Flexible cable automatic wiring method and system

Abstract

The invention discloses a flexible cable automatic wiring method and system, and relates to the technical field of automatic control and intelligent manufacturing. The method comprises the following steps: collecting multi-dimensional physical information; generating a current cable state vector through a multi-modal fusion recognition model; a current optimal action instruction is generated in an adaptive strategy optimization network based on deep reinforcement learning; issuing the current optimal action instruction to a motion controller of the wiring robot so as to control an actuator of the wiring robot to execute corresponding actions; and calculating an instant reward based on the new state of the cable after the wiring robot executes the corresponding action, storing the original state, the adopted action, the instant reward and the new state in the execution process into an experience playback database, and returning to the step of generating a current optimal action instruction until the wiring task is completed. According to the invention, automatic wiring adaptive adjustment and continuous evolution of the flexible cable are realized, and the success rate, the safety and the intelligent level of the wiring system are improved.

Description

Technical Field

[0001] This invention relates to the field of automation control and intelligent manufacturing technology, specifically to a method and system for automatic wiring of flexible cables. Background Technology

[0002] As industrial equipment evolves towards higher density and modularity, the complexity of cabling has increased significantly. Traditional manual cabling is inefficient and inconsistent, making automated cabling technology an important development direction. While some automated cabling systems already employ visual positioning and trajectory planning technologies, the following prominent issues still exist in practical applications:

[0003] (1. Insufficient force control accuracy: Most systems rely on preset force thresholds or fixed impedance parameters, which cannot respond to dynamic force changes during the plugging process in real time, which can easily lead to cable bending, socket damage or plugging failure.

[0004] (2. Poor adaptability: Different cables vary significantly in terms of material, diameter, and flexibility, making it difficult to generalize the impedance model with fixed parameters to diverse cabling tasks.)

[0005] (3. Lack of learning ability: The existing system does not have the ability to learn from historical tasks and optimize control parameters. It needs to be re-adjusted when facing new cables or new environments, and its level of intelligence is limited.

[0006] Although existing literature has proposed wiring methods based on machine learning or data, most of them focus on path planning or visual recognition. There is no evidence of applying online learning mechanisms to the real-time optimization of impedance model parameters to achieve dynamic compliant control and adaptive adjustment in the flexible cable splicing process. Summary of the Invention

[0007] This invention addresses the problems of unstable wiring quality and easy damage caused by the dynamic and unpredictable changes in the physical state of cables (such as tension, torsion, and curvature) in existing wiring technologies. It provides an automatic wiring method and system for flexible cables, which realizes the adaptive adjustment and continuous evolution of automatic wiring of flexible cables, thereby improving the success rate, security and intelligence level of the wiring system.

[0008] The present invention is achieved through the following technical solution.

[0009] In a first aspect, an automatic wiring method for flexible cables is provided, the method comprising:

[0010] Multiple sensing modules installed on the wiring robot collect multi-dimensional physical information at a preset frequency;

[0011] The multidimensional physical information is processed, and the processed multidimensional physical information is input into the multimodal fusion recognition model to generate the current cable state vector;

[0012] The current cable state vector is input into an adaptive policy optimization network based on deep reinforcement learning to generate the current optimal action command, wherein the current optimal action command includes commands for impedance control, adaptive compliance adjustment and force-position hybrid control.

[0013] The current optimal action command is sent to the motion controller of the wiring robot, so that the actuator of the wiring robot can be controlled by the motion controller to perform the corresponding action;

[0014] The instant reward is calculated based on the new state of the cable after the cabling robot performs the corresponding action. The original state, the action taken, the instant reward, and the new state during this execution process are stored in the experience replay database. The process returns to the step of: inputting the current cable state vector into the deep learning-based adaptive policy network to generate the current optimal action instruction until the cabling task is completed.

[0015] In some embodiments, multiple sensing modules mounted on a wiring robot collect multidimensional physical information at a preset frequency, including:

[0016] The tension of the cable along its length is collected by a distributed micro-tension sensor array mounted on the cabling robot.

[0017] Cable images are acquired using a multi-view vision system mounted on the wiring robot;

[0018] High-frequency stress waves and audible sound waves are collected by an acoustic sensing module installed on the wiring robot.

[0019] The contact pressure between the wiring robot and the cable is collected by a tactile sensor array installed on the wiring robot.

[0020] In some embodiments, the multidimensional physical information is processed, and the processed multidimensional physical information is input into a multimodal fusion recognition model to generate a current cable state vector, including:

[0021] An algorithm based on hardware timestamps and software interpolation is used to align the multidimensional physical information in time and space.

[0022] For the spatiotemporally aligned multidimensional physical information, a feature extraction algorithm is run in parallel to obtain the high-dimensional feature vector of the multidimensional physical information;

[0023] The high-dimensional feature vector is input into the multimodal fusion recognition model to generate the current cable state vector, wherein the multimodal fusion recognition model is a multi-head encoding-fusion decoding neural network.

[0024] In some embodiments, the current cable state vector is input into an adaptive policy optimization network based on deep reinforcement learning to generate the current optimal action instruction, including:

[0025] The current cable state vector is obtained in the Markov environment and input into the policy optimization network to obtain the action probability distribution output by the policy optimization network and the current optimal action instruction based on the action probability distribution.

[0026] In some embodiments, the method further includes: calculating an immediate reward based on the new state of the cable, and storing the original state, the action taken, the immediate reward, and the new state during the current execution process into an experience replay database, then using a digital twin simulation environment and a near-end policy optimization algorithm to train the policy network and value network of the policy optimization network, wherein training the policy network and the value network includes:

[0027] In a digital twin simulation environment, training data is collected through extensive trial and error and interaction with the environment. The training data includes: the original state, the actions taken, the immediate reward, and the new state.

[0028] The initial policy network and initial value network are trained using a proximal policy optimization algorithm.

[0029] Deploy the pre-trained policy network onto the wiring robot;

[0030] The policy network generates action information based on the actual cable status.

[0031] The wiring robot performs actions based on the motion information, collects new experience, and stores it in the experience replay buffer pool.

[0032] During idle or low-load periods, small batches of data are sampled from the buffer pool to fine-tune the policy network and the value network online, adapting them to the differences between the real world and the simulation.

[0033] The accumulated large amount of real data is regularly used for offline reinforcement learning or batch retraining to further optimize the policy network and value network.

