimate of the partial joint to the angle residual; Step S5: Based on the training dataset, the corrected joint angle estimate and the motor torque estimate are used to train and obtain the grouped torque residual model. The grouped torque residual model can obtain the torque residual estimate from the input corrected joint angle estimate and motor torque estimate. Step S6: Fit the model based on the training dataset and the torque feedforward estimation to obtain the impedance feedback model; the impedance feedback model can obtain the impedance feedback torque from the input torque feedforward estimation. The torque feedforward estimation represents the sum of the motor torque estimation and the torque residual estimation; Step S7: For the robotic arm to be controlled, real-time robot operation data is acquired and processed sequentially through the static transmission mapping model, dynamic angle residual model, whole arm rigid body main torque network, grouped torque residual model and impedance feedback model. The obtained impedance feedback torque controls the movement of the robotic arm.

[0007] Preferably, in step S1, the robotic arm has 7 joints, the first to the seventh joints being: arm raising joint, arm extending joint, shoulder twisting joint, elbow flexion joint, wrist turning A joint, wrist turning B joint, and wrist twisting joint; each joint has a motor, and the first to the seventh joints correspond to the first to the seventh motors respectively.

[0008] Preferably, in step S1, the step of collecting robot operation data and preprocessing it to obtain a training dataset specifically includes: (1) Parse the original text log, calculate the cumulative number of seconds within the day, and generate a sampling interval based on the difference between adjacent time periods, which is used as a timestamp; (2) For the first and second joints, the motor position is obtained from the motor side, and the motor speed and acceleration are obtained by differentiation to form the motor state; the motor position is used as the joint angle, and the joint speed and joint acceleration are determined by differentiation to form the joint state. For the third joint, the motor state is obtained, and the joint angle is calculated by converting the motor position using a fixed ratio and direction sign, thus obtaining the joint state; For joints 4 to 7, obtain the motor status, decode the joint angles using the joint-side decoder, and then obtain the joint status. Read the target joint commands and motor torque from the raw text log; (3) Perform low-pass filtering on the timestamped motor state, joint state, target joint command and motor torque sequence to form the training dataset.

[0009] Preferably, step S2 specifically includes: (1) The joint structure of the robotic arm is divided into configuration groups, including: single joint configuration group, coupled joint configuration group and end joint configuration group; In this embodiment, the single-joint configuration group uses the position of the fourth motor to predict the static joint angle estimate of the fourth joint; the coupled joint configuration group uses the positions of the fifth and sixth motors to jointly predict the static joint angle estimate of the fifth and sixth joints; and the end-joint configuration group uses the position of the seventh motor to predict the static joint angle estimate of the seventh joint. (2) Normalize the mean and variance of each motor position, construct first to fourth order candidate polynomial features, and perform ridge regression based on the first to fourth order candidate polynomial features. Among the multiple ridge regression regularization coefficients obtained, the mapping function corresponding to the ridge regression regularization coefficient with the smallest verification error is selected as the static transmission mapping model. The expression for the static transmission mapping model is:

[0010] in, This represents the static joint angle estimation. Indicates the motor position; This represents a static transmission mapping model.

[0011] Preferably, in step S3, the expression for the dynamic angle residual model is:

[0012] Indicates angular residual; Represents a dynamic angle residual model; Indicates from the first sampling time to the... Input features at each sampling time; Indicates the length of the timing window.

[0013] Preferably, in step S4, the corrected joint angle estimation input to the rigid body main moment network of the whole arm specifically includes: The measured joint angles are directly used as the corrected joint angle estimates for joints 1 to 3; The corrected joint angle expression for joints 4 to 7 is:

[0014] in, This indicates a correction to the joint angle estimation.

[0015] Preferably, in step S4, the joint angle estimation is corrected. As the input state to the main moment network of the rigid body of the entire arm, the expression for the main moment network of the rigid body of the entire arm is:

[0016] in, This indicates an estimate of the motor torque. The mapping function representing the main moment network of the rigid body of the entire arm; This indicates the input state of the main torque network; and Let represent the first and second derivatives corresponding to the input states of the main torque network, respectively. and The sine and cosine characteristics of the joint angles are represented.

[0017] Preferably, in step S5, the input features of the grouped torque residual model come from multiple torque residual groups, specifically including: shoulder double joint group, intermediate joint group, wrist single joint group, wrist coupling group and distal joint group; The shoulder dual-joint group uses the context of joints 1 to 7, the motor states of motors 1 to 3, the target joint command, the tracking error, the motor torque estimate, the joint velocity sign, and the absolute value of the joint velocity as input features; The intermediate joint group uses the context of joints 1 to 4 and the motor states of motors 1 to 3 as input features; The wrist single joint group, wrist coupled group, and distal joint group used the context and motor state related to the 4th, 5th, 6th, and 7th joints as input features, respectively. In step S5, the expression for the grouped moment residual model is:

[0018]

[0019] in, This represents the torque residual estimate; Represents the grouped moment residual model; Indicates from the first sampling time to the... Input features of the torque residual model at each sampling time; This indicates torque feedforward estimation.

[0020] Preferably, in step S6, the expression for the impedance feedback model is:

[0021] in, This indicates the impedance feedback torque. Indicates the stiffness coefficient. Indicates the damping coefficient. Indicates the bias term. Indicates the target joint angle in the target joint command; This indicates the target joint velocity in the target joint command.

[0022] On the other hand, the present invention provides a control system for a robotic arm for rope-driven and quasi-direct-drive robots, comprising: The acquisition and preprocessing module is used to acquire robot operation data and preprocess it to obtain a training dataset. The robot has a robotic arm with multiple joints. The data in the training dataset includes the motor state, joint state, target joint command, and motor torque of the robotic arm. The motor state includes motor position, motor speed, and motor acceleration, and the joint state includes joint angle, joint speed, and joint acceleration. The static drive mapping module is used to perform ridge regression processing based on the training dataset to obtain a static drive mapping model; the static drive mapping model can obtain static joint angle estimates from the input motor position; The dynamic angle residual module is used to train a dynamic angle residual model based on the training dataset and static joint angle estimation. The dynamic angle residual model adopts a GRU network and can obtain the angle residual from the input motor state and static joint angle estimation. The rigid body main torque network module is used to train the rigid body main torque network of the whole arm based on the training dataset and the corrected joint angle estimation of some joints. The rigid body main torque network of the whole arm adopts a multilayer perceptron and can obtain the motor torque estimation from the input corrected joint angle. The corrected joint angle estimate of the partial joint is obtained by adding the static joint angle estimate of the partial joint to the angle residual; The grouped torque residual module is used to train a grouped torque residual model based on the training dataset, the corrected joint angle estimate, and the motor torque estimate. The grouped torque residual model can obtain the torque residual estimate from the input corrected joint angle estimate and motor torque estimate. The impedance feedback module is used to fit the training dataset and the torque feedforward estimate to obtain the impedance feedback model; the impedance feedback model can obtain the impedance feedback torque from the input torque feedforward estimate. The torque feedforward estimation represents the sum of the motor torque estimation and the torque residual estimation; The control module is used to acquire real-time robot operation data for the robotic arm to be controlled. The data is then processed sequentially through the static transmission mapping model, dynamic angle residual model, whole arm rigid body main torque network, grouped torque residual model, and impedance feedback model. The obtained impedance feedback torque controls the movement of the robotic arm.