[0034] In some embodiments, the current optimal action command is sent to the motion controller of the wiring robot, so that the motion controller controls the actuator of the wiring robot to perform the corresponding action, including:

[0035] The current optimal action instruction is returned to the Markov environment, and the actuator of the wiring robot is controlled by the motion controller to perform the corresponding action based on the current optimal action instruction.

[0036] In some embodiments, the motion controller includes: an adaptive motion controller, a compliant gripping and tension management unit, and a safety monitoring and emergency stop module, wherein,

[0037] The adaptive motion controller is used to receive motion commands from the policy network and solve them into desired trajectory points in the robot joint space or Cartesian space. The adaptive motion controller embeds a feedforward-feedback composite control law. The feedforward part is based on the robot dynamics model to compensate for nonlinear coupling, and the feedback part is used to suppress model errors and external disturbances.

[0038] The compliant gripping and tension management unit is used to control the gripping force of the end gripper of the wiring robot;

[0039] The safety monitoring and emergency stop module is used for real-time monitoring, setting multiple safety thresholds, and triggering an emergency stop or a preset safety rollback action when an unrecoverable abnormal state is detected.

[0040] In some embodiments, the method further includes: monitoring operational anomalies, and adopting a tiered response strategy for the detected anomalies, wherein adopting the tiered response strategy includes:

[0041] Policy layer adaptive adjustment: Common and non-urgent state anomalies are handled by the policy network;

[0042] Rule-based rapid intervention: Set a safety threshold. When the state quantity exceeds the safety threshold, immediately interrupt the output of the policy network and switch to a preset, rule-based safety controller.

[0043] Hardware-level emergency stop: In the event of an extreme danger signal, the hardware emergency stop circuit of the wiring robot is triggered.

[0044] Secondly, an automatic flexible cable routing system is provided, the system comprising:

[0045] The multimodal sensing module is used to: collect multidimensional physical information at a preset frequency through multiple sensing modules installed on the wiring robot;

[0046] The state recognition module is used to: process the multidimensional physical information and input the processed multidimensional physical information into the multimodal fusion recognition model to generate the current cable state vector;

[0047] The decision module is used to: input the current cable state vector into a deep learning-based adaptive policy network to generate the current optimal action command, wherein the current optimal action command includes commands for impedance control, adaptive compliance adjustment and force-position hybrid control;

[0048] The execution and control module is used to: send the current optimal action command to the motion controller of the wiring robot, so as to control the actuator of the wiring robot to perform the corresponding action through the motion controller;

[0049] The learning and evolution module is used to: calculate the immediate reward based on the new state of the cable, and store the original state, the action taken, the immediate reward, and the new state during this execution process into the experience replay database, and return it to the state recognition module.

[0050] In some embodiments, the execution and control module includes: an adaptive motion controller, a compliant gripping and tension management unit, and a safety monitoring and emergency stop module, wherein,

[0051] The adaptive motion controller is used to receive motion commands from the policy network and solve them into desired trajectory points in the robot joint space or Cartesian space. The adaptive motion controller embeds a feedforward-feedback composite control law. The feedforward part is based on the robot dynamics model to compensate for nonlinear coupling, and the feedback part is used to suppress model errors and external disturbances.

[0052] The compliant gripping and tension management unit is used to control the gripping force of the end gripper of the wiring robot;

[0053] The safety monitoring and emergency stop module is used for real-time monitoring, setting multiple safety thresholds, and triggering an emergency stop or a preset safety rollback action when an unrecoverable abnormal state is detected.

[0054] Compared with existing technologies, this invention has the following advantages and beneficial effects: By constructing a closed-loop wiring control scheme that integrates high-precision perception, intelligent decision-making, adaptive learning, and precise execution, it solves the core problems in the automated wiring of flexible cables, such as unstable wiring quality and easy cable damage caused by the dynamic and unpredictable changes in the physical state of the cables (such as tension, torsion, and curvature). It comprehensively perceives the cable state through a multimodal sensor network and dynamically adjusts the robot's operation strategy using advanced learning algorithms, realizing a paradigm shift from "blind operation" to "state perception-intelligent decision-making". Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1This is a flowchart of an automatic wiring method for flexible cables according to an embodiment of the present invention.

[0057] Figure 2 The overall process of automatic flexible cable routing according to an embodiment of the present invention is shown.

[0058] Figure 3 The key process of flexible cable conformal insertion trajectory planning and real-time control according to an embodiment of the present invention is illustrated.

[0059] Figure 4 The overall control architecture of a flexible cable automatic cabling system according to an embodiment of the present invention is shown.

[0060] Figure 5 This is a structural diagram of an automatic flexible cable routing system according to an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0062] In this invention, a closed-loop control system for cabling is constructed, integrating high-precision perception, intelligent decision-making, adaptive learning, and precise execution. It aims to solve the core problems in the automated cabling of flexible cables, such as unstable cabling quality and easy cable damage caused by the dynamic and unpredictable changes in the physical state of the cables (such as tension, torsion, and curvature). The system comprehensively perceives the cable state through a multimodal sensor network and dynamically adjusts the robot's operation strategy using advanced learning algorithms, thereby realizing a paradigm shift from "blind operation" to "state perception-intelligent decision-making".

[0063] On the one hand, the present invention provides an automatic wiring method for flexible cables. Figure 1 This is a flowchart illustrating an automatic wiring method for flexible cables according to an embodiment of the present invention. (Reference) Figure 1 The automatic wiring method for flexible cables includes S10 to S50.

[0064] In S10, multiple sensing modules installed on the wiring robot collect multi-dimensional physical information at a preset frequency.

[0065] In S20, the multidimensional physical information is processed and then input into the multimodal fusion recognition model to generate the current cable state vector.

[0066] In S30, the current cable state vector is input into a deep learning-based adaptive policy network to generate the current optimal action command, which includes commands for impedance control, adaptive compliance adjustment, and force-position hybrid control.

[0067] In S40, the current optimal action command is sent to the motion controller of the wiring robot, so that the motion controller can control the actuator of the wiring robot to perform the corresponding action.

[0068] In S50, the instant reward is calculated based on the new state of the cable after the wiring robot performs the corresponding action. The original state, the action taken, the instant reward, and the new state during this execution process are stored in the experience playback database. Then, the following steps are returned: the current cable state vector is input into the deep learning-based adaptive policy network to generate the current optimal action instruction until the wiring task is completed.