[0023] Compared with the prior art, the present invention has at least the following beneficial effects: First, this invention proposes a framework for decoupling the backbone and modeling heterogeneous grouping residuals for force control of tethered and quasi-direct drive robots, which separates and models static transmission, dynamic hysteresis, the backbone of whole-arm dynamics, local torque residuals, and impedance feedback. This framework does not rely on "layering" itself as the sole distinguishing factor, but rather improves the stability of model debugging, migration, and deployment through signal boundaries, group structure, and real-time degradation mechanisms.

[0024] Second, the static transmission mapping model of this invention models configuration groups rather than individual joints independently, which can describe the joint mapping relationship in multi-rope coupled joints. In particular, for coupled joints such as the 5th and 6th joints, the joint input of multiple motor-side positions and the joint output of multiple joint-side positions are more consistent with the actual mechanical transmission structure than single-joint fitting, and provide a physically interpretable baseline for subsequent dynamic residuals.

[0025] Third, this invention uses a dynamic angle residual model to predict angle residuals after the static layer, enabling the model to capture historical effects, speed-related errors, and torque-related deformations in flexible transmission, while maintaining the interpretable basis of the static mapping.

[0026] Fourth, this invention distinguishes between the rigid body trunk torque network of the entire arm and the torque residual model. The rigid body trunk torque network of the entire arm does not use tracking error signals and target joint commands, avoiding overfitting of low-level controller behavior; the torque residual model, on the other hand, allows the introduction of commands, errors and trunk outputs for compensation tasks, improving the fitting ability of unmodeled terms, thus forming a combined structure that is different from existing joint angle calibration schemes and simple compensation schemes.

[0027] Fifth, the torque residual model of the present invention adopts different context ranges, local motor side features, temporal window lengths, hidden layer dimensions or loss functions for the shoulder, intermediate joint, wrist coupled joint and end joint, taking into account the error scale differences between high-load joints and low-inertia joints, avoiding mutual interference between different joints in the unified network, and supporting group diagnosis and degradation.

[0028] Sixth, this invention provides a complete training, verification, export, and real-time deployment process, capable of generating a usable controller directly from raw logs. The real-time controller features engineering mechanisms such as a circular buffer, history length detection, residual model degradation, missing feature checking, residual availability flag, torque limiting, and debugging information output, making it suitable for real-world robot applications.

[0029] Seventh, this invention can be applied not only to the upper limbs of humanoid robots, but also to dexterous hands, tendon-driven robotic arms, wire-driven parallel mechanisms, elastic joint robots, quasi-direct-drive quadruped robots, or other multi-degree-of-freedom robot systems with nonlinear transmission and compliance errors. In embodied intelligent systems, this invention can serve as a low-level force control interface between upper-level model predictive control, reinforcement learning, or vision-language-action strategies and the robot body, receiving the target joint state and outputting real-time torque commands, torque feedforward, impedance feedback, and residual availability status.

[0030] Eighth, this invention separates the upper-level task strategy from the lower-level flexible transmission compensation through a unified real-time sensor status interface, target joint status interface, reset and single-step call interface, as well as historical length detection, feature missing check, and degradation path. When the historical data is insufficient or the input features are missing, the controller can still output rigid body feedforward plus impedance feedback and return the residual usable state to the upper layer, thereby maintaining the continuity and diagnosability of the lower-level execution interface. Attached Figure Description

[0031] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0032] Figure 1 This is a schematic diagram of the raw data, preprocessing filtering, and curve preview interface provided by the present invention.

[0033] Figure 2 A flowchart illustrating the graphical training process provided by this invention.

[0034] Figure 3 A flowchart illustrating the workflow for packaging and deploying training products on a real machine, as provided by this invention.

[0035] Figure 4 This is a schematic diagram of the static transmission mapping model provided by the present invention.

[0036] Figure 5 This is a schematic diagram of the dynamic angle residual model provided by the present invention.

[0037] Figure 6 This is a schematic diagram of the main moment network and grouped moment residual model of the rigid body of the whole arm provided by the present invention.

[0038] Figure 7 This is a schematic diagram of real-time control provided by the present invention. Detailed Implementation

[0039] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0040] This invention decomposes complex mappings into multiple sub-problems with clear boundaries through a hierarchical structure of backbone decoupling and heterogeneous grouping: the configuration-related relationship from the motor side to the joint side is restored by static transmission mapping of configuration groups; transmission lag is compensated by a time-series dynamic residual network; the main gravity, inertia, and coupling feedforward are learned by a backbone network of whole arm torque that does not contain target commands and tracking errors; local unmodeled torques are learned by a grouped GRU residual network configured according to anatomical parts; and the final control closed loop is formed by impedance feedback with directional signs, thereby improving the generalization, stability, interpretability, and real-world usability of the force control model.

[0041] To illustrate the effectiveness of the method proposed in this invention, the following detailed description of the above-mentioned technical solution of this invention is provided through a specific embodiment of a seven-degree-of-freedom robotic arm. A control method for a robotic arm designed for rope-driven and quasi-direct-drive robots is disclosed, with the specific implementation steps as follows: Step S1: Collect robot operation data and preprocess it to obtain a training dataset; the robot has a robotic arm with multiple joints; the data in the training dataset includes the motor state, joint state, target joint command, and motor torque of the robotic arm; the motor state includes motor position, motor speed, and motor acceleration, and the joint state includes joint angle, joint speed, and joint acceleration; In this embodiment, the robot has a robotic arm, which is a multi-joint robotic arm configured for a rope-driven robot or a quasi-direct-drive robot. It includes multiple joints connected in sequence and motors corresponding to each joint. Each motor is used to drive the movement of the corresponding joint.