[0069] The following detailed description of S10 to S50, in conjunction with the accompanying drawings, illustrates the present invention. This invention adopts a modular and hierarchical design concept, primarily comprising the following four core layers: a multimodal perception layer, a state recognition and decision-making layer, an execution and control layer, and a learning and evolution layer, each responsible for performing different functions.

[0070] S10 is implemented by a multimodal perception layer. Multiple sensing modules mounted on the wiring robot collect multidimensional physical information at preset frequencies, including: collecting cable tension along its length using a distributed micro-tension sensor array mounted on the wiring robot; collecting cable images using a multi-view, high-frame-rate vision system mounted on the wiring robot; collecting high-frequency stress waves and audible sound waves using an acoustic sensor module mounted on the wiring robot; and collecting the contact pressure between the wiring robot and the cable using a tactile sensor array mounted on the wiring robot.

[0071] Multiple sensing modules act as the "sensory nerves" of the cabling robot, responsible for real-time, non-invasively collecting multi-dimensional physical information of the cable during the entire process of grasping, moving, and laying. These multiple sensing modules include at least: a distributed micro tension sensor array, a multi-view high frame rate vision system, an acoustic emission / triboacoustic sensor module, and a high-density tactile sensor array.

[0072] A distributed micro-tension sensing array, employing microelectromechanical systems (MEMS) technology, integrates micro-thin-film tension sensors as a mesh-like flexible circuit within the lining of a robot's end effector or as a wearable tag attached to key points on the cable surface. This array synchronously samples at a frequency ≥1kHz to acquire a cross-sectional view of the tension distribution along the cable's length, rather than focusing on single-point forces.

[0073] The multi-view, high-frame-rate vision system consists of multiple industrial cameras mounted at the robot's end effector (eyes on the hand) and at the top of the workspace (global overview). The end effector provides high-resolution images of cable end textures and surface defects; the global cameras reconstruct the cable's centerline shape in three-dimensional space in real time using binocular stereo vision or structured light technology. The vision system maintains a frame rate of at least 100Hz to capture rapid deformations.

[0074] The acoustic emission / triboacoustic sensing module integrates a piezoelectric ceramic acoustic emission sensor inside the robot's gripper to capture high-frequency stress waves (>100kHz) and audible sound signals generated during friction, bending, and scraping between the cable and the gripper, the cable and the guide mechanism, or the environment. This module is extremely sensitive to microscopic damage to the cable (such as microcracks in the insulation layer) and abnormal contact.

[0075] A high-density tactile sensor array covers the contact surface of the gripper with an array (such as 16x16) composed of capacitive or piezoresistive sensing units. It outputs a contact pressure distribution cloud map at a frequency of ≥200Hz to determine whether the grip is uniform, whether the cable has a slipping tendency, and to identify the texture and foreign objects on the cable surface.

[0076] S20 is implemented by the state recognition part in the state recognition and decision layer. It processes the multidimensional physical information and inputs the processed multidimensional physical information into a multimodal fusion recognition model to generate the current cable state vector. This includes: using an algorithm based on hardware timestamps and software interpolation to spatiotemporally align the multidimensional physical information; running a feature extraction algorithm in parallel on the spatiotemporally aligned multidimensional physical information to obtain a high-dimensional feature vector; and inputting the high-dimensional feature vector into the multimodal fusion recognition model to generate the current cable state vector. The multimodal fusion recognition model is a multi-head encoder-fusion decoder neural network.

[0077] For example, an algorithm based on hardware timestamps (such as PTP (Precise Time Protocol)) and software interpolation is used to unify sensor data from different physical locations and sampling rates into the same spatiotemporal coordinate system. Subsequently, feature extraction algorithms are run in parallel for each data mode (such as extracting mean, variance, and gradient from tension profiles; extracting centerline, curvature, and twist angle from visual point clouds; extracting spectral features and energy envelope from acoustic signals; and extracting pressure torque and centroid shift from tactile images), forming high-dimensional feature vectors.

[0078] Based on data acquired by a distributed micro-tension sensor array, tension field reconstruction and feature calculation are performed. Assume n micro-tension sensors are arranged along the length of the cable to be laid, and the reading of the i-th sensor at time t is... The obtained feature values ​​include the following:

[0079] Tension distribution vector: By using spline interpolation, the tension field function can be reconstructed on continuous arc length coordinates s. ;

[0080] Global tension level: This reflects the overall stress on the cable;

[0081] Tension uniformity index: The larger this value, the more uneven the stress on the cable, and the greater the risk of local stress concentration.

[0082] Tension gradient: ,in High gradients can be estimated by the difference between adjacent sensors, and may indicate jamming points or bends.

[0083] A multi-view, high-frame-rate vision system is used for 3D morphological visual perception and geometric feature extraction. Global camera point clouds are utilized. The three-dimensional centerline point sequence of the cable is obtained by using a skeleton extraction-based algorithm (such as the Voronoi diagram method). ,in The obtained feature values ​​include the following:

[0084] Discrete curvature: for three consecutive points on the center line Its approximate curvature is The curvature sequence is obtained. ;

[0085] Maximum curvature / minimum bending radius: , This is a direct indicator for determining whether the minimum bending radius constraint for cables has been violated;

[0086] Torsion Angle Estimation: Define an accompanying Frenet frame on the centerline or calculate the cumulative torsion angle of the cable around its centerline based on cable surface texture tracing.

[0087] Acoustic signal processing and anomaly detection are performed using an acoustic sensing module. (Original acoustic emission signal) and audio signals It has undergone bandpass filtering and noise reduction preprocessing. The obtained feature values ​​include the following:

[0088] Time-frequency domain features: Perform STFT (Short-Time Fourier Transform) on a short-segment signal to extract the Mel-frequency cepstral coefficients (MFCC), spectral centroid, zero-crossing rate, etc., forming a feature vector.

[0089] Event detection: Set an energy threshold to detect sudden energy events in the acoustic emission signal, and set the count rate. and average energy It can serve as a proxy indicator for friction or micro-damage.