[0042] The robotic arm has 7 joints, the first to the seventh joints being: arm raising joint, arm extending joint, shoulder twisting joint, elbow flexion joint, wrist turning A joint, wrist turning B joint, and wrist twisting joint; each joint has a motor, and the first to the seventh joints correspond to the first to the seventh motors respectively.

[0043] During the operation of the robotic arm, the motor side generates motor status information such as motor position, motor speed, motor acceleration, and motor torque, while the joint side generates joint status information such as joint angle, joint speed, and joint acceleration, and controls the movement of each joint according to the target joint command.

[0044] In this step, the motor status, joint status, target joint commands, and motor torque are collected during the robot's operation. The collected robot operation data is then time-aligned, differentiated, and filtered to obtain the training dataset, as detailed below.

[0045] (1) Time alignment First, robot operation data is read from the raw text logs. The time column in the logs can be in the format of "hour:minute:second" or "minute:second". After parsing, the cumulative number of seconds within the day is calculated, and a sampling interval is generated based on the difference between adjacent time periods, serving as a timestamp. When the sampling interval for the first line or an abnormal line of the file is unavailable, the median of the valid sampling intervals or a preset file interval is used to fill the gap. When concatenating multiple log files, the file source is preserved through the file source identifier for subsequent grouping and splitting.

[0046] (2) Obtain velocity and acceleration by differentiating position information, and read target joint commands and motor torque. The training dataset is obtained by constructing features based on the original text logs. For joints 1 and 2, the motor position generated by the motor side is obtained, and the motor speed and acceleration are obtained by differentiation to form the motor state. The motor position is used as the joint angle, and then the joint speed and joint acceleration are determined by differentiation to form the joint state. For joint 3, the motor state is obtained, and the joint angle is obtained by converting the motor position using a fixed ratio and direction sign. For joints 4 to 7, the motor state is obtained, and then the joint angle obtained by decoding the joint angle is obtained using the joint side decoder to obtain the joint state.

[0047] In some embodiments, the robot has a left arm and a right arm, and motor mapping tables can be established for the left and right arms respectively to mark the feature data of the two robotic arms.

[0048] In addition, target joint commands and motor torques are read from the original text log. The target joint commands are control commands used to indicate the desired motion state of each joint of the robotic arm. The motor torque is the torque output by the motor to drive the corresponding joint movement, and its value reflects the driving force applied by the motor to the joint of the robotic arm.

[0049] In some embodiments, more information can be read from the raw text log as feature data, such as motor speed feedback, motor acceleration feedback, tracking error, optional force sensor data, and robotic arm side identification.

[0050] (3) Filter the data to obtain the training dataset. In this step, low-pass filtering can be applied to the timestamped motor state, joint state, target joint command, and motor torque sequence. The low-pass filtering can use a preset cutoff frequency.

[0051] Examples of the fields after final preprocessing in this invention are as follows: they may include t, q_cmd_j, qm_j, qm_dot_diff, qm_dot_feedback, qm_ddot_diff, qm_ddot_feedback, tau_m, tau_m_sm, q_dec, q_j, q_dot, q_ddot, e_track, force_1 to force_3, and arm_side. Where t is the timestamp, q_cmd_j is the target joint command for the j-th joint, qm_j is the motor position of the j-th motor, qm_dot_diff and qm_dot_feedback are the motor speed and motor speed feedback, respectively, qm_ddot_diff and qm_ddot_feedback are the motor acceleration and motor acceleration feedback, respectively, tau_m and tau_m_sm are the motor torque and filtered motor torque, respectively, q_dec and q_j are the joint-side decoder information and joint angle of the j-th joint, q_dot and q_ddot are the joint speed and joint acceleration, e_track is the tracking error, force_1 to force_3 are optional force sensor data, arm_side is the arm side, and j is the joint number.

[0052] like Figure 2 As shown, the present invention uses a graphical interface to display the original data, the preprocessed fields, and the filtering and curve preview. The filtering and curve preview overlays the original signal and the processed signal to observe the suppression effect of low-pass filtering on spike noise, speed differential noise, and motor torque jitter.

[0053] This step transforms the raw robot operation data into a training dataset that features uniform fields, time alignment, and retains the data's origin. This training dataset is used for the training and validation of subsequent models.

[0054] Step S2: Perform ridge regression processing based on the training dataset to obtain the static transmission mapping model; the static transmission mapping model can obtain the static joint angle estimate from the input motor position; In this step, the invention divides the joint structure of the robotic arm into configuration groups, using the motor positions within each configuration group as input and the estimated static joint angles of the corresponding joints as output. A polynomial feature is constructed from the input and ridge regression is performed to obtain a static transmission mapping model.

[0055] For roped robots, the relationship between joint angles and motor positions often cannot be expressed using a fixed transmission ratio. For example, the wrist or elbow may be driven by multiple coupled ropes, with the 5th and 6th joints being mutually coupled.

[0056] Therefore, this invention divides the joint structure of the robotic arm into configuration groups and establishes a static transmission mapping model according to the configuration groups, such as... Figure 4 As shown, the configuration group includes: a single-joint configuration group, a coupled-joint configuration group, and an end-joint configuration group.

[0057] In this embodiment, the single-joint configuration group uses the position of the fourth motor to predict the static joint angle estimate of the fourth joint; the coupled joint configuration group uses the positions of the fifth and sixth motors to jointly predict the static joint angle estimate of the fifth and sixth joints; and the end-joint configuration group uses the position of the seventh motor to predict the static joint angle estimate of the seventh joint.

[0058] This invention constructs first to fourth-order candidate polynomial features after normalizing the mean and variance of the motor position, and selects the one with the smallest verification error from multiple ridge regression regularization coefficients after ridge regression processing.

[0059] The expression for the static transmission mapping model constructed in this invention is as follows:

[0060] in, This represents the static joint angle estimation. Indicates the motor position; The mapping function represents the static transmission mapping model.

[0061] In this embodiment, the static transmission mapping model was trained on approximately 437,000 training samples and 109,000 validation samples for the left arm. The candidate ridge regression regularization coefficients included 1e-8, 1e-6, 1e-4, 1e-2, and 1e-1. After training, a fourth-order polynomial with a regularization coefficient of 0.1 was selected for the single-joint configuration group, a third-order polynomial with a regularization coefficient of 1e-8 was selected for the coupled-joint configuration group, and a fourth-order polynomial with a regularization coefficient of 0.1 was selected for the distal end joint configuration group. The root mean square errors of validation for the single-joint configuration group, the coupled-joint configuration group, and the distal end joint configuration group were approximately 0.0083 rad, 0.0164 rad, and 0.0642 rad, respectively.

[0062] In some embodiments, the information stored in the static transmission mapping model may include: input joints, output joints, feature columns, mean, variance, polynomial power, regression coefficients, intercept, and training / validation error.

[0063] Through this step, the nonlinear transmission relationship is transformed into an interpretable and rapidly reasonable static transmission mapping model. The resulting static joint angle estimates are used for training the dynamic angle residual model in subsequent steps.

[0064] Step S3: A dynamic angle residual model is trained based on the training dataset and static joint angle estimation. The dynamic angle residual model adopts a GRU network, which can obtain the angle residual from the input motor state and static joint angle estimation. In this step, the present invention constructs a timing window for each configuration group, taking the motor state and static joint angle estimation as input. A GRU network is used to predict the angle residuals of the static joint angle estimation, and a dynamic angle residual model is trained to obtain the model.

[0065] The static transmission mapping model only describes the instantaneous relationship between the motor position and the joint angle, and cannot represent the rope elasticity, damping, friction, hysteresis, and controller lag. Therefore, this step establishes a dynamic angle residual model, which is used to predict the angle residual of the static joint angle estimate.

[0066] like Figure 5 As shown, each sample in the dynamic angle residual model has a length of The data is from a historical window. Input features can be selected by experimental combination, including motor position, motor speed feedback, speed obtained by the derivative of motor position, motor acceleration, motor torque, target joint command, static joint angle estimation, error of static joint angle estimation relative to target joint command, and selectable sampling interval.

[0067] In this step, different combinations of feedback features only, differential features only, and feedback features and differential features can be trained simultaneously, and the feature combination corresponding to the better experiment can be selected according to the root mean square error of the joint angle estimation corrected by the validation set or the improvement ratio.

[0068] The dynamic angle residual model of this invention employs a GRU network. In some embodiments, the time window length... The GRU network has a hidden layer dimension of 64, 2 layers, a dropout value of 0.1, a learning rate of 1e-3, a weight decay of 1e-5, and 15 early stopping rounds.

[0069] In some embodiments, the input features and angle residuals are standardized during training, and dropout and gradient clipping are used; the training set and validation set can also be divided by retaining the file source identifier, so that the model validation is closer to the generalization of the new trajectory, and finally a dynamic angle residual model is obtained.

[0070] The expression for the dynamic angle residual model is:

[0071] Indicates angular residual; Represents a dynamic angle residual model; Indicates from the first sampling time to the... Input features at each sampling time; Indicates the length of the timing window.

[0072] After the dynamic angle residual model outputs the angle residual, the corrected joint angle estimate is obtained through the following expression:

[0073] in, This indicates a correction to the joint angle estimation; This represents the static joint angle estimation; This represents the angular residual.

[0074] In this embodiment, multiple input combination experiments were trained for the single-joint configuration group, the coupled-joint configuration group, and the end-joint configuration group, and the optimal experiment was selected based on the validation set results. After using all feature inputs, the validation angle error of the single-joint configuration group was improved by approximately 24.17% relative to the static joint angle estimation, and the end-joint configuration group was improved by approximately 7.26%.

[0075] After training, the dynamic angle residual model, standardized parameters, feature list, static joint angle estimation column, and target column are saved as the model file of the dynamic angle residual model.

[0076] This step compensates for velocity-dependent hysteresis, backlash, flexible deformation, and historical state effects in the rope drive system using the dynamic angle residual model. The resulting corrected joint angle estimate is used for subsequent training of the whole-arm rigid body backbone torque network.

[0077] Step S4: Based on the training dataset and the corrected joint angle estimation, the rigid body main torque network of the whole arm is trained. The rigid body main torque network of the whole arm adopts a multilayer perceptron and can obtain the motor torque estimation from the input corrected joint angle estimation. The corrected joint angle estimate represents the sum of the static joint angle estimate and the angle residual; In this step, the modified joint angle estimate is used as the unified joint state input to the rigid body backbone torque network of the entire arm. The derivative of the modified joint angle estimate is calculated and periodic features are constructed. The motor torque in the training dataset is used as the supervision target to train the rigid body backbone torque network of the entire arm using a multilayer perceptron.

[0078] As mentioned above, in this embodiment, the static transmission mapping model and the dynamic angle residual model are used to predict the corrected joint angles of the 4th to 7th joints; while for the proximal joints, namely the 1st to 3rd joints, the present invention directly uses the measured joint angles as the estimates of the corrected joint angles of the 1st to 3rd joints.

[0079] The first and second derivatives of the corrected joint angle estimates for all joints are calculated to obtain the joint velocities and accelerations corresponding to the corrected joint angle estimates, and then smoothed using a rolling window. To handle the periodicity of the angles, sine and cosine features of the corrected joint angle estimates are added to the features.

[0080] This invention will correct the joint angle estimation As the input state of the main rigid body moment network of the entire arm, the input state of the main rigid body moment network of the entire arm is represented as follows: As mentioned earlier, for proximal joints, The joint angle is estimated using measured joint angles or existing joint angles. Find the first derivative and second derivative and construct and feature.

[0081] The rigid body backbone network of the whole arm consists of the joints , , , and The feature is the input.

[0082] The rigid arm backbone torque network employs a multilayer perceptron with two hidden layers, each containing 256 neurons. Input features are standardized, and training is completed using motor torque as the supervised objective, resulting in an output motor torque estimate. The expression for the rigid arm backbone torque network is as follows:

[0083] in, This indicates an estimate of the motor torque. The mapping function representing the main moment network of the rigid body of the entire arm; This indicates the input state of the main torque network; and Let represent the first and second derivatives corresponding to the input states of the main torque network, respectively. and The sine and cosine characteristics of the joint angles are represented.

[0084] The overall rigid body backbone moment network of this invention is positioned to learn the main dynamic feedforward. Therefore, the input to the overall rigid body backbone moment network does not include target joint commands, tracking errors, or other signals that clearly reflect the behavior of the underlying controller; it only uses modified joint angle estimates and their corresponding joint velocities, joint accelerations, sine characteristics, and cosine characteristics. This enhances the generalization ability of the overall rigid body backbone moment network to new target trajectories, new error distributions, and different control gains.

[0085] In this step, the main rigid body torque network of the entire arm obtains the motor torque estimate based on the corrected joint angle estimate. The obtained motor torque estimate serves as both the baseline for torque residual fitting and the input feature for subsequent steps.