[0090] Tactile image analysis and contact state determination are performed using a tactile sensor array. The tactile sensor array outputs a pressure matrix. The obtained feature values ​​include the following:

[0091] Grip stability characteristics: Calculation of pressure torque (Total pressure) (Coordinates of the center of mass of pressure);

[0092] Centroid offset velocity: This is a key indicator for determining whether a cable is slipping.

[0093] Contact pattern classification: using a lightweight CNN Real-time classification is performed to determine the contact mode as uniform gripping, edge contact, point contact, or detection of surface protrusions / foreign objects.

[0094] Then, a deep learning-based multimodal fusion recognition model generates the current cable state vector. Specifically, a multi-branch neural network is constructed. Each branch (visual branch, tension branch, acoustic branch, tactile branch) encodes the raw or feature data, and then the data is fused at the feature level or decision level in the fusion layer. The model ultimately outputs a structured cable state vector S, along with a state classification label (e.g., {State: Normal, Substate: Slight Bending, Risk Level: Low}). This model is pre-trained offline using a large amount of labeled data and supports online fine-tuning after deployment.

[0095] For example, the multimodal fusion recognition model can be a neural network architecture of multi-head encoder-fusion decoder, whose encoder branches include: tension branch, vision branch, acoustic branch, and tactile branch.

[0096] Tension branch: Wait for the input to be a fully connected network (FCN).

[0097] Visual branch: Wait for another FCN to be input, or directly input the centerline point cloud. Input a PointNet network.

[0098] Acoustics branch: The input is a one-dimensional convolutional neural network (1D-CNN).

[0099] Tactile branch: Integrating tactile images or extracted features Wait for the input to be a small CNN or FCN.

[0100] The fusion layer concatenates the high-dimensional feature vectors encoded from each branch, and then performs deep fusion through several fully connected network layers to learn the correlation between modalities.

[0101] Decoding / Output Layer: Outputs a continuous state vector S(t), for example, S(t)=[s1,s2,...,sd], where each dimension can represent the normalized tension level, degree of bending, degree of torsion, risk of slippage, etc.

[0102] State regression head: can output state class probability distributions in parallel.

[0103]

[0104] Supervised learning training of the multimodal fusion recognition model can be performed using a large amount of labeled data collected in simulation and real-world scenarios. The loss function is a weighted sum of regression loss (such as MSE) and classification loss (such as cross-entropy):

[0105] ;

[0106] in, Total loss value: The goal of model training is to minimize this value so that the model's predictions are as close to the true situation as possible;

[0107] The weighting coefficients for regression loss. Used to adjust the proportion of regression error in the total loss;

[0108] : Predicted cable state vector;

[0109] : The ground truth value of the cable state vector. This is the actual physical state value from the labeled data, serving as the "standard answer" for training;

[0110] The weighting coefficients for classification loss. Used to adjust the proportion of classification error in the total loss;

[0111] Cross-entropy loss function. Commonly used to measure the difference between the predicted probability distribution and the true label distribution in classification tasks;

[0112] The predicted probability distribution for each state category. The probability output by the model for each state category;

[0113] : The truth label for the category, indicating the actual category to which the cable marked in the data belongs.

[0114] S30 corresponds to the decision-making part of the state recognition and decision-making layer. It obtains the current cable state vector in the Markov environment and inputs the current cable state vector into the policy optimization network to obtain the action probability distribution output by the policy optimization network and the current optimal action instruction based on the action probability distribution, thereby realizing the mapping from "perception" to "decision".

[0115] The reinforcement learning-based adaptive policy optimizer (Adaptive Policy Optimization Network) models the cabling process as a Markov Decision Process (MDP). The state is the cable state vector S identified above. An action is a multi-dimensional continuous control command, defined as the adjustment amount of the robot actuator, for example:

[0116] ;

[0117] in, For the increase in clamping force, and These are the linear and angular velocity increments of the end effector in the base coordinate system. Cable release rate coefficient. The reward function is carefully designed to simultaneously optimize cabling quality, efficiency, and security, for example, using a reward function of the following form:

[0118] ;

[0119] in, Punish cables for excessive bending and twisting. Encourage quick task completion. Punishing high-tension risks, A large positive reward is provided for successful wiring. Advanced deep reinforcement learning (DRL) algorithms such as Proximal Policy Optimization (PPO) are used to train a policy network π(a|s) through interaction with a simulation environment or a real system. This network can directly output the optimal adjustment action A based on the real-time state S.

[0120] S40 corresponds to the execution and control layer, which returns the current optimal action command to the Markov environment, and controls the actuator of the wiring robot to perform the corresponding action based on the current optimal action command through the motion controller.

[0121] The motion controller includes: an adaptive motion controller, a compliant gripping and tension management unit, and a safety monitoring and emergency stop module. The adaptive motion controller receives motion commands A from the DRL strategy and calculates them into desired trajectory points in the robot's joint space or Cartesian space. This controller incorporates a feedforward-feedback composite control law. The feedforward part compensates for nonlinear coupling based on the robot's dynamics model, while the feedback part (such as PID or impedance control) suppresses model errors and external disturbances to ensure tracking accuracy.

[0122] Compliant gripping and tension management unit: Specifically controls the end effector grippers. It not only adjusts according to ΔF... g Dynamically adjustable clamping force prevents slippage or cable damage, and may also integrate a miniature servo-driven active cable feed / recruitment mechanism, adjusting λ... r It can proactively manage the tightness of cables and achieve "hand-cable coordination".

[0123] Safety monitoring and emergency stop module: Real-time monitoring of joint torque, motor current, collision sensor signals, etc., and setting multiple levels of safety thresholds. Once an unrecoverable abnormal state is detected (such as a sudden increase in tension exceeding the fracture threshold, or a collision with a rigid obstacle), an emergency stop or a preset safety retraction action is immediately triggered to protect the equipment and cables.

[0124] S50 corresponds to the learning and evolution layer. It calculates the immediate reward based on the new state of the cable after the wiring robot performs the corresponding action, and stores the original state, the action taken, the immediate reward, and the new state in the experience replay database during this execution process, and returns to S20 until the wiring task is completed.

[0125] For adaptive policy optimization networks, the state is represented by a structured state vector. Alternatively, it can be combined with its classification labels to serve as an intelligent agent's observation of the environment.

[0126] Action: Defined as a multi-dimensional continuous action space A as described above. After the action is output, it is converted into robot control instructions by the execution layer.

[0127] Reward: A well-designed reward function is key to guiding the agent's learning. An example reward function is as follows:

[0128] ;

[0129] in, Curvature penalty: penalizes excessive bending of the cable;

[0130] : The maximum curvature of the cable at the current moment;

[0131] : Preset safety curvature threshold;

[0132] The cumulative torsional angle of the cable around its centerline;

[0133] : Safety torsion angle threshold;

[0134] The global average tension level of the cable along its length;

[0135] Safety tension threshold;

[0136] : The velocity of the pressure centroid shift in a tactile image;

[0137] : The count rate of acoustic emission signals;

[0138] : Rewards for progress in cabling;

[0139] A large, one-time reward or penalty upon completion of the task;

[0140] Weighting coefficient.

[0141] A large positive reward is given when the task is successfully completed, and a large negative reward is given when the task fails (such as a broken cable or a severe jam).

[0142] The Proximal Policy Optimization (PPO) algorithm is used to update the policy optimization network because it has excellent stability and sample efficiency in continuous control tasks.

[0143] Policy optimization network structures include: Policy Network (Actor) And value networks (Critic) .

[0144] Typically, a shared feature extraction layer is used, followed by bifurcation to output the action probability distribution (mean and variance) and state value estimate.

[0145] Optimization objective: PPO ensures stability by limiting the magnitude of each policy update. Its objective function is: ,in, It is the probability ratio of the new strategy to the old strategy. It is the advantage function estimated by the Critic network.

[0146] The flexible cable automatic routing method further includes: calculating immediate rewards based on the new cable state, and storing the original state, actions taken, immediate rewards, and new states during the current execution process in an experience replay database; then, using a digital twin simulation environment and a proximal policy optimization algorithm, training the policy network and value network of the policy optimization network. The training of the policy network and value network includes: collecting training data through extensive trial and error and interaction with the environment in the digital twin simulation environment, where the training data includes: the original state, actions taken, immediate rewards, and new states; training the initial policy network and initial value network using the proximal policy optimization algorithm; deploying the pre-trained policy network to the routing robot; generating action information based on the real cable state through the policy network; executing actions based on the action information through the routing robot, collecting new experience, and storing it in an experience replay buffer; sampling small batches of data from the buffer during idle or low-load periods to fine-tune the policy network and value network online to adapt them to the differences between the real world and the simulation; and periodically using the accumulated large amount of real data for offline reinforcement learning or batch retraining to further optimize the policy network and value network.

[0147] The experience replay database can continuously store the entire process data of each cabling task, including raw sensor data, extracted features, identified state S, actions taken A, immediate rewards obtained R, and the final task result (success / failure and reason). The data is stored in time series format, forming a rich set of state-action-reward sample pairs.

[0148] By using an offline batch processing and model fine-tuning engine, batch data can be sampled from the experience replay database periodically or during system idle time to jointly fine-tune or retrain the state recognition model and DRL policy network. This allows the system to digest historical experience, correct model biases, refine better policies, and achieve incremental performance improvements.

[0149] A high-fidelity robot-cable-environment digital twin simulation model can be built using a digital twin simulation accelerator. New strategies or model parameters can be trained and validated in the simulation environment on a massive scale, at high speed, and without risk. Mature strategies can then be transferred to the physical system, greatly accelerating the learning process and reducing trial-and-error costs.

[0150] As described above, the training of the policy network and value network is implemented offline in combination with online methods. First, in a digital twin simulation environment, the agent interacts with the environment through extensive trial and error to collect data (s, a, r, s'), which is then pre-trained using ground-based simulation. The PPO algorithm is then used to train the policy network and value network until the policy converges to an initial policy that can effectively complete the wiring task. Physical transfer and online fine-tuning are then performed: the pre-trained policy network... Deployed to a real robotic system. During real-world operations: 1. The policy network generates action A(t) based on the real state S(t); 2. The robot executes the action, and the system collects new experience and stores it in an experience replay buffer; 3. When the system is idle or under low load, small batches of data are sampled from the buffer to fine-tune the policy network and value network online, adapting them to the differences between the real world and simulation (i.e., Sim-to-Real transfer); 4. The accumulated large amount of real data is periodically used for offline reinforcement learning or batch retraining to further optimize the policy.

[0151] The automatic cabling method for flexible cables also includes: monitoring for operational anomalies and employing a tiered response strategy for detected anomalies. The tiered response strategy includes: adaptive adjustment at the policy level, rule-based rapid intervention, and hardware-level emergency stop.

[0152] Policy layer adaptive adjustment: Common and non-urgent state anomalies are handled by the policy network. For example, if an increased risk of bending is detected, the policy will automatically output a speed reduction. and adjusting posture The action.

[0153] Rule-based rapid intervention: Setting safety thresholds:

[0154] ;

[0155] When the state variable exceeds the safety threshold, the output of the policy network is immediately interrupted, and the system switches to a preset, rule-based safety controller to perform conservative but safe actions such as "immediate stop, slight backtracking, and slight relaxation of clamping force".

[0156] Hardware-level emergency stop: In the event of extreme danger signals (such as a violent collision detected by the joint torque sensor), the hardware emergency stop circuit of the wiring robot is triggered to ensure physical safety.

[0157] The flexible cable automatic routing method of this invention enables system integration and real-time performance assurance. Its software architecture uses a Robot Operating System (ROS 2) as middleware, leveraging its advantages in node communication, data logging, and toolchain. The perception, recognition, decision-making, and control modules are encapsulated as independent, parallel-computable ROS nodes. Real-time performance assurance employs critical control loops (such as state recognition to action generation) running on a Real-Time Operating System (RTOS) or Linux with a real-time kernel. A Time-Triggered Time (TTT) architecture or carefully designed priorities ensures deterministic control cycles (e.g., 10ms) even in worst-case scenarios. Data streams (sensor data) are aggregated via high-bandwidth buses (e.g., EtherCAT, PCIe). Feature extraction and state recognition modules may be deployed on dedicated hardware (e.g., GPU, NPU) to accelerate inference. Action commands generated by the decisions are sent to the robot actuator via real-time Ethernet.