[0086] Step S5: Based on the training dataset, the corrected joint angle estimate and the motor torque estimate are used to train and obtain the grouped torque residual model. The grouped torque residual model can obtain the torque residual estimate from the input corrected joint angle estimate and motor torque estimate. In this step, the joints are divided into multiple torque residual groups based on the robot's anatomical location, transmission structure, and error sources. Using features such as corrected joint angle estimates and motor torque estimates as inputs, and the difference between the motor torque and the estimated motor torque as the supervision target, grouped torque residual models corresponding to each torque residual group are trained.

[0087] Relying solely on the rigid main torque network of the entire arm is insufficient to cover the torque residuals caused by flexible transmission, local friction, large shoulder load, wrist coupling, and the servo inner loop. Therefore, this invention provides a grouped torque residual model.

[0088] First, the torque residual groups of the seven joints are divided as follows: shoulder double joint group, intermediate joint group, wrist single joint group, wrist coupling group, and terminal joint group.

[0089] The shoulder dual-joint group uses the context of joints 1 to 7, the motor states of motors 1 to 3, the target joint command, the tracking error, the motor torque estimate, the joint velocity sign, and the absolute value of the joint velocity as input features.

[0090] The intermediate joint group uses the context of joints 1 through 4 and the motor states of motors 1 through 3 as input features.

[0091] The wrist single-joint group, wrist coupled group, and distal joint group used the context and motor state associated with the 4th, 5th, 6th, and 7th joints, respectively, as input features.

[0092] Based on the above division, different torque residual groups use different context ranges and different characteristic intensities.

[0093] In some embodiments, different context joint ranges, local motor states, timing window lengths, hidden layer dimensions, or loss functions can be set for different torque residual groups. The timing window for the shoulder dual-joint group can be set to 35, the hidden layer dimension can be set to 96, and the Huber loss function can be used; the timing window for the intermediate joint group can be set to 30, and the hidden layer dimension can be set to 80; the wrist single-joint group, wrist coupled group, and distal joint group use the context and motor states associated with the group and adjacent wrist joints, and the timing window can be set to 25.

[0094] Group modeling has three advantages. First, it avoids the torque residuals of small joints being overwhelmed by the large torque error in the shoulder, making the loss scale of each torque residual group more matched. Second, it allows for different inputs to be configured for different mechanical structures, with the torque residuals in the shoulder relying more on the overall arm context and error compensation, and the torque residuals in the wrist relying more on the local motor state and adjacent coupling relationships. Third, it facilitates start-up, shutdown, diagnosis, and degradation according to torque residual groups during actual machine deployment.

[0095] The grouped moment residual model of this invention uses a GRU network. The grouped moment residual model outputs the moment residual estimate of the joint corresponding to each group, expressed as follows:

[0096] in, This represents the torque residual estimate; Represents the grouped moment residual model; Indicates from the first sampling time to the... Input features of the torque residual model at each sampling time.

[0097] The torque residual estimate is added to the motor torque estimate to obtain the torque feedforward estimate, expressed as:

[0098] in, This indicates torque feedforward estimation; This indicates an estimate of the motor torque. This indicates the torque residual estimate.

[0099] Figure 6 The process of processing the grouped moment residual model was also demonstrated.

[0100] The training results of the grouped moment residual model of this invention are as follows: The root mean square error (RMS) of the rigid body main torque network relative to the motor torque is approximately 0.3394. After superimposing the torque residual estimation, the RMS of the torque feedforward estimation is approximately 0.2695, representing an overall improvement of approximately 20.59%. Among these improvements, the improvement rates for joints 4, 5, 6, and 7 are approximately 23.49%, 25.53%, 19.36%, and 21.67%, respectively.

[0101] To quantify the technical effect of heterogeneous grouping based on anatomical location, transmission structure, and error source, the training summary of this invention simultaneously outputs the error when only the whole-arm rigid trunk torque network is enabled, and the error after superimposing the grouped torque residual model. The above-mentioned improvement results for the wrist and distal end-related 4th, 5th, 6th, and 7th joints demonstrate that local context and adjacent coupling characteristics have a significant effect on torque residual compensation for small inertia joints.

[0102] In further implementation, the system outputs an ablation comparison table for each torque residual group. This table includes four types of results: those using only the whole-arm rigid body backbone torque network, those using a unified whole-arm torque residual model, those using a unified local torque residual model, and those using the grouped torque residual model of this invention. For the shoulder dual-joint group, the focus is on comparing the influence of the whole-arm context and local context on the root mean square error of the high-load joints; for the wrist coupling group, the focus is on comparing the influence of joint input and joint output versus the independent torque residual of a single joint on the root mean square error of the coupled joints. These comparisons demonstrate that heterogeneous grouping is determined by the robot's anatomical location, transmission structure, and error sources.

[0103] The static transmission mapping model and the dynamic angle residual model restore the joint state; the rigid body backbone torque network of the whole arm, excluding target joint commands and tracking errors, learns transferable gravity, inertia, and coupling feedforward; the grouped torque residual model then incorporates compensation information such as target joint commands, tracking errors, and motor torque estimation according to the joint structure; the impedance feedback model performs small-amplitude feedback fitting with the residual torque as the target. The synergy between the above models enables the model to absorb the fitting ability of data-driven methods for nonlinearity and hysteresis, while retaining the physical meaning, fault location interface, and real-time degradation path of each model.

[0104] In this step, the grouped torque residual models focus on learning the effects of local flexibility, friction, hysteresis, and residual effects of the servo inner loop. The sum of the motor torque estimate and the torque residual estimate forms the torque feedforward estimate, which is used for impedance feedback model fitting in subsequent steps.

[0105] Step S6: Fit the model based on the training dataset and the torque feedforward estimation to obtain the impedance feedback model; the impedance feedback model can obtain the impedance feedback torque from the input torque feedforward estimation. The torque feedforward estimation represents the sum of the motor torque estimation and the torque residual estimation; In practical force control, torque feedforward estimation cannot completely eliminate tracking errors and external disturbances. Therefore, an impedance feedback model is set after torque feedforward estimation. The fitting objective is the difference between the motor torque and the torque feedforward estimation, i.e., the residual torque that the torque feedforward estimation cannot explain. The expressions for joint error and joint speed error are:

[0106] in, Indicates joint error; Indicates the target joint angle in the target joint command; This indicates the joint speed error, expressed in radians per second. This indicates the target joint velocity in the target joint command.