[0158] This invention constructs a closed-loop control scheme that integrates high-precision perception, intelligent decision-making, adaptive learning, and precise execution. It aims to solve the core problems in the automated wiring of flexible cables, such as unstable wiring quality and easy cable damage caused by the dynamic and unpredictable changes in the physical state of the cables (such as tension, torsion, and curvature). By comprehensively perceiving the cable state through a multimodal sensor network and dynamically adjusting the robot's operation strategy using advanced learning algorithms, it achieves a paradigm shift from "blind operation" to "state perception-intelligent decision-making".

[0159] Figure 2 The overall process of automatic flexible cable routing according to the present invention is illustrated. After system startup, the system first completes the initialization of multiple modules, including the robotic arm and various sensors, and performs self-checks on the communication link and power status to ensure that the control system enters the routing process under stable and safe operating conditions. After initialization, the system acquires the spatial position and attitude characteristics of the cable end through the end-effector vision imaging camera, and constructs a unified coordinate system based on this to provide accurate geometric reference for subsequent trajectory planning. Subsequently, the system generates the end-effector insertion force trajectory based on the socket calibration position and cable morphological characteristics, and pre-configures various parameters of the impedance control model (each controller, module, unit, or actuator) in conjunction with the cable flexibility, bending trend, and insertion force direction to ensure smoothness and stability during insertion.

[0160] During the insertion phase, the robotic arm's end effector advances towards the socket along the planned trajectory. Simultaneously, the system collects key feedback information such as insertion force and end effector displacement in real time. Through impedance control, it achieves dual closed-loop adjustment of force and position, ensuring a smooth and stable insertion process even in the presence of nonlinear interference, friction, or orifice deviation. If the system detects excessive attitude error, abnormally increased insertion force, or end effector position deviation exceeding a threshold, it immediately activates an automatic attitude correction mechanism. This mechanism restores the insertion direction through actions such as yielding, attitude adjustment, and trajectory fine-tuning, thereby preventing cable bending, socket damage, or actuator overload. If the system determines that the insertion action may pose a risk of jamming, it will further implement anti-jamming measures, including reducing the advancement speed, increasing compliance flexibility, and replanning the local trajectory, to ensure the insertion task can continue safely and continuously.

[0161] When the insertion depth of the end reaches the preset maximum value or the system detects that the mechanical state is stable, the insertion action is declared complete. The system then performs an end release or reset operation and prepares to process the next cable. The entire process forms a closed-loop operation system through multi-source sensing, real-time dynamic modeling, attitude self-correction, and compliant force control mechanisms, enabling flexible cable insertion tasks to maintain high reliability and high success rate even in complex environments.

[0162] Figure 3 This paper illustrates the key processes of flexible cable conformal insertion trajectory planning and real-time control. First, a vision sensor acquires the three-dimensional pose information of the cable end, including the attitude rotation matrix and spatial translation vector in the base coordinate system, thus constructing the initial spatial state for the insertion task. Subsequently, based on the geometric position, orientation, and attitude of the socket orifice, the system establishes the insertion target point and target attitude, giving the insertion trajectory clear end-effector constraints. To improve the safety and adaptability of the insertion process, the system automatically selects the conformal direction based on the real-time force direction and dynamically adjusts the equivalent stiffness of the end effector along this direction, enabling the robotic arm to exhibit higher compliance in the force direction and effectively reducing insertion resistance caused by friction, attitude deviation, or orifice error.

[0163] When generating the insertion trajectory, the system comprehensively considers the initial posture of the cable end, the socket direction, and the natural bending trend of the cable in space to construct a continuous and smooth conformal trajectory curve. It also predicts potential resistance points to reduce peak forces and dynamic impacts during insertion, achieving a natural and smooth transition from free space to contact space. The entire trajectory not only satisfies positional constraints but also dynamically adjusts local paths based on real-time force feedback, giving it greater environmental adaptability.

[0164] During trajectory execution, the system sets up real-time force-position closed-loop control points at multiple key nodes. By monitoring changes in insertion force and end-effector position deviation, it adaptively corrects the end-effector's motion commands, ensuring that the insertion force remains within a safe range and is evenly distributed. If an increase in local attitude deviation, a sudden change in force direction, or obstruction in insertion is detected, the closed-loop control system immediately adjusts compliance parameters or corrects the attitude to prevent jamming, scraping, or cable end deformation. Ultimately, through continuous force-position coordinated control, the insertion process remains stable, smooth, and shock-free, ensuring that the flexible cable can be accurately connected to the target socket, achieving highly reliable automated cabling operations.

[0165] Figure 4 The overall control architecture of the flexible cable automatic cabling system of the present invention is shown. This architecture adopts a layered design, modularizing sensor information acquisition, model calculation, strategy decision-making, and actuator control, and achieving low-latency data synchronization across modules through a high-speed industrial communication network. The entire control system consists of four core parts: a perception layer, a fusion layer, a decision-making layer, and an execution layer, forming a closed-loop control link that runs through "state acquisition—model inference—control command generation—execution feedback".

[0166] In the perception layer, a multi-source sensing system is deployed, which is responsible for collecting key data such as contact force information, robotic arm motion status, cable spatial orientation, and end-effector force changes. All sensors perform high-speed sampling through a PLC or industrial gateway and transmit the data in real time to the host computer controller via EtherCAT or TSN industrial Ethernet to ensure that the data reaches the computing module within a millisecond timescale.

[0167] In the fusion layer, a pose estimation model based on extended Kalman filtering, strong tracking filtering, and deep learning is employed to perform unified coordinate system transformation and fusion estimation on heterogeneous data from different sensor sources. The fusion layer not only constructs the real-time six-DOF pose of the cable end but also estimates the cable's potential bending tendency, force direction, and friction state, providing reliable state inputs for compliant control and trajectory optimization. Simultaneously, this layer also includes a digital twin model, which uses high-precision dynamic simulation to predict future force changes and motion trajectories of the robotic arm, providing a forward-looking reference for the control layer.

[0168] At the decision-making level, impedance control, adaptive compliance adjustment, and force-position hybrid control work in tandem. Impedance control dynamically adjusts control parameters M (mass), B (damping), and K (stiffness) based on real-time force and displacement errors, ensuring the robotic arm's compliance and anti-interference capabilities during insertion. Compliance adjustment automatically generates a compliant axial direction based on the force direction and reduces the equivalent stiffness at the end effector when insertion encounters resistance, improving safety. Force-position hybrid control integrates model prediction and real-time feedback signals to achieve automatic switching or fusion of force control and position control, ensuring smooth insertion without sacrificing precision. Anomaly detection continuously monitors the system's operating status; upon detecting anomalies such as excessive force, jamming, or posture deviation, it immediately triggers yielding and safety strategies to prevent equipment damage.