[0107] The fitted expression for the impedance feedback torque is:

[0108] in, This indicates the impedance feedback torque. Indicates the stiffness coefficient. Indicates the damping coefficient. This indicates the bias term.

[0109] To ensure the gain is physically reasonable, stiffness coefficients, damping coefficients, and bias terms can be fitted in the joint impedance body coordinates using least squares with non-negative constraints or non-negative least squares methods.

[0110] In some embodiments, mapping can be performed based on the sign relationship between the impedance feedback torque direction of different joints and the positive direction of the joint. In this embodiment, the sign of the impedance feedback torque direction of joints 3, 4, and 7 is set to -1, and the sign of the other joints is set to +1.

[0111] The fitting results of the impedance feedback model can be exported as JSON or CSV format and loaded directly during real-time control.

[0112] The impedance feedback model provided by this invention provides small, stable, and interpretable error feedback based on the basic accuracy of the torque feedforward estimation.

[0113] In some embodiments, if more compliant control is required, it can be adjusted... and Apply scaling; if higher trajectory accuracy is required, the feedback gain can be increased within the limits allowed by stability. Because , and By fitting the impedance body coordinates and converting them according to the direction signs of the joint coordinates and motor torque coordinates, the impedance feedback model can be compatible with different joint drive directions and multi-stage transmission directions, reducing the risk of incorrect gain sign configuration in actual operation.

[0114] Through this step, the present invention provides an impedance feedback model that can provide small-amplitude, stable, and interpretable error feedback based on torque feedforward estimation. After the direction sign conversion between joint coordinates and motor torque coordinates, the resulting impedance feedback model can be used for real-time control.

[0115] Step S7: For the robotic arm to be controlled, real-time robot operation data is acquired and processed sequentially through the static transmission mapping model, dynamic angle residual model, whole arm rigid body main torque network, grouped torque residual model and impedance feedback model. The obtained impedance feedback torque controls the movement of the robotic arm.

[0116] In this step, the real-time controller loads the static transmission mapping model, the dynamic angle residual model, the main torque network of the whole arm rigid body, the grouped torque residual model, and the impedance feedback model. After acquiring real-time robot operation data each time, it is processed in the order of the above models to obtain the real-time torque used to control the movement of the robotic arm.

[0117] The real-time controller provides a reset interface and a single-step interface. Each single-step interface receives the latest real-time robot operation data and target joint commands. The real-time robot operation data includes at least the current joint angle, joint velocity, joint acceleration, motor position, motor velocity, motor acceleration, optional motor torque, time, or sampling interval. The target joint commands include at least the target joint angle, and optionally also the target joint velocity and target joint acceleration.

[0118] In some embodiments, recent historical data can be stored in a circular historical buffer, with a buffer length not less than the maximum time window and derivative smoothing window of each GRU network.

[0119] Figure 7 This invention demonstrates the real-time control process provided by the present invention. The specific processing steps include: First, during each real-time processing step, the sensor fields are normalized to field names consistent with the training dataset; then, a static transmission mapping model is executed to obtain static joint angle estimates; next, a dynamic angle residual model is executed to obtain angle residuals and corrected joint angle estimates; subsequently, joint velocity, joint acceleration, sine characteristics, and cosine characteristics corresponding to the corrected joint angle estimates are constructed, and motor torque estimates are obtained through the rigid body backbone torque network of the entire arm; after satisfying the time window length of each torque residual group, a grouped torque residual model is executed to obtain torque residual estimates; the motor torque estimate is added to the torque residual estimate to obtain the torque feedforward estimate; finally, the impedance feedback torque is obtained through an impedance feedback model.

[0120] In some embodiments, torque limiting, torque rate of change limiting, non-numerical and infinite value handling strategies, whether to enable grouped torque residual models, whether to enable dynamic angle residual models, and feature missing checks can be configured. When the length of the annular history buffer is insufficient, static joint angle estimation or the current joint angle is temporarily used; when the time window length of the grouped torque residual model is insufficient, the stress torque residual estimation is set to zero.

[0121] In some embodiments, when exporting the model, the following files are saved: static transmission mapping model parameter file, dynamic angle residual model file, whole-arm rigid body main moment network file, grouped moment residual model file, impedance feedback model parameter file, unified field specification, feature list, training summary, and offline playback results. The static transmission mapping model parameter file is in JSON format, the dynamic angle residual model file, whole-arm rigid body main moment network file, and grouped moment residual model file are in PT format, and the impedance feedback model parameter file is in JSON or CSV format. The static transmission mapping model parameter file saves the input joints, output joints, feature columns, mean, variance, polynomial power, regression coefficients, and intercept; each network model file saves the network weights, network structure, standardized parameters, input feature list, output target list, and corresponding time window length.

[0122] like Figure 3 As shown, the training output is first subjected to integrity verification to confirm that the static transmission mapping model, the dynamic angle residual models of each configuration group, the main torque network of the rigid body of the whole arm, the grouped torque residual models of each torque residual group, and the parameters and field specifications of the impedance feedback model all exist. Then, it is packaged into a real machine deployment file. After the real machine control program loads the deployment file, the real-time robot running data is filtered by state and enters each model for inference through a circular history buffer. The data is then processed by torque limiting, torque change rate limiting, and processing of non-numerical and infinite values ​​to form a safe real-time torque output.

[0123] The engineering value of this invention's real-time control lies in the following: During the initial setup phase or when sensors are temporarily unavailable, the system can degenerate into a control path that adds the estimated motor torque output from the rigid body main torque network of the entire arm to the impedance feedback torque, avoiding the unavailability of models using timing windows due to insufficient historical data; once sufficient historical data is available, the dynamic angle residual model and grouped torque residual model are automatically activated to improve the accuracy of corrected joint angle estimation and torque feedforward estimation; and the system can return to its previous state during debugging. , , The system provides joint error, residual availability flags, and missing feature reports to help robot control programs, upper-level controllers, or intelligent strategies determine the availability status of the lower-level model.

[0124] like Figure 1 As shown, the graphical interface for raw data processing, preprocessing filtering, and curve preview of this invention is illustrated. The workflow of the graphical interface includes the raw data processing, preprocessing filtering, and curve preview process.