[0169] In the execution layer, optimized and corrected control commands are sent to the servo drives of each joint via the EtherCAT network, driving the robotic arm to perform flexible plugging movements. This layer is not only responsible for translating control strategies into actual mechanical motion, but also for transmitting execution results and the internal state of the drives back to the host computer in real time, allowing the system to update its state and make control decisions for the next cycle. The execution layer forms a high-speed closed loop with other layers, enabling the entire system to iterate and update at a high frequency and respond instantly to external disturbances.

[0170] Figure 4 The control architecture achieves deep integration of multi-source sensing, model prediction, adaptive control, and execution feedback, providing a highly robust, precise, and safe control foundation for flexible cable routing tasks. The entire architecture ensures that the system maintains stable force-position coordination insertion capabilities even in complex environments, significantly improving the intelligence level and success rate of the flexible cable routing process.

[0171] On the other hand, the present invention provides an automatic wiring system for flexible cables. Figure 5 This is a structural diagram of a flexible cable automatic wiring system according to an embodiment of the present invention. (Reference) Figure 5 The flexible cable automatic cabling system includes: a multimodal sensing module, a status recognition module, a decision-making module, an execution and control module, and a learning and evolution module.

[0172] The multimodal sensing module is used to collect multidimensional physical information at a preset frequency through multiple sensing modules installed on the wiring robot.

[0173] The state recognition module is used to process the multidimensional physical information and input the processed multidimensional physical information into the multimodal fusion recognition model to generate the current cable state vector.

[0174] The decision module is used to: input the current cable state vector into a deep learning-based adaptive policy network to generate the current optimal action command, wherein the current optimal action command includes commands for impedance control, adaptive compliance adjustment and force-position hybrid control.

[0175] The execution and control module is used to: send the current optimal action command to the motion controller of the wiring robot, so as to control the actuator of the wiring robot to perform the corresponding action through the motion controller.

[0176] The learning and evolution module is used to: calculate the immediate reward based on the new state of the cable, and store the original state, the action taken, the immediate reward, and the new state during this execution process into the experience replay database, and return it to the state recognition module.

[0177] In some embodiments, the execution and control module includes: an adaptive motion controller, a compliant gripping and tension management unit, and a safety monitoring and emergency stop module. The adaptive motion controller receives motion commands from a policy network and calculates them as desired trajectory points in the robot's joint space or Cartesian space. The adaptive motion controller embeds a feedforward-feedback composite control law; the feedforward part compensates for nonlinear coupling based on the robot's dynamics model, and the feedback part suppresses model errors and external disturbances. The compliant gripping and tension management unit controls the gripping force of the end effector of the wiring robot. The safety monitoring and emergency stop module performs real-time monitoring, sets multi-level safety thresholds, and triggers an emergency stop or a preset safety retraction action when an unrecoverable abnormal state is detected.

[0178] This invention proposes a method for flexible cable routing that, by introducing force / torque sensors, adaptive impedance control algorithms, and a real-time path correction mechanism, achieves safe and high-precision cable deployment in complex environments. This method can be widely applied to automated production scenarios such as electrical cabinets, wire harness assembly, and automotive cable splicing, solving the technical problems of unstable splicing, susceptibility to overload damage, and large wiring errors inherent in traditional rigid trajectory control. Specific technical effects are as follows:

[0179] 1. Achieve compliant connection and stress protection

[0180] By integrating a triaxial force sensor and a torque detection module, the system collects real-time force information on the cable during insertion or winding. When a sudden increase in external force or force deviation is detected, the control system can automatically adjust the attitude of the actuator and the direction of the insertion force to achieve smooth avoidance, preventing cable end bending, connector jamming, and interface damage, thus significantly improving the reliability of insertion and operational safety.

[0181] 2. Possesses adaptive impedance control capability

[0182] By constructing a dynamic impedance control model, the system can automatically adjust the damping and stiffness coefficients according to changes in cable stiffness, insertion force, and operating conditions, ensuring that the actuator maintains a stable posture and force balance under different flexible cable conditions. This mechanism effectively prevents excessive insertion force or posture deviation caused by differences in cable flexibility, improving the system's versatility and intelligence.

[0183] 3. Significantly improves wiring accuracy and stability

[0184] The proposed path correction algorithm can adjust the motor control trajectory based on real-time force deviation signals, dynamically correct the insertion path, and achieve "force-position" coordinated control. Compared with traditional fixed trajectory wiring methods, this method can maintain high-precision operation even in complex cable bending or path-restricted environments, with wiring errors controlled within the millimeter range.

[0185] 4. Enhance system security protection and anomaly response capabilities.

[0186] A dynamic force threshold monitoring and emergency stop protection mechanism is implemented. When the rate of force change exceeds a set value, the system automatically enters a safety mode or triggers a buffer stop to prevent equipment impact caused by sudden interference or jamming. This mechanism effectively protects the actuators and wiring tools, extends equipment life, and reduces maintenance costs.

[0187] 5. Possesses learning optimization and adaptive evolution capabilities.

[0188] By recording stress and position data during multiple cabling tasks, this invention can self-learn and optimize the impedance model parameters, enabling the control algorithm to gradually adapt to different cable materials, connector diameters, and assembly environments, thus forming a continuously optimized control strategy. This characteristic gives the system evolutionary capability and high robustness during long-term operation.

[0189] 6. Improve overall automation level and production efficiency

[0190] Through force feedback closed-loop control and path self-correction, the flexible cable splicing task is fully automated, eliminating the need for manual alignment or calibration. The system operates stably and reliably, enabling rapid switching between multiple cable specifications, significantly shortening the assembly cycle, and improving production continuity and flexibility.

[0191] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A flexible cable automatic routing method, characterized by, The method includes: Multiple sensing modules installed on the wiring robot collect multi-dimensional physical information at a preset frequency; The multidimensional physical information is processed, and the processed multidimensional physical information is input into the multimodal fusion recognition model to generate the current cable state vector; The current cable state vector is input into an adaptive policy optimization network based on deep reinforcement learning to generate the current optimal action command, wherein the current optimal action command includes commands for impedance control, adaptive compliance adjustment and force-position hybrid control. The current optimal action command is sent to the motion controller of the wiring robot, so that the actuator of the wiring robot can be controlled by the motion controller to perform the corresponding action; The instant reward is calculated based on the new state of the cable after the cabling robot performs the corresponding action. The original state, the action taken, the instant reward, and the new state during this execution process are stored in the experience replay database. The process returns to the step of: inputting the current cable state vector into the deep learning-based adaptive policy network to generate the current optimal action instruction until the cabling task is completed.

2. The method of claim 1, wherein, Multiple sensing modules mounted on the wiring robot collect multi-dimensional physical information at preset frequencies, including: The tension of the cable along its length is collected by a distributed micro-tension sensor array mounted on the cabling robot. Cable images are acquired using a multi-view vision system mounted on the wiring robot; High-frequency stress waves and audible sound waves are collected by an acoustic sensing module installed on the wiring robot. The contact pressure between the wiring robot and the cable is collected by a tactile sensor array installed on the wiring robot.

3. The method according to claim 2, characterized in that, The multidimensional physical information is processed, and the processed multidimensional physical information is input into a multimodal fusion recognition model to generate a current cable state vector, including: An algorithm based on hardware timestamps and software interpolation is used to align the multidimensional physical information in time and space. For the spatiotemporally aligned multidimensional physical information, a feature extraction algorithm is run in parallel to obtain the high-dimensional feature vector of the multidimensional physical information; The high-dimensional feature vector is input into the multimodal fusion recognition model to generate the current cable state vector, wherein the multimodal fusion recognition model is a multi-head encoding-fusion decoding neural network.

4. The method according to any one of claims 1 to 3, characterized in that, The current cable state vector is input into an adaptive policy optimization network based on deep reinforcement learning to generate the current optimal action instruction, including: The current cable state vector is obtained in the Markov environment and input into the policy optimization network to obtain the action probability distribution output by the policy optimization network and the current optimal action instruction based on the action probability distribution.

5. The method according to claim 4, characterized in that, The method further includes: calculating the immediate reward based on the new state of the cable, and storing the original state, actions taken, immediate reward, and new state during this execution process into an experience replay database; then, using a digital twin simulation environment and a near-end policy optimization algorithm, training the policy network and value network of the policy optimization network, wherein training the policy network and the value network includes: In a digital twin simulation environment, training data is collected through extensive trial and error and interaction with the environment. The training data includes: the original state, the actions taken, the immediate reward, and the new state. The initial policy network and initial value network are trained using a proximal policy optimization algorithm. Deploy the pre-trained policy network onto the wiring robot; The policy network generates action information based on the actual cable status. The wiring robot performs actions based on the motion information, collects new experience, and stores it in the experience replay buffer pool. During idle or low-load periods, small batches of data are sampled from the buffer pool to fine-tune the policy network and the value network online, adapting them to the differences between the real world and the simulation. The accumulated large amount of real data is regularly used for offline reinforcement learning or batch retraining to further optimize the policy network and value network.

6. The method according to claim 4, characterized in that, The current optimal action command is sent to the motion controller of the wiring robot, so that the motion controller controls the actuator of the wiring robot to perform the corresponding action, including: The current optimal action instruction is returned to the Markov environment, and the actuator of the wiring robot is controlled by the motion controller to perform the corresponding action based on the current optimal action instruction.

7. The method according to claim 6, characterized in that, The motion controller includes: an adaptive motion controller, a compliant gripping and tension management unit, and a safety monitoring and emergency stop module, wherein... The adaptive motion controller is used to receive motion commands from the policy network and solve them into desired trajectory points in the robot joint space or Cartesian space. The adaptive motion controller embeds a feedforward-feedback composite control law. The feedforward part is based on the robot dynamics model to compensate for nonlinear coupling, and the feedback part is used to suppress model errors and external disturbances. The compliant gripping and tension management unit is used to control the gripping force of the end gripper of the wiring robot; The safety monitoring and emergency stop module is used for real-time monitoring, setting multiple safety thresholds, and triggering an emergency stop or a preset safety rollback action when an unrecoverable abnormal state is detected.

8. The method according to claim 1, characterized in that, The method further includes: monitoring operational anomalies, and adopting a tiered response strategy for the detected anomalies, wherein the tiered response strategy includes: Policy layer adaptive adjustment: Common and non-urgent state anomalies are handled by the policy network; Rule-based rapid intervention: Set a safety threshold. When the state quantity exceeds the safety threshold, immediately interrupt the output of the policy network and switch to a preset, rule-based safety controller. Hardware-level emergency stop: In the event of an extreme danger signal, the hardware emergency stop circuit of the wiring robot is triggered.

9. A flexible cable automatic wiring system, characterized in that, The system includes: The multimodal sensing module is used to: collect multidimensional physical information at a preset frequency through multiple sensing modules installed on the wiring robot; The state recognition module is used to: process the multidimensional physical information and input the processed multidimensional physical information into the multimodal fusion recognition model to generate the current cable state vector; The decision module is used to: input the current cable state vector into a deep learning-based adaptive policy network to generate the current optimal action command, wherein the current optimal action command includes commands for impedance control, adaptive compliance adjustment and force-position hybrid control; The execution and control module is used to: send the current optimal action command to the motion controller of the wiring robot, so as to control the actuator of the wiring robot to perform the corresponding action through the motion controller; The learning and evolution module is used to: calculate the immediate reward based on the new state of the cable, and store the original state, the action taken, the immediate reward, and the new state during this execution process into the experience replay database, and return it to the state recognition module.

10. The system according to claim 9, characterized in that, The execution and control module includes: an adaptive motion controller, a compliant gripping and tension management unit, and a safety monitoring and emergency stop module, wherein... The adaptive motion controller is used to receive motion commands from the policy network and solve them into desired trajectory points in the robot joint space or Cartesian space. The adaptive motion controller embeds a feedforward-feedback composite control law. The feedforward part is based on the robot dynamics model to compensate for nonlinear coupling, and the feedback part is used to suppress model errors and external disturbances. The compliant gripping and tension management unit is used to control the gripping force of the end gripper of the wiring robot; The safety monitoring and emergency stop module is used for real-time monitoring, setting multiple safety thresholds, and triggering an emergency stop or a preset safety rollback action when an unrecoverable abnormal state is detected.

Images