[0125] This invention also provides a control system for a robotic arm used in rope-driven and quasi-direct-drive robots, comprising: The acquisition and preprocessing module is used to acquire robot operation data and preprocess it to obtain a training dataset. The robot has a robotic arm with multiple joints. The data in the training dataset includes the motor state, joint state, target joint command, and motor torque of the robotic arm. The motor state includes motor position, motor speed, and motor acceleration, and the joint state includes joint angle, joint speed, and joint acceleration. The static drive mapping module is used to perform ridge regression processing based on the training dataset to obtain a static drive mapping model; the static drive mapping model can obtain static joint angle estimates from the input motor position; The dynamic angle residual module is used to train a dynamic angle residual model based on the training dataset and static joint angle estimation. The dynamic angle residual model adopts a GRU network and can obtain the angle residual from the input motor state and static joint angle estimation. The rigid body main torque network module is used to train the rigid body main torque network of the whole arm based on the training dataset and the corrected joint angle estimation of some joints. The rigid body main torque network of the whole arm adopts a multilayer perceptron and can obtain the motor torque estimation from the input corrected joint angle. The corrected joint angle estimate of the partial joint is obtained by adding the static joint angle estimate of the partial joint to the angle residual; The grouped torque residual module is used to train a grouped torque residual model based on the training dataset, the corrected joint angle estimate, and the motor torque estimate. The grouped torque residual model can obtain the torque residual estimate from the input corrected joint angle estimate and motor torque estimate. The impedance feedback module is used to fit the training dataset and the torque feedforward estimate to obtain the impedance feedback model; the impedance feedback model can obtain the impedance feedback torque from the input torque feedforward estimate. The torque feedforward estimation represents the sum of the motor torque estimation and the torque residual estimation; The control module is used to acquire real-time robot operation data for the robotic arm to be controlled. The data is then processed sequentially through the static transmission mapping model, dynamic angle residual model, whole arm rigid body main torque network, grouped torque residual model, and impedance feedback model. The obtained impedance feedback torque controls the movement of the robotic arm.

[0126] This invention can also serve as a low-level ontological force control model or a motor-level forward and inverse dynamics interface for an embodied intelligent agent. Upper-level model predictive control, reinforcement learning strategies, or vision-language-action strategies transmit target joint commands through a unified target state interface. The real-time controller generates torque feedforward estimates, impedance feedback torques, and real-time torques based on real-time robot operation data, a circular history buffer, a static transmission mapping model, a dynamic angle residual model, a rigid body main torque network, a grouped torque residual model, and an impedance feedback model. It then returns residual availability flags and missing feature reports. Thus, the upper-level strategy drives the roped, collimated, or hybrid transmission robot through a unified low-level interface. Rope drive hysteresis, collimated drive elasticity, and multi-stage transmission errors are compensated by the aforementioned models. When the history length is insufficient or input features are missing, the real-time controller maintains the low-level execution by adding the motor torque estimate and the impedance feedback torque according to the aforementioned degradation path.

[0127] Without departing from the technical concept of this invention, the static transmission mapping model can be replaced by multinomial ridge regression with spline regression, Gaussian process, radial basis function network or small multilayer perceptron; the GRU network in the dynamic angle residual model and grouped torque residual model can be replaced with LSTM, TCN, Transformer Encoder, one-dimensional convolutional network or state space model; the rigid body backbone torque network of the whole arm can use the output of the analytical rigid body dynamics model as a priori features, but still maintain the isolation between the input of the network and the target joint command and tracking error; the impedance feedback model can be fitted using recursive least squares, Bayesian optimization, constrained quadratic programming or online adaptive algorithm; the configuration group and torque residual group can be determined based on the robot's degrees of freedom, rope structure, actuator layout, error correlation or data clustering results.

[0128] This invention is also applicable to robots with right arms, both arms, legs, waists, dexterous hands, or multi-branched systems. For robots without joint-side encoders, joint angles can be inferred using vision, inertial navigation, external motion capture, model observers, or force sensors as supervisory data in the training dataset. For robots with contact tasks, end-effector forces, wrist triaxial forces, haptic array data, or environmental contact states can be added to the input of the dynamic angle residual model or grouped torque residual model to extend to contact force control and compliant manipulation tasks. In embodied intelligence applications, this invention can provide a unified target joint command input, real-time torque output, and residual availability state for model predictive control, reinforcement learning closed-loop training, cross-entity transfer, domain randomization, and online continuous learning.

[0129] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A control method for a robotic arm in rope-driven and quasi-direct-drive robots, characterized in that, Includes the following steps: Step S1: Collect robot operation data and preprocess it to obtain a training dataset; the robot has a robotic arm with multiple joints; the data in the training dataset includes the motor status, joint status, target joint commands, and motor torque of the robotic arm; Motor status includes motor position, motor speed, and motor acceleration; joint status includes joint angle, joint speed, and joint acceleration. Step S2: Perform ridge regression processing based on the training dataset to obtain the static transmission mapping model; The static transmission mapping model can obtain the static joint angle estimate from the input motor position; Step S3: Based on the training dataset and static joint angle estimation, a dynamic angle residual model is trained, which can obtain the angle residual from the input motor state and static joint angle estimation. Step S4: Based on the training dataset and the corrected joint angle estimation of some joints, the rigid body main torque network of the whole arm is trained. The rigid body main torque network of the whole arm adopts a multilayer perceptron and can obtain the motor torque estimation from the input corrected joint angle. The corrected joint angle estimate of the partial joint is obtained by adding the static joint angle estimate of the partial joint to the angle residual; Step S5: Based on the training dataset, the corrected joint angle estimate and the motor torque estimate are used to train and obtain the grouped torque residual model. The grouped torque residual model can obtain the torque residual estimate from the input corrected joint angle estimate and motor torque estimate. Step S6: Fit the model based on the training dataset and the torque feedforward estimation to obtain the impedance feedback model; The impedance feedback model can estimate the impedance feedback torque from the input torque feedforward. The torque feedforward estimation represents the sum of the motor torque estimation and the torque residual estimation; Step S7: For the robotic arm to be controlled, real-time robot operation data is acquired and processed sequentially through the static transmission mapping model, dynamic angle residual model, whole arm rigid body main torque network, grouped torque residual model and impedance feedback model. The obtained impedance feedback torque controls the movement of the robotic arm.

2. The control method for a robotic arm for rope-driven and quasi-direct-drive robots according to claim 1, characterized in that, In step S1, the robotic arm has 7 joints, the first to the seventh joints being: arm raising joint, arm extending joint, shoulder twisting joint, elbow flexion joint, wrist turning A joint, wrist turning B joint, and wrist twisting joint; each joint has a motor, and the first to the seventh joints correspond to the first to the seventh motors respectively.

3. The control method for a robotic arm for rope-driven and quasi-direct-drive robots according to claim 2, characterized in that, In step S1, the step of collecting robot operation data and preprocessing it to obtain a training dataset specifically includes: (1) Parse the original text log, calculate the cumulative number of seconds within the day, and generate a sampling interval based on the difference between adjacent time periods, which is used as a timestamp; (2) For the first and second joints, the motor position is obtained from the motor side, and the motor speed and acceleration are obtained by differentiation to form the motor state; the motor position is used as the joint angle, and the joint speed and joint acceleration are determined by differentiation to form the joint state. For the third joint, the motor state is obtained, and the joint angle is calculated by converting the motor position using a fixed ratio and direction sign, thus obtaining the joint state; For joints 4 to 7, obtain the motor status, decode the joint angles using the joint-side decoder, and then obtain the joint status. Read the target joint commands and motor torque from the raw text log; (3) Perform low-pass filtering on the timestamped motor state, joint state, target joint command and motor torque sequence to form the training dataset.

4. The control method for a robotic arm for rope-driven and quasi-direct-drive robots according to claim 3, characterized in that, Step S2 specifically includes: (1) The joint structure of the robotic arm is divided into configuration groups, including: single joint configuration group, coupled joint configuration group and end joint configuration group; In this embodiment, the single-joint configuration group uses the position of the fourth motor to predict the static joint angle estimate of the fourth joint; the coupled joint configuration group uses the positions of the fifth and sixth motors to jointly predict the static joint angle estimate of the fifth and sixth joints; and the end-joint configuration group uses the position of the seventh motor to predict the static joint angle estimate of the seventh joint. (2) Normalize the mean and variance of each motor position, construct first to fourth order candidate polynomial features, and perform ridge regression based on the first to fourth order candidate polynomial features. Among the multiple ridge regression regularization coefficients obtained, the mapping function corresponding to the ridge regression regularization coefficient with the smallest verification error is selected as the static transmission mapping model. The expression for the static transmission mapping model is: in, This represents the static joint angle estimation. Indicates the motor position; This represents a static transmission mapping model.

5. The control method for a robotic arm for rope-driven and quasi-direct-drive robots according to claim 4, characterized in that, In step S3, the dynamic angle residual model uses a GRU network, and its expression is: Indicates angular residual; Represents a dynamic angle residual model; Indicates from the first sampling time to the... Input features at each sampling time; Indicates the length of the timing window.

6. The control method for a robotic arm for rope-driven and quasi-direct-drive robots according to claim 5, characterized in that, In step S4, the corrected joint angle estimation input to the rigid body main moment network of the whole arm specifically includes: The measured joint angles are directly used as the corrected joint angle estimates for joints 1 to 3; The corrected joint angle expression for joints 4 to 7 is: in, This indicates a correction to the joint angle estimation.

7. The control method for a robotic arm for rope-driven and quasi-direct-drive robots according to claim 6, characterized in that, In step S4, the expression for the main moment network of the rigid body of the entire arm is: in, This indicates an estimate of the motor torque. The mapping function representing the main moment network of the rigid body of the entire arm; This indicates the input state of the main torque network; and Let represent the first and second derivatives corresponding to the input states of the main torque network, respectively. and The sine and cosine characteristics of the joint angles are represented.

8. The control method for a robotic arm for rope-driven and quasi-direct-drive robots according to claim 7, characterized in that, In step S5, the input features of the grouped torque residual model come from multiple torque residual groups, specifically including: shoulder double joint group, intermediate joint group, wrist single joint group, wrist coupling group and distal joint group; The shoulder dual-joint group uses the context of joints 1 to 7, the motor states of motors 1 to 3, the target joint command, the tracking error, the motor torque estimate, the joint velocity sign, and the absolute value of the joint velocity as input features; The intermediate joint group uses the context of joints 1 to 4 and the motor states of motors 1 to 3 as input features; The wrist single joint group, wrist coupled group, and distal joint group used the context and motor state related to the 4th, 5th, 6th, and 7th joints as input features, respectively. In step S5, the expression for the grouped moment residual model is: in, This represents the torque residual estimate; Represents the grouped moment residual model; Indicates from the first sampling time to the... Input features of the torque residual model at each sampling time; This indicates torque feedforward estimation.

9. The control method for a robotic arm for rope-driven and quasi-direct-drive robots according to claim 8, characterized in that, In step S6, the expression for the impedance feedback model is: in, This indicates the impedance feedback torque. Indicates the stiffness coefficient. Indicates the damping coefficient. Indicates the bias term. Indicates the target joint angle in the target joint command; This indicates the target joint velocity in the target joint command.

10. A control system for a robotic arm used in rope-driven and quasi-direct-drive robots, characterized in that, include: The data acquisition and preprocessing module is used to collect robot operation data and preprocess it to obtain a training dataset. The robot has a robotic arm with multiple joints; The training dataset includes the motor status, joint status, target joint commands, and motor torque of the robotic arm; Motor status includes motor position, motor speed, and motor acceleration; joint status includes joint angle, joint speed, and joint acceleration. The static transmission mapping module is used to perform ridge regression processing based on the training dataset to obtain a static transmission mapping model. The static transmission mapping model can obtain the static joint angle estimate from the input motor position; The dynamic angle residual module is used to train a dynamic angle residual model based on the training dataset and static joint angle estimation. It can obtain the angle residual from the input motor state and static joint angle estimation. The rigid body main torque network module is used to train the rigid body main torque network of the whole arm based on the training dataset and the corrected joint angle estimation of some joints. The rigid body main torque network of the whole arm adopts a multilayer perceptron and can obtain the motor torque estimation from the input corrected joint angle. The corrected joint angle estimate of the partial joint is obtained by adding the static joint angle estimate of the partial joint to the angle residual; The grouped torque residual module is used to train a grouped torque residual model based on the training dataset, the corrected joint angle estimate, and the motor torque estimate. The grouped torque residual model can obtain the torque residual estimate from the input corrected joint angle estimate and motor torque estimate. The impedance feedback module is used to fit the training dataset and the torque feedforward estimate to obtain the impedance feedback model. The impedance feedback model can estimate the impedance feedback torque from the input torque feedforward. The torque feedforward estimation represents the sum of the motor torque estimation and the torque residual estimation; The control module is used to acquire real-time robot operation data for the robotic arm to be controlled, and process it sequentially through the static transmission mapping model, dynamic angle residual model, whole arm rigid body main torque network, grouped torque residual model and impedance feedback model, and control the movement of the robotic arm by the obtained impedance feedback torque.